Latent Space

“If the news is fake — imagine history.” — AmuseChimp via Naval

1. The first news item prompting this editorial is the unofficial but credible reporting that Cursor is now at $2B ARR, and raising at $50B, contra a couple weeks of headlines that Cursor churn is ramping up. In this case, a filter bubble cropped up on X, where novelty and scandal is rewarded and truth was hard to glean.

2. The second news item is the Ars-Technica-Scott-Shambaugh saga, which, there is no polite way to say this, is a veritable clusterfuck of open source abuse, AI Clawbots, and the journalist covering it publishing made up AI quotes, eventually getting fired. As a bonus, human commentors also hallucinated more untruths.

3. The third is a recent episode of a top podcast featuring the idea that the best way to launch products is to make 20 different fake TikTok videos of an app and only building it after the videos go viral, instead of building first and then marketing. (We first covered this idea in the Hyperstitions of Moloch, and it obviously works.)

Left unchecked, the net result of all this is a declining trust in all media, which shatters consensus reality and therefore civil society — I do not have to care what you think is true if I can simply assert my truth louder.

It would be different if we voted with our thumbs. Dead Internet Theory is not solely caused by AI, but is accelerated by it. I am reminded of an old social media evolution framework that goes like this:

• Before “social media”, people only got their news entertainment from magazines and newspapers — designed for mass distribution, so not at all personalized. Gossip was only by literal word of mouth, from a chatty neighbor or coworker.

• Your Friends kill Generic Celebrities: Then Facebook happened, and suddenly you could see news about people you know/might meet and learn all the gossip about them. So much more interesting!

• Professional Friends kill Real Friends: Your real friends are pretty boring. My Instagram posts about what I had for lunch aren’t as interesting as someone who treats their posts like a job; every photo filtered, every story boarded. The influencers best at this become superstars. Great engagement for everyone! But also - is it a surprise that one of the biggest vloggers turned out to be… scripted?

• Recsys Long Tail Content kills Professional Friends: The problem with professional friends is that they had the same problem as the generic celebrities of old — they converge on “merely good enough” for most. They also bring new problems: even when being paid millions they still don’t produce enough content , and, yet, they are also divas who wield too much power over the platforms. So: the platforms stop prioritizing your follow graph (stated preferences) and start feeding your worst impulses (revealed preferences). Another huge jump in engagement, and 100x more creators are given the lottery ticket to Make It Big!

• AI Content kills Human Content: The final stage — personalized creation replacing curation. Everyone lives in a Truman Show cage of their own making, happily swiping away their free will and connection to quotidian reality.

We’re surely stepping closer and closer to this every day — as newswriters attempting to Scale Without Slop, this is a problem we’re trying to navigate and create new solutions for.

At the same time it has never been easier to signal taste and human effort — a lesser newsletter might leave you with a slop image like this:

But for now it still takes a human to make a truly useful graphic:

AI News for 2/27/2026-3/2/2026. We checked 12 subreddits, 544 Twitters and 24 Discords (264 channels, and 31899 messages) for you. Estimated reading time saved (at 200wpm): 2895 minutes. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies!

Qwen 3.5 “small” open models: long-context + multimodal on-device is getting real

• Qwen3.5-0.8B / 2B / 4B / 9B released (Base + Instruct): Alibaba launched a compact series positioned as “more intelligence, less compute,” with native multimodal and scaled RL, explicitly targeting edge + lightweight agent deployments (Alibaba_Qwen). Community amplification highlights 262K native context (extendable to 1M) and competitive scores reported in tweet summaries (e.g., “82.5 MMLU-Pro,” “78.4 MMMU,” “97.2 CountBench”)—treat these as vendor/secondary claims until you read the model cards (kimmonismus).

• Architecture notes emerging via commentary: Multiple tweets converge on Qwen’s move toward hybrid / non-orthodox attention, with “hybrid models” coming back in 3.5 vs the earlier “Thinking vs Instruct” split in Qwen3 updates (nrehiew_). A more detailed (but still unofficial) breakdown claims a Gated DeltaNet hybrid pattern: “3 layers linear attention : 1 layer full attention” to keep memory flat while preserving quality (LiorOnAI).

• Practical deployment caught up fast:

  • Ollama: ollama run qwen3.5:9b|4b|2b|0.8b, with tool calling + thinking + multimodal surfaced in the packaging (ollama, ollama).

  • LM Studio: Qwen3.5-9B touted as ~7GB local footprint (Alibaba_Qwen).

  • iPhone on-device demo: Qwen3.5 2B 6-bit running with MLX on “iPhone 17 Pro” is getting framed as an “edge breakthrough” (adrgrondin, kimmonismus).

• Gotcha for evaluators: “Reasoning disabled by default” on the small models; enable via chat-template kwargs (example given for llama-server / Unsloth docs) (danielhanchen).

Coding agents + reliability + “availability is the new frontier”

• Codex 5.3 and coding eval chatter: Anecdotal reports of Codex 5.3 solving “promising” tasks and pushing benchmarks like WeirdML (79.3% claim, leading v. Opus 4.6 at 77.9%) while noting Gemini peak performance may still be higher (theo, htihle). Also speculation about nearing saturation on WeirdML v2 (teortaxesTex).

• “We’re about to hit 1 9 of availability”: The emerging ops pain point is not only model quality but downtime and degraded UX; the theme repeats across memes and serious complaints about Claude outages and productivity impacts (ThePrimeagen, Yuchenj_UW, Yuchenj_UW).

• Agent observability / evaluation becomes a first-class problem:

  • “Since we’re all agent managers now, what’s your favourite way to get observability?” (_lewtun).

  • Agent reliability is cross-functional (can’t “engineer” your way out of bad eval criteria; PMs/domain experts must own success definitions) (saen_dev).

  • Practical eval advice: define success before building; start with deterministic graders; use LLM judges for style; grade the produced artifact not the path (_philschmid).

• AGENTS.md / SKILL.md as “guardrails,” not magic:

  • A reported Codex study across 10 repos / 124 PRs: AGENTS.md reduced median runtime ~28.6% and tokens ~16.6%, mostly by reducing worst-case thrashing rather than uniform gains (omarsar0).

  • Carnegie Mellon-style loop for SKILL.md improvement in production: “log → evaluate → monitor → improve” with an OSS example (PR review bot) (gneubig).

• Anthropic-as-coding-org tension: A viral datapoint claims “80%+ of all code deployed is written by Claude Code,” paired with concern that speed may be coming with reliability regressions (GergelyOrosz). Separate threads discuss Claude Code adoption inside major companies and “supervision” replacing manual coding (_catwu, Yuchenj_UW).

Infra + local AI hardware: Apple Neural Engine cracks, Docker/vLLM on macOS, and “AI infrastructure year”

• Reverse-engineering Apple’s Neural Engine for training: A highly engaged thread claims a researcher built a transformer training loop on the ANE using undocumented APIs, bypassing CoreML; heavy ops on ANE, some gradients still on CPU. Also contains efficiency claims like “M4 ANE 6.6 TFLOPS/W vs 0.08 for A100” and “38 TOPS is a lie—real throughput 19 TFLOPS FP16”—these specifics should be verified against the repo/paper, but the meta-point is: on-device training/fine-tuning might be opened up (AmbsdOP, plus ecosystem note AmbsdOP; additional technical summary LiorOnAI).

• macOS local serving gets smoother: Docker Desktop “Model Runner” adds support to run MLX models with OpenAI-compatible API workflows; positioned as a practical unlock for Apple Silicon dev loops (Docker).

• Inference hardware divergence: A GPU vs Taalas HC explainer contrasts software-executed models on GPUs (HBM streaming + kernel scheduling bottlenecks) vs “model-as-hardware” ASIC with weights in mask ROM; claims 16–17k tok/s per user for HC1 with tradeoff “one chip = one model” (TheTuringPost).

• Open-source perf tooling: AMD open-sourced rocprof-trace-decoder (SQTT trace defs) enabling deeper instruction-level timing traces; framed as AMD tracing infra being “better than NVIDIA’s” (tinygrad).

• AI infra as strategic theme: Zhipu’s “2026 is the year of AI infrastructure” is more slogan than spec, but fits the overall signal: reliability + cost + tooling now dominate marginal model improvements (Zai_org).

New research + benchmarks: transformer scaling theory, MuP edge cases, CUDA-kernel RL, and “bullshit detection”

• Transformer scaling theory refresher: “Effective Theory of Wide and Deep Transformers” (Meta) re-circulated as a 60+ page analysis of forward/backward signal propagation, width scaling rules, hyperparameter scaling, NTK analysis, and optimizer behavior (SGD vs AdamW), with validation on vision/language transformers (TheTuringPost, arXiv link tweet).

• Beyond MuP / Muon stability corner cases: Discussion of stability metrics for Embedding / LM head / RMSNorm layers and why embedding + LM head can “not play well with Muon” (Jianlin_S).

• CUDA Agent (ByteDance): Widely shared as a meaningful step beyond “code that compiles” toward “code that’s fast,” using agentic RL with real profiling-based rewards. Claimed SOTA on KernelBench, big gains vs torch.compile, and competitive vs frontier LLMs on hardest kernels (HuggingPapers, deep thread BoWang87).

• BullshitBench v2: Benchmark update adds 100 new questions split across coding/medical/legal/finance/physics, tests 70+ model variants, and claims reasoning often hurts; Anthropic models allegedly dominate and OpenAI/Google are “not improving” on this benchmark (petergostev, reaction scaling01).

• Scheming eval realism: Advice that “contrived environments” can invalidate scheming results; emphasizes careful environment design (NeelNanda5).

Agents + product/toolchain releases: repo graphs, Stripe LLM billing proxy, LangChain refresh, Llama.cpp packaging

• GitNexus (browser-only repo knowledge graph + “graph RAG” via Cypher): Parses repos into an interactive D3 graph, stores relations in embedded KuzuDB, and answers queries via graph traversal (Cypher) instead of embeddings; notable for doing it in-browser with Web Workers and MIT licensing (MillieMarconnni).

• Stripe-style billing for LLMs: Launches “billing for tokens” where you pick models, set markup, route calls via Stripe’s LLM proxy, and record usage automatically—an indicator that “LLM ops” is moving into standard SaaS finance plumbing (miles_matthias).

• LangChain rebrand / consolidation: “Meet our final form” relaunch of LangChain’s web presence (signal is primarily product/positioning, not a spec drop) (LangChain).

• llama.cpp distro packaging: Request for feedback on official Debian/Ubuntu packages—small, but meaningful for mainstreaming local inference tooling (ggerganov).

• MCP vs “Agent Skills” clarification + Weaviate skills repo: Clean distinction: MCP servers as deterministic API interfaces vs markdown “skills” as behavior guidance; Weaviate publishes skills-based integration patterns for common agent tools (weaviate_io).

US DoW–OpenAI–Anthropic “supply chain risk” saga: contract language, surveillance loopholes, and policy trust boundaries (high-level)

• Stratechery frames a standoff: Anthropic vs DoW is positioned as a misalignment between legitimate concerns and government reality (stratechery).

• Reporting disputes OpenAI’s “red lines” framing: The Verge claims DoD didn’t agree to the red lines the way OpenAI implied (haydenfield). Separate threads emphasize: without full contract text, it’s hard to validate any public claim about enforceability or “freezing” laws in time (jeremyphoward).

• Sam Altman posts contract amendment language: Adds explicit prohibition on “intentional” domestic surveillance of US persons, including via commercially acquired identifiers, and says intelligence agencies (e.g., NSA) are excluded without follow-on modification; also acknowledges Friday announcement was rushed (sama, additional principles post sama).

• Pushback: “intentional/deliberate” may preserve the classic “incidental collection” loophole: Multiple legal-minded threads argue the amendment may still allow broad collection if framed as incidental, and that “metadata/hashed identifiers” can evade “personal or identifiable” definitions. Repeated call: independent red-teaming by counsel, and ideally full contract review (j_asminewang, David_Kasten, justanotherlaw, _NathanCalvin).

• Anthropic safeguards claims: Anthropic-adjacent staff dispute a narrative that Anthropic offered an unconstrained “helpful-only” natsec model; claim Claude Gov includes additional training + safeguards + classifier stack (sammcallister).

• Policy meta: A recurring engineering-relevant point is that governance and contract semantics are becoming production constraints on model deployment—no longer “PR side quests.” See also the “AI politics fissure is taking advanced AI seriously vs not” framing (deanwball).

• Qwen 3.5 Small Model Series launch (0.8B/2B/4B/9B, multimodal, scaled RL, Base models too) — @Alibaba_Qwen

• Reverse-engineered Apple Neural Engine; training loop on ANE — @AmbsdOP

• Qwen3.5 small models now in Ollama — @ollama

• Sam Altman: DoW contract amendment language re domestic surveillance + intel agency scope — @sama

• CUDA Agent: RL for high-performance CUDA kernel generation via profiler-based reward — @BoWang87

• “80%+ of code deployed is written by Claude Code” + reliability concern — @GergelyOrosz

• GitNexus: in-browser repo → knowledge graph + Cypher graph-RAG agent — @MillieMarconnni

Second wave speakers for AIE Europe and CFP for AIE World’s Fair are announced today, and OpenCode is confirmed for Miami! We’ll also be in Melbourne & Singapore.

Editor: This is the latest in our guest post program, where we will publish AI Engineering essays worth considering, even if we don’t personally agree with them — having just shipped an AI review tool, this is one of those cases where I am not there yet, but is clearly on the horizon, and am happy for Ankit to argue the case!

Humans already couldn’t keep up with code review when humans wrote code at human speed. Every engineering org I’ve talked to has the same dirty secret: PRs sitting for days, rubber-stamp approvals, and reviewers skimming 500-line diffs because they have their own work to do.

We tell ourselves it is a quality gate, but teams have shipped without line-by-line review for decades. Code review wasn’t even ubiquitous until around 2012-2014, one veteran engineer told me, there just aren’t enough of us around to remember.

And even with reviews, things break. We have learned to build systems that handle failure because we accepted that review alone wasn’t enough. This shows in terms of feature flags, rollouts, and instant rollbacks.

Teams with high AI adoption complete 21% more tasks and merge 98% more pull requests, but PR review time increases 91%, based on data from over 10,000 developers across 1,255 teams.

Two things are scaling exponentially: the number of changes and the size of changes. We cannot consume this much code. Period. On top of that, developers keep saying that reviewing AI-generated code requires more effort than reviewing code written by their colleagues. Teams produce more code, then spend more time reviewing it.

There is no way we win this fight with manual code reviews. Code review is a historical approval gate that no longer matches the shape of the work.

AI code review tools are just buying us time. If AI writes the code and AI reviews it, why do we need a pretty review UI to display that? As much as AI code reviews can be valuable, these will shift left in the dev cycle. There’s no reason to waste CI resources and manage versioning in between the review cycles.

Post-PR review made sense when humans wrote code and needed fresh eyes. When agents write code, “fresh eyes” is just another agent with the same blind spots. The value is in the iteration loop here, not as an approval gate.

We know from experience agents are not always reliable, and it’s very human to think, I caught the AI doing something dumb once; therefore, I must always check it. That instinct made sense when manual verification was feasible. At the current scale, it’s not anymore. And it’s just going to get worse.

The answer is to move the human checkpoint upstream. If the thought of not reviewing code seems scary, let me remind you that checkpoints have moved before in software development. We moved from waterfall sign-offs to continuous integration. We can move them again.

Spec-driven development is becoming the main way of working with AI. Humans should review specs, plans, constraints, and acceptance criteria—not 500-line diffs.

In this new paradigm, Specs become the source of truth. Code becomes an artifact of the spec. You don’t need to review the code. You review the steps. You review the verification rules. You review the contract the code must fulfill.

Human-in-the-loop approval moves from “Did you write this correctly?” to “Are we solving the right problem with the right constraints?” The most valuable human judgment is exercised before the first line of code is generated, not after.

How comfortable do we need to get before we stop reading the code?

In rule form:

Code must not be written by humans

Code must not be reviewed by humans

LLMs are not great at following commands. They deviate. Frequently. And they’re unreliable at self-verification—they’ll confidently tell you the code works while it’s on fire. The fix isn’t to ask the LLM to verify. It’s to ask it to write a script that verifies. Shift from judgment to artifact.

Trust is layered. This is the Swiss-cheese model: no single gate catches everything. You stack imperfect filters until the holes don’t align. So, where else can we put approval gates?

Instead of asking one agent to get it right, ask three agents to try differently and pick the best outcome. Let them compete. The cost of optionality is the lowest in the history of software engineering.

The selection doesn’t have to be manual either. You can rank outputs by which one passes the most verification steps, which one produces the smallest diff, which one doesn’t introduce new dependencies. Competition creates a signal you wouldn’t get from a single attempt.

There should be a deterministic way to verify the work. Tests, type checks, contract verification - things that don’t have opinions, just facts.

Instead of asking an LLM “Did this work?” you define verification steps that produce a series of pass/fail artifacts. The agent can’t negotiate with a failing test. It either meets the specification or it doesn’t.

These guardrails can be defined as layers themselves:

• Coding guidelines - these can be custom linters

• Organization-wide invariants - the non-negotiables, e.g. No hardcoded credentials, API keys, or tokens

• Domain Contracts - specific to a framework, a service, or a part of the code base, e.g. Payments domain: All amounts use the Money type

• Acceptance Criteria - specific to the task

Verification steps should be defined before the code is written, not invented after to confirm what’s already there. If the agent writes both the code and the tests, you’ve just moved the problem—now you’re trusting the agent to test the right things. Verification criteria need to come from the spec, not from the implementation.

So where do humans add value? Upstream, defining what success looks like.

This is where Behavior-Driven Development becomes newly relevant. BDD was always a good idea—write specifications in natural language that describe expected behavior, then automate those specs as tests. But it never fully caught on because writing specs felt like extra work when you were also going to write the code.

With agents, the equation flips. The spec isn’t extra work; it’s the primary artifact. You write:

The agent implements. The BDD framework verifies. You never have to read the implementation unless something fails.

This is humans doing what humans are good at: defining what “correct” means, encoding business logic and edge cases, thinking about what could go wrong. The agent handles the translation from intent to code. The BDD specs become your verification layer—deterministic, automated, and defined before the first line is written.

Acceptance criteria authored by humans, verified by machines. That’s the gate that actually matters.

What can this agent touch? What requires escalation? These become architectural decisions, not afterthoughts.

Most agent frameworks treat permissions as an all-or-nothing setting. The agent either has shell access or it doesn’t. But granularity matters. An agent fixing a bug in a utility function doesn’t need access to your infrastructure configs. An agent writing tests doesn’t need to modify CI pipelines.

Scope should be as narrow as possible while still letting the agent do useful work. If the task is “fix the date parsing bug in utils/dates.py,” the agent’s filesystem access should be limited to that file and its test file. Not the whole codebase. Not “src/ and tests/”. Just the files that matter for this task.

Escalation triggers are equally important. Certain patterns—touching auth logic, modifying database schemas, adding new dependencies—should automatically flag for human review regardless of how confident the agent is.

Separation of responsibilities: One agent does the work, another verifies. They don’t trust each other, and that’s the point.

This is an old pattern—it’s why your QA team shouldn’t report to your engineering manager, and why the person who writes the code shouldn’t be the only one who reviews it.

With agents, you can enforce this architecturally. The coding agent has no knowledge of what the verification agent will check. The verification agent has no ability to modify the code to make its own job easier. They’re adversarial by design.

You can take this further: a third agent attempts to break what the first agent built, specifically targeting edge cases and failure modes. Red team, blue team—but automated and running on every change.

The incentive of an agentic system is simple: given a task, can I complete it? Can I please the person who gave it to me? The agent’s success is never inherently driven by long-term accuracy or business requirements.

It’s our job to encode in the constraints.

For code generated by agents and read by agents, what “good code” looks like will become more standardized. For a new codebase, you’ll have to provide less direction because the defaults will be more consistent.

The future is ship fast, observe everything, revert faster.

Not: review slowly, miss bugs anyway, debug in production.

We’re not going to outread the machines. We need to outthink them—upstream, where the decisions actually matter.

Ultimately, if agents can handle the code just fine, what does it matter if we can read it or not?

Ankit Jain is the founder and CEO of Aviator, where he’s building the infrastructure for AI-native engineering teams. Aviator’s platform helps modern organizations improve AI adoption while maintaining high engineering standards.

AI News for 2/26/2026-2/27/2026. We checked 12 subreddits, 544 Twitters and 24 Discords (263 channels, and 12529 messages) for you. Estimated reading time saved (at 200wpm): 1189 minutes. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies!

Against the backdrop of nonstop positioning with the Department of War (Anthropic refusing terms vs OpenAI doing a deal), OpenAI finally closed the much debated Big Round that had been started since December. In the post, they make several interesting new disclosures:

• Weekly Codex users have more than tripled since the start of the year to 1.6M

  • was 1M on Feb 4 (!!?!?!)

• More than 9 million paying business users rely on ChatGPT for work

  • was 3M in June 2025

• ChatGPT is where people start with AI, with more than 900M weekly active users, and we now have more than 50 million consumer subscribers (monetization continuing to accelerate in Jan/Feb)

  • was 800M in Oct 2025, 35M paying in July

All this justifies $110B in new investment at a $730B pre-money valuation:

• $30B from SoftBank (“advancing our own ASI strategy”),

• $30B from NVIDIA (including the use of 3 GW of dedicated inference capacity and 2 GW of training on Vera Rubin systems) - down from “up to $100B”, still with circular funding concerns

• $50B from Amazon with increased partnership (analysis) involving:

  • an initial $15 billion investment and followed by another $35 billion in the coming months when certain conditions are met — leaving Amazon with a large stake in both OpenAI and Anthropic

  • “Stateful Runtime Environment” powered by OpenAI on Amazon Bedrock

  • AWS will be the exclusive third-party cloud provider for OpenAI Frontier

  • 2 gigawatts of Trainium capacity through AWS infrastructure worth “$100 billion over 8 years”, spanning both Trainium3 and next-gen Trainium4 chips

Close watchers might notice the absence of Microsoft, which continues the existing reduced partnership and gets the stateless APIs.

To put this in perspective, 118 countries/economies have a nominal GDP below $100B — roughly 61% of all world economies. Because the consecutive “largest fundraises in history” are too big to fit in a human head, here’s a chart worthy of wtfhappened2025.com:

and outside of AI, a 10 year history:

and here from OpenAI Deep Research + ChatGPT Canvas, sorted by descending amount:

or a timeline perspective:

Hypernetworks for instant LoRA “compilation”: Doc-to-LoRA + Text-to-LoRA

• Doc-to-LoRA / Text-to-LoRA (Sakana AI): Sakana introduces two related methods that amortize customization cost by training a hypernetwork to generate LoRA adapters in a single forward pass, turning what would be fine-tuning / distillation / long-context prompting into “instant weight updates.” The core claim: instead of keeping everything in an expensive active context window, you can compile task descriptions or long documents into adapter weights with sub-second latency, enabling rapid adaptation and “durable memory”-like behavior (SakanaAILabs, hardmaru).

  • Text-to-LoRA: specializes to unseen tasks from just a natural language description (SakanaAILabs).

  • Doc-to-LoRA: internalizes factual documents; on needle-in-a-haystack, reports near-perfect accuracy on sequences ~5× longer than the base model context window, and even demonstrates a cross-modal trick: transferring visual information from a VLM into a text-only model via internalized weights (SakanaAILabs; recap thread omarsar0).

  • Positioning vs long-context: explicitly framed as a way to reduce quadratic attention costs and avoid rereading long docs at every call—store knowledge in adapters rather than tokens (omarsar0).

• Credit / prior art tension: One researcher complains that Hypersteer (hypernetworks producing steering vectors from text descriptions) did not get sufficient credit in later similar work (aryaman2020). There’s also broad community excitement / “hypernetworks are back” reactions (willdepue, zhansheng).

• Open question raised: why not just use attention with an extremely long KV cache—i.e., is Doc-to-LoRA mainly about efficiency/serving cost? (hyhieu226)

OpenAI financing + deployment transparency tooling

• $110B funding round: OpenAI announces a $110B raise with backing from Amazon, NVIDIA, SoftBank, framed as scaling infra “to bring AI to everyone” (OpenAI, sama). A separate note from Epoch AI contextualizes the scale: The round would nearly triple total capital raised to date; The Information reportedly projects $157B cash burn through 2028, and this round + existing cash would roughly match that projection (EpochAIResearch).

• Deployment Safety Hub: OpenAI launches a searchable site to browse “system cards” (previously PDFs) as a more accessible interface to deployment safety documentation (dgrobinson).

US DoD (“Department of War”) vs Anthropic saga: supply-chain designation, backlash, and industry implications

• Anthropic draws a line; tech reacts: A central flashpoint is Anthropic’s public refusal to enable mass domestic surveillance and fully autonomous weapons (as characterized by posters reacting to Anthropic’s statement), which drew rare cross-competitor praise and heightened attention to “red lines” in frontier deployment (mmitchell_ai, ilyasut).

• Designation shock + legal scope debate: Posts circulate a claimed DoW move to designate Anthropic a “Supply-Chain Risk to National Security” and to pressure contractors/partners—sparking arguments about legality, precedent, and chilling effects (kimmonismus, deanwball). One legal clarification: DoD can restrict what contractors do on DoD contract work, but likely can’t legally ban contractors from using Anthropic in their private/commercial work (petereharrell).

• Economic/strategic fallout framing: The sharpest critiques argue this would damage US credibility as a business partner and potentially force hyperscalers/investors into impossible tradeoffs (deanwball); others note uncertainty until full details are known but still see a supply-chain designation as ill-fitting (jachiam0).

• Public sentiment spike: Posts highlight strong public outrage at the idea of a DoD-backed domestic surveillance program and punishment for refusal (quantian1, janleike). Many users signal “solidarity subscriptions” to Claude (willdepue, Yuchenj_UW).

• Anthropic statement and intent to litigate: Anthropic posts an official statement responding to Secretary Hegseth’s comments (AnthropicAI). Commentary highlights the line “challenge any supply chain risk designation in court” and emphasizes the dispute over restricting customers outside DoD contract scope (iScienceLuvr).

• Meta-point: Regardless of where one lands on Anthropic’s choices, many posts treat this as a governance precedent moment: who decides acceptable use, what due process exists, and how contracts interact with fast-moving model capabilities (kipperrii).

Models + leaderboards: Qwen3.5 expansion and “open model” rankings

• Qwen3.5 new releases (Artificial Analysis summary): Alibaba expands Qwen3.5 with 27B dense, 122B A10B MoE, and 35B A3B MoE, all Apache 2.0, 262K context (extendable to 1M via YaRN per the post). Artificial Analysis reports Intelligence Index scores: 27B = 42, 122B A10B = 42, 35B A3B = 37, with notable agentic/task metrics like GDPval-AA 1205 for 27B, plus detailed tradeoffs (hallucination/accuracy and token usage—27B used 98M output tokens to run the index) (ArtificialAnlys).

• Arena leaderboards (Feb 2026): Arena posts Top Open Models for text and code. Text top-3: GLM-5 (1455), Qwen-3.5 397B A17B (1454), Kimi-K2.5 Thinking (1452) (arena). Code Arena top includes GLM-5 (1451) at #1, with Kimi-K2.5 and MiniMax-M2.5 tied at #2 (arena). Arena also highlights Arena-Rank, their open-source ranking package for reproducible leaderboards (arena).

• Perplexity open-sources bidirectional embedding models (claim): A thread claims Perplexity open-sourced bidirectional “Qwen3-retrained” embedding models (0.6B/4B; standard vs context-aware embeddings; MIT licensed) to improve document-level understanding for retrieval; treat as a third-party summary rather than primary release notes (LiorOnAI).

Systems, inference, kernels, and RL training: bandwidth, ROCm, and off-policy RL

• vLLM ROCm attention backends (AMD): vLLM announces 7 attention backends for vLLM on ROCm with KV-cache layout changes, batching tricks, and model-specific kernels; reported up to 4.4× decode throughput on AMD GPUs with an env var switch (VLLM_ROCM_USE_AITER=1) (vllm_project). A follow-up details MLA KV compression claims (e.g., ~8K → 576 dims) and throughput wins on MI300X/MI325X/MI355X (vllm_project).

• DeepSeek DualPath I/O paper (third-party explainer): A ZhihuFrontier summary describes a DeepSeek+THU+PKU paper proposing system-level redesign of Prefill/Decode to exploit idle storage NIC bandwidth on decode nodes via RDMA, aiming at KV-cache movement bottlenecks for agentic long-context inference; includes claimed speedups (e.g., 1.87× on DS-660B) with caveats for smaller models (ZhihuFrontier).

• Kernel/infra chatter (“quack”, Liger): A thread points to Dao-AILab’s quack writeup on memory hierarchy bandwidth, plus a note that Liger not using cluster-level reductions for xentropy could explain slower performance in some settings (fleetwood___).

• Off-policy RL for reasoning (Databricks MosaicAI): Databricks promotes OAPL (Optimal Advantage-based Policy Optimization with lagged inference policy) as a stable off-policy alternative that can match/beat GRPO while using ~3× fewer training generations, positioned as operationally simpler than strict on-policy loops (DbrxMosaicAI, jefrankle).

• ERL vs RLVR (Turing Post explainer): A long “workflow breakdown” contrasts standard RLVR (scalar verifiable rewards) with Experiential Reinforcement Learning (ERL) inserting within-episode reflection/retry + distillation; cites reported gains (e.g., +81% Sokoban) and tradeoffs (pipeline complexity/compute) (TheTuringPost).

• Mamba-2 / GDN initialization bug discussion: Albert Gu clarifies a viral plot debate: main takeaway is an init bug materially affecting some results; also notes nuanced interactions in hybrids (e.g., “stronger” components can make others “lazy,” with a related reference) (_albertgu, _albertgu).

Top tweets (by engagement, technical / industry-relevant)

• OpenAI raises $110B (sama, OpenAI)

• Sakana AI Doc-to-LoRA / Text-to-LoRA (SakanaAILabs, hardmaru)

• Anthropic–DoD supply-chain designation critique / governance precedent (deanwball, quantian1, janleike)

• Karpathy on coding workflow evolution (tab → agents → parallelism) (karpathy)

• Karpathy on “programming a research org” with multi-agent workflows; limitations observed (karpathy)

• Anthropic official statement (AnthropicAI)

We’re excited to announce our new show In-Context Cooking where guests cook while chatting about cool topics.

For our first episode, we have the Founder & CEO of SemiAnalysis Dylan Patel. Dylan went from anonymous Silicon Twitter chip poster and rural Georgia beekeeper to one of the most cited analysts in AI infrastructure, advising top labs, hyperscalers, hedge funds, and semiconductor giants while tracking everything from GitHub commit share to semiconductor fab bottlenecks.

This Substack is reader-supported. To receive new posts and support my work, consider becoming a free or paid subscriber.

Dylan joins us in the kitchen to recreate restaurant-style chicken fried rice while unpacking Taiwan endgames, export controls, Nvidia’s paranoia advantage, why Amazon and Google are about to light $200B/year on fire (intentionally), and why the biggest AI risk might not be China but voters.

We discuss:

• Dylan’s origin story: rural Georgia → Minnesota beekeeper → anonymous chip blogger → building SemiAnalysis into a 60-person global research firm

• Starting a Substack because “I could do better”: anonymous WordPress posts, Silicon Twitter, Doug’s push to charge, and turning consulting into a company

• Taiwan endgame scenarios: status quo vs. KMT vs. DPP vs. political destabilization, why a “soft” takeover is more plausible than invasion, and how U.S. export controls still bind TSMC regardless of party

• Export controls & AI sovereignty: Dario-style hard containment vs. Nvidia-style ecosystem leverage, whether chips or models matter more, and how far behind China really is

• AI escape velocity: why Claude Code adoption jumping from 2% → 4% of GitHub commits in a month matters, and why coding agents are the first real trillion-dollar unlock

• The hyperscaler capex explosion: $180B (Google) and $200B (Amazon) in AI infrastructure, why the market hates it, and why Dylan thinks profits disappear by 2027

• The innovator’s dilemma at planetary scale: why Meta, Microsoft, Google, and Amazon must spend or die which is Pascal’s Wager for digital god

• Debt-financed AI clusters: Meta’s Louisiana buildout, leverage as strategy, and what happens when the most profitable companies ever stop caring about free cash flow

• The coming AI backlash: general public hostility, labor displacement, income inequality optics, and why “anti-AI” could become a winning political platform

• Nvidia vs. vertical integration: Jensen Huang’s paranoia, embracing heterogeneity (CPX, GPUs, Groq), and why moats are shallower when everyone has $200B

• Where the real bottleneck is: from CoWoS shortages (2023) to data centers and power (2024–25) back to semiconductor fabs (2026+)

• Energy improvisation: diesel generators, reciprocating engines, breaking grid constraints and why fabs, not data centers, now bottleneck growth

• Capital vs. labor in the AI era: trillions in developer wages, automation of knowledge work, and whether GDP growth will accrue to people or infrastructure

• Fried rice technique: velveting chicken, day-old rice, wok envy on induction, forgetting the soy sauce, and the eternal MSG debate

Show Notes:

• Transistor Radio Episode 43

• Claude Code is the Inflection Point (SemiAnalysis article)

• “This is crazy”: Anthropic CEO blasts AI chip sales to China (Axios)

• Artificial Analysis Coding Index

• “Only the Paranoid Survive” by Andy Grove

—

Dylan Patel

• LinkedIn: https://www.linkedin.com/in/dylanpatelsa

• X: https://x.com/dylan522p

SemiAnalysis

• https://semianalysis.com/

• https://www.linkedin.com/company/semianalysis/

00:00 – Intro
00:16 – Guest Introduction
01:16 – Guessing the Dish
01:46 – From Beekeeper to Semiconductor Powerhouse
03:08 – Starting the SemiAnalysis Blog
05:37 – Part 1: Cooking
05:45 – Velveting Chicken & Talking Taiwan/TSMC Endgames
06:52 – China, Taiwan & Semiconductor Geopolitics
10:57 – AI Talent, Export Controls & U.S. China Tensions
18:19 – Is AI a Bubble? Hyperscaler CapEx Explosion
22:26 – Claude Code, GitHub Commits, & AI Adoption Acceleration
24:54 – Why Markets Hate $200B AI Spending
30:26 – The Hyperscaler Innovator’s Dilemma
38:49 – Who Wins the Chip War? Nvidia vs Vertical Integration
41:52 – Jensen Huang & The Paranoid Founder Advantage
45:32 – What’s the Real Bottleneck in AI Progress?
49:01 – The Semiconductor Fab Constraint
50:55 – Part 2: Tasting
52:01 – Slop vs Technique: Flavor Philosophy Debate
53:02 – Hiring at SemiAnalysis & AI Infrastructure Alpha
54:24 – Final Verdict: Whose Fried Rice Wins?

[00:00:00] Dylan Patel: I’m not crying because of the, because of the AI [00:00:03] researchers leaving. I’m crying because Run. I promise. I swear to God, if [00:00:06] Uncle Roger finds this video, I’m gonna cry. It’s about to be [00:00:09] sweeter than Pan to express. [00:00:12] Wait, what do you vote? What do you vote? [00:00:15]

[00:00:15] Allen Park: Hey guys. Welcome to In-Context Cooking, where we [00:00:18] take one dish, taste it, and try to recreate it with [00:00:21] minimal health.

Today we have a very special guest. [00:00:24] The founder and CEO of semi analysis is Dylan [00:00:27] Patel. Welcome, Dylan.

[00:00:27] Dylan Patel: Hello. Thanks for having me.

[00:00:29] Allen Park: Yeah, thank you for being here. [00:00:30] I guess the question to start off is how would you rate [00:00:33] yourself on a scale of one to 10 of one being awful? 10. 10. 10. [00:00:36] 10. Okay. So you’re an amazing cook.

So look, [00:00:39] as a cook, how would you rate yourself on a [00:00:42] scale of one to 10?

[00:00:42] Dylan Patel: Uh, I’d probably like. Five or [00:00:45] six.

[00:00:45] Allen Park: Okay. Five or six is not bad. I feel like you

[00:00:47] Dylan Patel: have maybe three. Maybe three. Lemme, lemme [00:00:48] revise.

[00:00:49] Allen Park: Okay. So low expectations and then [00:00:51] overdelivering, you’re gonna overdeliver, the presentation

[00:00:53] Dylan Patel: is great, just not

[00:00:53] Allen Park: the

[00:00:53] Dylan Patel: [00:00:54] plate.

[00:00:54] Allen Park: Okay. Okay. Honestly, like we will work with three [00:00:57] or five, whatever, but yeah, I guess in front of us, [00:01:00] we’ll, we have a bunch of ingredients. Do you have an idea of what we’re gonna make? I feel like it’s kind [00:01:03] of. Just looking off of it.

[00:01:04] Dylan Patel: Eggs, [00:01:06] eggs, rice, chicken. This is like

[00:01:08] Allen Park: very,

[00:01:08] Dylan Patel: [00:01:09] this is like very like, yeah.

Anything, [00:01:12] anything. But then the right side, you know, like the peas really throw me off.

[00:01:14] Allen Park: [00:01:15] Yeah.

[00:01:15] Dylan Patel: The ginger though. So I and soy sauce. Okay, so this is like fried rice.

[00:01:17] Allen Park: [00:01:18] Okay. Yeah. So today we have [00:01:21] chicken fried rice, which will be recreating, [00:01:24] uh, this is restaurant chicken fried rice. So we’ll try to [00:01:27] recreate it as close as possible.

Mm-hmm. Based off tasting it. So here’s a [00:01:30] spoon for you. Cheers.

[00:01:31] Dylan Patel: Cheers. [00:01:33]

[00:01:34] Allen Park: That was, that was very [00:01:36] good.

[00:01:36] Dylan Patel: I’m shocked it’s still warm.

[00:01:37] Allen Park: Yeah.

[00:01:37] Dylan Patel: I assume this has been sitting here for an [00:01:39] hour.

[00:01:39] Allen Park: Yeah, we nuked it a lot and then kept it [00:01:42] as is. So, yeah. Just to get started, do you wanna introduce yourself? I know [00:01:45] that, um, you went to Minnesota briefly [00:01:48] after, and was it, you were a beekeeper for two years.

You [00:01:51] did a lot of these kind, like side quests. You have like a lot of, [00:01:54] but now you’re, you know, the leading voice [00:01:57] on, um, chips and everything that, you know, whether it be hedge [00:02:00] funds or even people in ai.

[00:02:01] Dylan Patel: So I did live in Minnesota briefly after [00:02:03] college. Um, I’m from rural Georgia. [00:02:06] Um, yeah, we, uh, I did [00:02:09] beekeep for like a, a year and a half basically.

I feel like I’ve just [00:02:12] kind of, sort of went through a lot of life. Uh, just next step, next [00:02:15] step, next step. Doesn’t seem like there’s a clear, immediate path.

[00:02:17] Allen Park: Yeah. [00:02:18]

[00:02:18] Dylan Patel: Uh, looking back, I can spin a narrative like, oh, obviously I [00:02:21] would be doing this. Yeah. Because my interest when I was eight [00:02:24] was this, and my interest when I was 12 was this.

Mm-hmm. But, uh, you know, [00:02:27] like moderating forms related to chips. Yeah. But, you know, I thought it was just like a [00:02:30] serendipitous thing, you know?

[00:02:30] Allen Park: Yeah.

[00:02:31] Dylan Patel: Um, but then [00:02:33] eventually sort of everything culminated. And like [00:02:36] blogging and doing consulting and doing research, [00:02:39] being interested in AI and data science, being interested in chips, and then it all sort of like [00:02:42] culminated in like, oh my God, everything blew up all together.

And, [00:02:45] and so I guess right time, right moment.

[00:02:47] Allen Park: Yeah.

[00:02:47] Dylan Patel: Maybe, [00:02:48] maybe some foresight to have your passion be. The thing that [00:02:51]everyone cares about now.

[00:02:51] Allen Park: Yeah. No, I think it turned out very great. You [00:02:54] know, doing amazing. And rumor has it, [00:02:57] you started your substack because you read Doug’s [00:03:00] and thought you could do a lot better, and he was like, Hey, you should start a substack.

Is that [00:03:03] true or is there more behind that story? [00:03:06]

[00:03:06] Dylan Patel: Yeah, so what, what happened was I, um, I had an [00:03:09] anonymous blog

[00:03:09] swyx: mm-hmm.

[00:03:10] Dylan Patel: On the internet for many years. [00:03:12] Yeah. Uh, I was moderating Reddit and all these things anonymously mm-hmm. [00:03:15] For around like hardware, Nvidia and tele and v this kind of stuff. [00:03:18]And I was posting all this stuff.

And I had an [00:03:21] anonymous Twitter in the sil in silicon Twitter, right? Yeah. Which people on like [00:03:24]teapot and like tech Twitter, don’t understand.

[00:03:25] Allen Park: Yeah.

[00:03:26] Dylan Patel: [00:03:27] Um, and so I was doing this anonymously, [00:03:30] Doug, Doug started posting and I was like, well, this is [00:03:33] interesting, but like, I think I could do it a lot better. And then he is like, dude, why are you [00:03:36] like, why are you like posting on WordPress, right?

Mm-hmm. Like, do it, do [00:03:39] it non anonymously. Do it on Substack and start charging for it. So then, [00:03:42]but then I was like, too, I was like, hi. I was like, I’m not gonna charge for this. I’m so [00:03:45] great. And then, and then one day I was like, you know what? Screw this. I’m gonna start [00:03:48] charging for it. ‘cause Doug told me too many, many times.

Yeah. Um, and [00:03:51] then a few years go by and then Doug, Doug joins the company. [00:03:54] Um, so it was, it was really a great moment that, yeah, full circle. He told [00:03:57] me to, uh, do it because otherwise it might still just be [00:04:00] like a, a niche anonymous blog where I’m still just doing random [00:04:03] consulting rather than actually, uh, you know, a company with 60 [00:04:06] people.

[00:04:06] Allen Park: Gotcha. Okay. So at that time you were still doing consulting, right? [00:04:09] And you just had a separate blog, A thousand million. Well,

[00:04:10] Dylan Patel: it was consulting related to the [00:04:12] blog in the industry.

[00:04:13] Allen Park: Gotcha. Okay. Yeah, that’s amazing. I [00:04:15] guess now we could kind of look at the ingredients and maybe try [00:04:18] little, so here we have ginger, [00:04:21] garlic, carrots, cheese.

We’re gonna, we’re just gonna eat

[00:04:22] Dylan Patel: the whole garlic. [00:04:24]

[00:04:24] Allen Park: Yeah. We could all take a piece. And then this I think [00:04:27] is, and we also have tasting spoons of refund. Try. Um, I think this [00:04:30] is sugar. Yeah.[00:04:33]

So

[00:04:35] Dylan Patel: sweet. [00:04:36] Okay. Let’s see. I really hope this is a salt.

[00:04:38] Allen Park: Okay. It’s [00:04:39] sugar. Yeah. So this is sugar. And we’ll know ‘cause [00:04:42] everything will be laid out for you exactly like this. Um, [00:04:45] so we’ll have to eyeball and we’ll be going fast. This, I don’t think we need taste. [00:04:48] I’m pretty sure this is MSG baking powder.

[00:04:50] swyx: Oh

[00:04:50] Allen Park: [00:04:51] yeah.

So I’m What kind

[00:04:51] Dylan Patel: of, what kind of fried rice does have?

[00:04:53] swyx: M mss [00:04:54]

[00:04:54] Allen Park: g Yeah.

[00:04:54] swyx: Oh yeah,

[00:04:54] Allen Park: that’s, that’s true. We YouTuber [00:04:57] Uncle Roger. Yes. But yeah, we sauce here [00:05:00]baking powder, sugar, onion. Of course. I think this is

[00:05:02] Dylan Patel: one, [00:05:03] if, if he discovers this, he’s gonna roast. The fuck.

[00:05:04] Allen Park: Yeah, he will. But hopefully, [00:05:06] you know, he will give us some grace, [00:05:09] but Okay.

And then eggs. Salt. Yeah.

[00:05:11] Dylan Patel: Wait, is that [00:05:12] the tactic? Intentionally don’t put m ms. G in your rice.

[00:05:14] Allen Park: [00:05:15] Yeah.

[00:05:15] Dylan Patel: So rage bait Uncle Rogers.

[00:05:16] Allen Park: Yeah. This is also short green rice, which [00:05:18] is also a red flag, but we just have it. And this is [00:05:21] day old. So this is the important [00:05:24] ingredient to have the day old, not fresh.

[00:05:25] Dylan Patel: Why does that matter?

[00:05:25] Allen Park: Um, it’s a little [00:05:27] drier and it’ll make the fried rice, just like the [00:05:30] grains be a little more separated compared to being a motion mess. But [00:05:33] yeah. I think we’re good. Are you ready to show off your cooking skills?

[00:05:35] Dylan Patel: [00:05:36] Sure.

[00:05:36] Allen Park: Great. Let’s get started. [00:05:39] Dylan, are you ready to make some [00:05:42] fried rice?

[00:05:43] Dylan Patel: Absolutely.

[00:05:44] Allen Park: Okay. So the first [00:05:45] thing, let’s grab the chicken and then we just need to [00:05:48]marinate it, so, or velvet it rather.

So take the [00:05:51] cornstarch, it’s one of the white powders that we saw [00:05:54] before, and the baking powder. And then [00:05:57] add like a spoonful, maybe half a spoon of corn [00:06:00] starch, um, to the chicken. [00:06:03]And then just a hint of the baking powder. [00:06:06] Um, and so [00:06:09] yeah, once we add that

[00:06:11] Dylan Patel: a pinch or what [00:06:12] are we talking?

[00:06:12] Allen Park: I just did like, yeah, a pinch or like a very [00:06:15] small amount with my spoon.

And then we can add a [00:06:18] little bit of salt and soy [00:06:21] sauce. I think this is salt. Make sure it’s salt [00:06:24] and not sugar.

[00:06:24] Dylan Patel: Yeah.

[00:06:25] Allen Park: Okay. There’s a little bit of salt [00:06:27] and then soy sauce. And so what we [00:06:30] wanna do is we wanna mix it kind of vigorously. [00:06:33] So it just looks like

[00:06:34] Dylan Patel: you said salt and soy sauce,

[00:06:35] Allen Park: olive, [00:06:36] yeah. Um, just like a little bit of each to give it some [00:06:39] flavor because we’re gonna marinate it and yeah, we just wanna [00:06:42] mix it vigorously so that it looks like a slop of chicken.[00:06:45]

But I [00:06:48] guess one question on my mind that you’ve also discussed about [00:06:51] is end game scenarios with araki, [00:06:54] or in this case Taiwan to N-T-S-M-C, I guess, like [00:06:57] have you thought about what that would look like? [00:07:00]

[00:07:00] Dylan Patel: Bro wants to make fried rice and talk about China [00:07:03] politics. Let’s go

[00:07:04] Allen Park: indeed. Very on [00:07:06] topic.

[00:07:06] Dylan Patel: It’s vaguely racist.

[00:07:08] Allen Park: [00:07:09] Just, just vaguely, just vaguely,

[00:07:10] Dylan Patel: vaguely racist. Yeah. [00:07:12] Um, okay, so, so endgame scenarios are, are kind of [00:07:15]insane, right? Mm-hmm. Like, um, there, there’s, there’s a [00:07:18] variety of things that one could, uh, that could happen.

[00:07:20] Allen Park: Yeah.

[00:07:20] Dylan Patel: [00:07:21] Um,

[00:07:21] Allen Park: such as,

[00:07:22] Dylan Patel: so, so in, [00:07:24] in, in some cases, right? It’s like. You know, status quo is [00:07:27] the best, right?

Mm-hmm. Um, you know, no war [00:07:30] happens. I’m done with the, uh, chicken, by the way. Okay. No war happens [00:07:33]

[00:07:33] Allen Park: and just start cutting the onion to like small dices, but continue. [00:07:36]

[00:07:36] Dylan Patel: We have no war, no invasion. There’s no blockade. [00:07:39] Yeah. Sort of status quo. China continues to, China [00:07:42] continues to, you know, industrialize itself with [00:07:45] the industry that Taiwan has, but there’s no, there’s no major [00:07:48] event that occurs.

[00:07:49] Allen Park: Mm-hmm.

[00:07:49] Dylan Patel: You know, that’s sort of [00:07:51] option one, option two, which is what a lot of people. Seem [00:07:54] to, like, think is the best, which is Taiwan actually, like [00:07:57] more stringently claims. It’s independent. [00:08:00] Yeah. At least a lot of like westerners and, and then that seems like the, like [00:08:03] actually a poor option. Um, just because that, [00:08:06] that, uh, potentially causes China to, uh, move, [00:08:09] move much more, um, aggressively on Taiwan.

Mm-hmm. [00:08:12] And sort of like, there’s like another set of options which is like, okay, well [00:08:15] what if, what if China moves close or Taiwan moves closer to China, [00:08:18] um, in a political way. Right? And so like, that’s an [00:08:21] example of that is like, hey, like they elect. The KMT. [00:08:24] Mm-hmm. Um, and the KMT ends up, [00:08:27] uh, winning. So there’s two parties in Taiwan, the D-P-P-K-M-T.

Yeah. [00:08:30] Um, if the DDPP [00:08:33] wins, they’re sort of, they cont they have won for the last decade and they’re sort of [00:08:36] more pro us, more anti-China, more, more [00:08:39] Taiwan independence. Okay. Um, and then there’s sort of like the [00:08:42] last scenario, which is just a full on invasion.

[00:08:43] swyx: Yeah.

[00:08:44] Dylan Patel: Um, [00:08:45] or, or at least a, a, a political [00:08:48] coup or takeover of some sort.

And so there’s like a [00:08:51] variety of options possible. Um, what I think is most [00:08:54] likely, um, is that there is some sort [00:08:57] of political coup or action mm-hmm. [00:09:00] That, um, destabilizes Taiwan in some way.

[00:09:02] swyx: Mm-hmm. [00:09:03]

[00:09:03] Dylan Patel: Um, but doesn’t act doesn’t necessarily [00:09:06] lead to an invasion, a full scale, um, [00:09:09] invasion. And so sort of, this is the best of the both worlds for at least [00:09:12] China.

Right. Like they don’t have to actually enter an [00:09:15] invasion. Yeah. Uh, but they get to continually creep more and more on Taiwan [00:09:18] without actually having to deal with the repercussions of a war. [00:09:21] And subsequent blockades and so on and so forth.

[00:09:23] Allen Park: Mm. [00:09:24]

[00:09:24] Dylan Patel: The like kind of galaxy brain thing for an American [00:09:27] to want is for actually the Taiwanese, um, [00:09:30] government, uh, the pro-US party to lose.[00:09:33]

Right.

[00:09:33] Allen Park: Oh, interesting.

[00:09:34] Dylan Patel: Um, so you want KMT to [00:09:36] win? Yeah. Um, and, and if KMT wins, then that means China will [00:09:39] be placated more. Yeah. And you don’t have Taiwan [00:09:42] fully move into China’s orbit.

[00:09:44] Allen Park: Mm-hmm.

[00:09:44] Dylan Patel: [00:09:45] Um, but you do have the government sort of placated in [00:09:48] China. Yeah. Um, at the same time, uh, even if, even if the [00:09:51] KMT wins, it’s not like TSMC starts disobeying [00:09:54] American export restrictions because mm-hmm.

The way American export restrictions are [00:09:57] upheld is that Taiwan utilizes American banking [00:10:00] systems. American equipment industries. Yeah. And so they’ll [00:10:03] still have to uphold any US export restrictions. So sort of like, you [00:10:06] know, China’s placated by the fact that they have a friendly government in [00:10:09] power, and yet China doesn’t actually have any of the [00:10:12] chips, and the US continues to get to access the chips,

[00:10:14] Allen Park: I guess, for [00:10:15] export controls as well.

So. I know that right now [00:10:18] there’s like stances, for example, Dario right is very [00:10:21] anti, uh, China having access to things. [00:10:24] Um, and I think supposedly one of the top [00:10:27] researchers, SHNU also left because of that. There could [00:10:30] be maybe a lot

[00:10:30] Dylan Patel: of, he left, he left, uh, inro. He

[00:10:32] Allen Park: left one [00:10:33] of the top labs.

[00:10:33] Dylan Patel: Got it. Got it.

[00:10:34] Allen Park: And I guess, do you have opinions on [00:10:36] like, if the US takes how

[00:10:36] Dylan Patel: much of the dice, by the

way?

[00:10:38] Allen Park: You could just cut [00:10:39] a dice like half and then after that just start cutting the [00:10:42]carrots to like a similar size. Do you have any thoughts [00:10:45] on, you know, the danger of maybe a lot of Chinese [00:10:48] talent fleeing due to the very staunchly, [00:10:51] anti-Chinese stance that a lot of these air labs or top [00:10:54] figures may take?

Or do you think this is a, like a non-issue? [00:10:57]

[00:10:57] Dylan Patel: You know, um, I’m, I’m not crying because of [00:11:00] the, because of the AI researchers leaving. I’m crying on dead, I [00:11:03] promise. No, but it would be, it’d be a travesty, right? [00:11:06] If like a lot of like Chinese researchers left. Yeah. Um, [00:11:09] American labs. You know what, I think it’s probably like, like.

Half [00:11:12] or at least a third of researchers [00:11:15] at labs are Chinese.

[00:11:16] Allen Park: Mm.

[00:11:16] Dylan Patel: Um, so obviously there’s [00:11:18] like a level of like, you know, why, why are you anta Don’t antagonize [00:11:21] too much.

[00:11:21] Allen Park: Yeah.

[00:11:22] Dylan Patel: Um, at the same time, [00:11:24] there is a level of like, you know, this is, [00:11:27] this is the greatest, uh, technology to ever, [00:11:30] uh, fall into humanity’s hands.

Mm-hmm. [00:11:33] Um, you know, obviously we think we’re the good guys and by we, I [00:11:36] mean Americans. Um, and so Americans think that they should [00:11:39] control it and maybe, maybe Anthropic thinks they’re the good guys, not [00:11:42] Americans as a whole, um, but whatever the, whatever the [00:11:45] moral justification is, like, you know, whether it’s, [00:11:48] Hey, we’re the good guys and AI’s gonna be super powerful, or, and we’re [00:11:51] the only ones who can steward it, or, Hey, AI would be a great weapon, so let’s [00:11:54] make sure we’re the only one with that weapon.

Um, there’s [00:11:57] certainly some level of control that needs to be had.

[00:11:59] swyx: Mm-hmm.

[00:11:59] Dylan Patel: Right [00:12:00] now the question is like, where do you, where does that control, like start [00:12:03] and stop, right? Mm-hmm. Um, ‘cause one could say, okay, well let’s just [00:12:06] control the chips. Um, uh, or let’s just [00:12:09] control the ai. But then it’s like, okay, well then you let them buy the chips and [00:12:12] they’re able to, they’re able to do everything they want to [00:12:15] do anyways.

[00:12:15] swyx: Yeah.

[00:12:16] Dylan Patel: Right. And this is, this has existed across many [00:12:18] technologies, right. Um, you know, China, China has great [00:12:21] engineering. You only remove one piece of the puzzle. They’re able to [00:12:24] re-engineer that last. A bit of the puzzle, you know, that’s, that’s [00:12:27] one context, right? And so, so that, that is like some [00:12:30] people’s argument, right?

Mm-hmm. So for example, if you look at [00:12:33] Nvidia or you look at like David Sachs, I think [00:12:36] their argument is like, you know, hey, um, you know, [00:12:39] we should not let them have the models, but we should let them have the chips and everything [00:12:42] else because then, uh, they’re still relying on American [00:12:45]ecosystem, American talent, American, um, [00:12:48] you know, technology.

Yeah. American platforms, [00:12:51] event resources. The other argument is like, Hey, like look, this stuff, [00:12:54]um. You know, if, if they don’t have access to our [00:12:57] models, but they have access to our chips, that alone gives them, [00:13:00] um, shoot. [00:13:03] Um,

[00:13:03] Allen Park: yeah.

[00:13:03] Dylan Patel: That alone gives them all the control and [00:13:06] capabilities they need, uh, to, [00:13:09] to basically, um, be on [00:13:12] par or just slightly behind.

Mm-hmm. Right. And so we see that in the current regime [00:13:15] today, right? Yeah. China has effectively. [00:13:18] Uh, great access to chips. Um, not as much, [00:13:21] not completely unfiltered access. Right. They either have to [00:13:24] rent it mm-hmm. Or they have to smuggle it, or there’s some chips that are [00:13:27] allowed. Um, and, and China’s not that far behind, right?

You look [00:13:30] at like Kim k, k two, five, uh, agent [00:13:33] swarms, and it’s like, well, this is like not [00:13:36]worse than Codex by a marginal amount. Maybe 5.3, right? Yeah. But [00:13:39] 5.2, it’s like not worse than Codex by a mar, like a [00:13:42] large amount.

[00:13:42] Allen Park: Mm-hmm.

[00:13:43] Dylan Patel: Um. And so there [00:13:45] is, there is like an argument to be made that like, hey, current [00:13:48] regime ha does not have the US leading in AI by enough.[00:13:51]

[00:13:51] Allen Park: Mm.

[00:13:51] Dylan Patel: Right?

[00:13:51] Allen Park: Mm-hmm.

[00:13:52] Dylan Patel: Um, and, and that’s sort of the argument that [00:13:54] like Ario would make, I think, is that look like current [00:13:57] regime of export controls, China is, is still way [00:14:00] caught up. Uh, China is still, [00:14:03] um, you know, not that far behind. Yeah. Um, and it would be [00:14:06] disastrous if they are, um, in, in, [00:14:09] in, in that, in their eyes, right?

Mm-hmm. And so the question is how does [00:14:12] one, how does one, um. Write that circle. [00:14:15] Right? So one could either a antagonize China [00:14:18] more. Yeah. Ban more things. Um, does that risk a [00:14:21] Taiwan evasion or does that, um, does that lead [00:14:24] to something, say, say catastrophic? Does [00:14:27] that, does that alienate researchers in, in America?

Um, that’s [00:14:30] one argument. Another one is like, well look, let’s, is like, we’ve got like two years, right? [00:14:33] You know, AI 27 bros. Right? Um, and I’m not [00:14:36] fully there, right, in terms of AI 27, but I’m pretty bullish on AI [00:14:39] generally. Yeah. And so then the other agreement is like, look, we only have a few years. [00:14:42] Until the, um, capabilities of [00:14:45] AI accelerate GDP growth.

And it’s, it’s, by the [00:14:48] way, am I cooking something? Am I cutting something else?

[00:14:49] Allen Park: Um, yeah. [00:14:51] Okay.

[00:14:51] Dylan Patel: Oh, you did mash sticks, huh?

[00:14:52] Allen Park: So you, you [00:14:54] can start cutting. We need to mince the garlic and then the ginger, [00:14:57] um, the garlic and the ginger has to be a lot [00:15:00] finer.

[00:15:00] Dylan Patel: Okay.

[00:15:01] Allen Park: Um,

[00:15:02] Dylan Patel: um, I. [00:15:03]

[00:15:03] Allen Park: Yeah, continue.

[00:15:04] Dylan Patel: I felt, I felt a little awkward just [00:15:06] talking and not doing anything.

Okay. And I feel like, [00:15:09]

[00:15:09] Allen Park: so we’re just cutting all the vegetables

[00:15:10] Dylan Patel: already. Yeah. Yeah. Sounds good. [00:15:12] Sounds good. The one stance that like a lot of like, [00:15:15] um, folks have is that, you know Yeah. Like, let’s, let’s [00:15:18] be more anti-China, um, [00:15:21] ‘cause there’s only a few years left. Mm-hmm. [00:15:24] And so, like, you know, until like super powerful AI [00:15:27]systems and those super powerful AI systems will make next generation AI [00:15:30] systems, right?

Mm-hmm. And that’s like. That’s like a lot of [00:15:33] the, um, argument that like, let’s say the [00:15:36] Dario’s of the world would make. Yeah. Um, and I’m sympathetic to that [00:15:39] argument to some degree as well. Right. Okay. Um, if we start looking at [00:15:42] like, Hey, um, this [00:15:45] year Google’s spending two, uh, Amazon’s spending $200 [00:15:48] billion.

Google’s spending $180 billion. On, on [00:15:51] AI infrastructure primarily. Right?

[00:15:52] Allen Park: Yeah.

[00:15:52] Dylan Patel: Um, this is, [00:15:54] you know, four x what they were doing just not too long ago.

[00:15:56] Allen Park: Mm-hmm. [00:15:57]

[00:15:57] Dylan Patel: Um, and if they get returns on that of any degree, then we’re talking [00:16:00] about trillions of dollars of economic value Yeah. Being added, [00:16:03] um, in, in just the next, uh, [00:16:06] handful of years.

[00:16:06] Allen Park: Yeah.

[00:16:07] Dylan Patel: And so, you know, the, the, the [00:16:09] risk here is that like, okay, um, whatever AI [00:16:12] is capable of doing, it’s obviously adding hugely to, to the [00:16:15] economy and. You wanna do [00:16:18]anything and everything you can to slow China down. We’ve never had an exp explos [00:16:21] explosive growth like this. Mm-hmm. We’re on the cusp of it.

Right, right. Right [00:16:24] now we only have like, you know. Um, you know, the AI industry [00:16:27] maybe does 50 bill of revenue. Mm-hmm. Um, you know, [00:16:30] across and, but like, we’re seeing it explode. Right. [00:16:33]Philanthropics, you know, adding two, 3 billion of revenue a month now. Mm-hmm. Versus they were [00:16:36] just adding a few hundred million of revenue a month earlier.

Um, so [00:16:39] clearly we’re in the takeoff period. Yeah. Um, and so, [00:16:42] so the argument there is that like, let, let’s just limit them [00:16:45] completely.

[00:16:45] swyx: Mm-hmm.

[00:16:45] Dylan Patel: Um. So they don’t end up [00:16:48] with, um, all the, all these super powerful AI systems. [00:16:51] I, I, I, I’m, I’m, I’m a little bit like, you know, [00:16:54] it’s, it’s hard to rationalize, you know, in every specific like [00:16:57] argument just because like you.

You know, I could, I could, [00:17:00] I could argue, or I could steelman, uh, any of the arguments. Right? [00:17:03]Yeah. Actually. Okay. No. Maybe, maybe, maybe China [00:17:06] is, um, should be sold our chips because at [00:17:09]the end of the day, they’re still relying on our chips now for their AI [00:17:12] systems.

[00:17:12] Allen Park: Mm-hmm.

[00:17:12] Dylan Patel: Um, and, and then they have less incentive to invade [00:17:15] Taiwan.

Yeah. And if they invade Taiwan, then, you know, all of a sudden, [00:17:18] um, the whole party stops and we can’t do anything. And when you think [00:17:21] about what’s the capability of, uh, China versus [00:17:24] the US to, um. [00:17:27] To, to have a, a vertical supply chain [00:17:30] in, um, chips or AI or in [00:17:33] really anything. Mm-hmm. Uh, the China, China has by far the most [00:17:36] advanced, uh, supply chain in semiconductors if you [00:17:39] just look at China itself.

Mm-hmm. Right. In which case, like [00:17:42] if the, the lifeblood of AI is compute to some extent, [00:17:45] uh, to a large extent, then China would win if we didn’t [00:17:48] have Taiwan. Right?

[00:17:49] Allen Park: Yeah.

[00:17:49] Dylan Patel: And maybe that timescale would be way [00:17:51] longer because Taiwan is so far ahead in the production capacity. [00:17:54] And China doesn’t have the equipment ecosystem like the rest of the [00:17:57] world does.

Um, but at the end of the day, that’s, that’s [00:18:00] exactly like sort of the argument that that one would make [00:18:03] is his hay. Like, you know, if you push China too [00:18:06] far, now they might invade, um, [00:18:09] they might invade Taiwan and that ends up with. [00:18:12] With this catastrophic scenario.

[00:18:13] Allen Park: Yeah. Double clicking on something you [00:18:15] mentioned already with these big AI labs and even [00:18:18] hyperscalers, somewhat overextending on, [00:18:21] um, future spend with these data centers.[00:18:24]

Would there ever be something that would prompt [00:18:27] you to get a little concerned? ‘cause it seems right now we’re in a very, [00:18:30] um, acceleration type of moment for AI where [00:18:33] still adoption for what even like cloud code and a lot of [00:18:36] these great tools, um, isn’t as mainstream, but. Are [00:18:39] there things that if you see happening in the [00:18:42] next, what, six months or year that would kind of [00:18:45] concern you in terms of where the future lies [00:18:48] of these companies maybe overextending a little too much and [00:18:51] whether concerns of being in a bubble may [00:18:54] actually have some merit to it?

[00:18:55] Dylan Patel: I mean, I think, I think sort of, not [00:18:57] answering your question, but talking about something else I wanna talk about Yeah. Which I’d love to do. [00:19:00] Um,

[00:19:01] Allen Park: that’s fair.

[00:19:01] Dylan Patel: Is, is sort of like you mentioned [00:19:03] like these doomsdayers and these believers. I think the [00:19:06] biggest risk is actually just like the general public hates ai.

[00:19:08] Allen Park: Mm.

[00:19:08] Dylan Patel: [00:19:09] Right. Um, you know, I think, I think if you go literally [00:19:12] anywhere, the general public absolutely [00:19:15] hates AI so much. They, they have [00:19:18] literally no, um, [00:19:21] they don’t, yeah. They, I mean, like, you know, you go to your [00:19:24] random artist and they like hate ai. You go to your random, like person in [00:19:27] rural America, they’re like, screw ai.

It’s like, you know, taking [00:19:30] all the water. Um, you know, completely like nonsense [00:19:33] arguments. Yeah. But like, it doesn’t matter. [00:19:36] Um, so when you look across the ecosystem, you’ve got that [00:19:39] problem.

[00:19:39] swyx: Mm-hmm.

[00:19:40] Dylan Patel: Um, and so like, and then you’ve got the [00:19:42] doomsdayers, right? There’s also the general public, which like doesn’t quite understand [00:19:45] ai or maybe they do, um, you know, and, [00:19:48] and, and they’re just like so worried about ai, like taking, taking her [00:19:51] gers and like all sorts of other things.

Right. Um, [00:19:54] and so taking Gers, I think, I think that’s, that’s [00:19:57] another aspect of this that’s like quite interesting is, um. [00:20:00] You know, general public hates it. Um, yeah. The [00:20:03] doomsdayers, you know, look, I’m, I’m, I’m, I’m a, I’m a bit of a [00:20:06] to, uh, live in the moment person to like, [00:20:09] think, you know, Hey, what exactly is alignment mean?

And, [00:20:12] and does AI kill us all? Um, you know, [00:20:15] obviously there’s huge risk to that. Mm-hmm. Uh, that’s not what I’m [00:20:18] an expert in, so I don’t, I don’t really care to opine too [00:20:21] much. Um, but, [00:20:24]um. A as as far as like, you know, is is it a [00:20:27] bubble? Are people, are people, are we, are we doing too much? Mm-hmm. Like, [00:20:30] what’s going on?

Are you good? [00:20:33] Um, is there a bubble? Are we doing too much? Um, [00:20:36] you know, the question is, um, [00:20:39] you know, if, if AI model progress slows down, then of [00:20:42] course we’re in a bubble. Yeah. That’s, that’s obvious. Um, [00:20:45] but you know, the, the, the way that progress is accelerating, [00:20:48] uh, so fast, right? Um, you know, you, you, you [00:20:51] see it, uh, month on, month on month, right?

I mean, you [00:20:54] know, the new models are coming out every month, every week, almost [00:20:57] it feels like nowadays. Yeah. Um, or new capabilities, right? Just think [00:21:00] about like, like this year so far, right? [00:21:03]Um, you know, obviously Claude 4.5 came out and Claude [00:21:06] Code came out, uh, last year. [00:21:09] But adoption really uptick in the beginning of this year.

[00:21:12] Um, and so we saw a huge uptick, uh, just in this [00:21:15] month, uh, just in January. It went from, uh, 4% [00:21:18] of um, or 2% of, uh, [00:21:21] commits on GitHub to 4% of GitHub commits were done by [00:21:24] cloud code. Yeah, right. That does not mean, you know, a lot of people use cloud code without [00:21:27] having cloud commit for them. So you’re like, you’re talking about [00:21:30] like, or, and then they use Codex, they use cognition, they use, you know, [00:21:33] or sorry, Devon, they use all these other platforms.

They use GitHub, [00:21:36] copilot, they use Codex. You know, we’re probably at 10% or so of total [00:21:39] code as being committed, um, or written by ai, if not [00:21:42] more.

[00:21:42] swyx: Mm-hmm.

[00:21:42] Dylan Patel: Um, but at least for Claude code itself, it went from two to [00:21:45] 4% in one month.

[00:21:46] swyx: Mm-hmm.

[00:21:46] Dylan Patel: Right. Um, [00:21:48] this, this acceleration is, is like. [00:21:51] I think it’s the thing that we’ve all been waiting for.

Yeah. Um, ‘cause [00:21:54] like find people who are using chat GPT. That’s great. Find people who are [00:21:57] using like image gen. That’s great. These were not like things that [00:22:00] add trillions of dollars to the economy. These are social networks. These [00:22:03] are like, you know, he’ll help me on my, uh, homework. These are like, [00:22:06] chill things.

[00:22:06] Allen Park: Yeah.

[00:22:07] Dylan Patel: Um, but now we’re, we’re in the [00:22:09] stage where it’s like, no, no, no. Like these are trillions of dollars of, [00:22:12] uh, economic value that could be added. Um, and if [00:22:15] you think about worldwide software developer wages, $2 trillion in [00:22:18] wages, you end up with, uh, a pretty [00:22:21] incredible amount of, um, [00:22:24] spend that could happen.

Um, I feel like it’s my [00:22:27] worst interview ever because I’m, I’m, I’m like so focused on [00:22:30] chopping garlic and not cutting up off my fingers. Um. [00:22:33] I wonder if people are gonna like, trash me for my [00:22:36] lack of claw like grip. Because like I’ve tried to [00:22:39] do it, but then I’ve failed and I’ve done like more dangerous cutting [00:22:42] grips.

[00:22:42] Allen Park: Mm-hmm.

[00:22:43] Dylan Patel: I, I do wonder how, how, [00:22:45] how vicious is, is the audience.

[00:22:47] Allen Park: We’ll see. [00:22:48]

[00:22:48] Dylan Patel: Um,

[00:22:49] Allen Park: but yeah, you’re talking about claw coded [00:22:51] option and GitHub commits to being significantly more, especially [00:22:54] this, uh, past like couple months [00:22:57] and like the, uh, [00:23:00] value that cloud code will have towards the economy. Do you wanna [00:23:03] elaborate more on that or was that a finished

[00:23:04] Dylan Patel: Yeah, yeah.

I mean, I think, I think [00:23:06] like there’s, there’s quite a bit of, [00:23:09] um, I’m definitely claw gripping now. You know, [00:23:12] this is, I gotta make up for it now that I’m cognizant. [00:23:15] Um, I. [00:23:18] So I think, I think, um, the adoption [00:23:21] of AI has been so accelerated over the last [00:23:24] month. Yeah. You know, just, just think about everything that’s happened in the last month.

We’ve had [00:23:27] Claude Code happen, we had Claude Bot, we had, uh, malt [00:23:30] book. Uh, now we have, uh, you know, we had Kim K 2.5 [00:23:33] swarms. We had Codex, um, [00:23:36] 5.3, which is, you know, a significant step up as well. [00:23:39] Um, and seems better in some specific areas. [00:23:42] And, and these are, these are, there’s so many areas, right? Mm-hmm. And it’s like, [00:23:45] um, at least internally, like at my company, we’ve, we’ve [00:23:48] completely, um, flipped over, right?

Like about, about a third [00:23:51] of the company’s engineers, about a third of the company’s hedge fund people, and about a third of the company is [00:23:54] like passionate individuals. Mm-hmm. Um, you know, and so [00:23:57] the X hedge fund people, they are, they’re all in on cloud code [00:24:00] now too, right? They scrape data, they do pro, they do financial [00:24:03] modeling, they do proforma financial modeling with cloud [00:24:06] code as the assistant.

Um, and so there’s a variety [00:24:09] of like, sort of like, um, [00:24:12] you know, I think, I think like we, we’ve hit like sort [00:24:15] of escape velocity and all these things. And so, [00:24:18] you know, over the last month or over the last two weeks, we’ve had, [00:24:21] um, you know, the hyperscalers report earnings. Mm-hmm. [00:24:24] Um, and everyone’s stocks have gone down, right?

Mm-hmm. Um, [00:24:27] Google announced 180 billion of CapEx. The stock went [00:24:30] down and then Amazon announced 200 billion of CapEx and their stock went [00:24:33] down, like, I wanna say like 10%. So the market hates it, [00:24:36] but they don’t realize, like, you know, these [00:24:39] CapEx decisions are because they see the light at [00:24:42] the end of the tunnel tunnel, if you will.

Right? Yeah. Um, the amount of, [00:24:45] the amount of adoption is just insane. So now that we’ve had these [00:24:48]companies like report earnings and they disclose what their plans are for the [00:24:51] year, and they’re much higher than almost anyone predicted. [00:24:54] Mm-hmm. Um, you, you, you’ve got, you’ve got. [00:24:57] You’ve got the market just hating it.[00:25:00]

Um, yeah. And, and, and, [00:25:03] and, and so that, that, that brings like this interesting conundrum, [00:25:06] which is that like, okay, the market is mad, they’re spending all this money [00:25:09] on compute CapEx. Yeah. But these companies know much better than, [00:25:12] than than you, right? Like, and by you I mean the [00:25:15] investor. Yeah. Right. Um, in reality, they’re spending this much because [00:25:18] they see insane amounts of demand, [00:25:21] right?

Mm-hmm. Um, you know, and Andro doesn’t just add [00:25:24] $2 billion of revenue in one month. Uh, you [00:25:27]know, with without having, you know, huge demand. [00:25:30] Mm-hmm. Um, and they’re doing it at positive margins, right?

[00:25:32] Allen Park: Yeah.

[00:25:32] Dylan Patel: [00:25:33] Um, and they’re doing it, you know, and, and, and when they go to everyone, they’re like, go look guys. [00:25:36] We need more compute.

We need more compute. We need more compute. Um, and [00:25:39] so, so it, it’s, it’s, you know, after three years [00:25:42] straight of everyone of the AI lab saying we need more compute, [00:25:45] the hyperscalers are now seeing it’s not just we need more compute. ‘cause [00:25:48] we wanna train bigger models and we wanna do more research. It’s [00:25:51] actually, we need more compute because we need, we need to serve our users, [00:25:54] right?

Mm-hmm. Um, we need to add hundreds [00:25:57] of millions of dollars or billions, actually billions of dollars of [00:26:00] compute, right? If philanthropic added two and a half billion of revenue, um, [00:26:03] and their gross margin is 40%. They added like one and a half billion dollars [00:26:06] of compute in one month. Mm-hmm. Right. To just to serve [00:26:09] that.

You extrapolate that line out a little bit and you’re like, holy [00:26:12] crap. You know, they actually need hundreds of billions of dollars of [00:26:15] compute. Um, okay, fine. Well we, we need to build this. [00:26:18] All right. ‘cause it’s a front run of, you know, you build it and then they can rent it. [00:26:21]

[00:26:21] Allen Park: Yeah.

[00:26:21] Dylan Patel: Um,

[00:26:22] Allen Park: so the bet here is [00:26:24] that growth will continue to accelerate and the amount of.

[00:26:27] Money that Anthropic and open Eye makes will just continue to go up. [00:26:30]

[00:26:30] Dylan Patel: Yeah, exactly. And, and, and, and I don’t, [00:26:33] you know, like, look, the party can stop at some point, like, you know, [00:26:36] that’s for sure. Mm-hmm. Um, at any point, you know, I think my [00:26:39]favorite thing was, uh, I, I tweet, I tweeted about like, people were [00:26:42] like, oh, who expected this CapEx?

And then I was like, [00:26:45] well, we did. Right? And then someone replies, [00:26:48] the whale watcher told you that you’re gonna see whales. Wow. [00:26:51] Surprising. I was like, wow, that’s a pretty good, uh, [00:26:54] uh, reply to me, you know? But like, anyways, like, it’s like, yeah. [00:26:57] You know, you know, obviously, obviously I’m the whale watcher here.

[00:27:00] Um mm-hmm. You know, we see the CapEx coming. [00:27:03]

[00:27:03] Allen Park: Yeah.

[00:27:03] Dylan Patel: So the market doesn’t like it, but it’s clearly obvious that it’s [00:27:06] needed. Mm-hmm. Um, and, and, and now the [00:27:09] discussion is sort of like, you know, I think a couple years ago I said the [00:27:12]hyperscalers would have no free cash flow. Yeah. Right. IE they would [00:27:15] not have any, uh, they would not be generating profits and [00:27:18] buying back their stock in a short amount of time.

[00:27:20] Allen Park: [00:27:21] You could also transfer your onions and some veg to the [00:27:24] bull white bull on your left, I believe, or right.

[00:27:26] Dylan Patel: [00:27:27] Yes. Um,

[00:27:27] Allen Park: so on my cutting board, I just have garlic, ginger in the green [00:27:30] part of the scallions. So leave the green part for a [00:27:33] garnish for later. So just keep it on. [00:27:36] Um, I guess at what point do [00:27:39] you think the market would [00:27:42] be, um, satisfied or okay with [00:27:45] this?

[00:27:45] Dylan Patel: I think, I think the market is gonna get really mad [00:27:48] at the hyperscalers. Um, they haven’t really yet. Um, [00:27:51] mm-hmm. But we saw the, we saw the signs of it at the beginning, uh, [00:27:54] about mid last year. Right.

[00:27:55] Allen Park: Okay.

[00:27:55] Dylan Patel: Um, for example, [00:27:57] Oracle peaked when they announced within a week after [00:28:00]announcing, uh, the open A, that they’re gonna do, you know, 300 plus [00:28:03] billion dollars of deals with, uh, open ai.

[00:28:05] Allen Park: Mm-hmm. [00:28:06]

[00:28:06] Dylan Patel: Um, and the market really, [00:28:09] uh, peaked, uh, around then. [00:28:12]

[00:28:12] Allen Park: Mm-hmm.

[00:28:14] Dylan Patel: Um, [00:28:15] and, and, and then like since then they’ve gone down. [00:28:18] And, and like other darlings that were like doing AI [00:28:21] infrastructure like, uh, have also peaked, right? Mm-hmm. [00:28:24] So now we’ve got this like, interesting, um, [00:28:27] conundrum where now the [00:28:30] hyperscalers are starting to say how much CapEx they’re gonna do.

[00:28:33] Yeah. Um, [00:28:36] and, and, and, and so when we, when we think about like, Hey, [00:28:39] what’s gonna, what’s gonna end up happening is. You know, these hyperscalers are [00:28:42] gonna keep spending, right? Mm-hmm. What is their biggest advantage? Right? [00:28:45] It’s, it’s, it’s that they can build the most infrastructure in the world. Mm-hmm. [00:28:48] They’ve built the organization to build, uh, infrastructure [00:28:51] faster than anyone else.

Um, [00:28:54] wow. This like, got kicked, cooked super fast.

[00:28:56] swyx: Show the [00:28:57] cameras. [00:29:00]

[00:29:00] Dylan Patel: It’s just eggs, bro. [00:29:03] Um, I just don’t generally use an, uh, induction, right? [00:29:06]

[00:29:07] Allen Park: Yeah. They heat up very fast. [00:29:09] So just as a heads up.

[00:29:10] Dylan Patel: Yeah. Um, so, so, [00:29:12] so the hyperscalers are, you know, have been the most profitable [00:29:15] companies to ever exist in the Humana, in humanity.

Mm-hmm. Right? Whether [00:29:18] it’s meta through ads, uh, whether it’s, [00:29:21] um, Google through search, uh, whether it’s Amazon through [00:29:24] AWS and, uh, amazon.com, [00:29:27] um, so on and so forth, right. Uh, Microsoft [00:29:30] through, you know, windows Plus Office 3 6 5 plus [00:29:33] Azure, right? Um, they, they’ve all been the most profitable [00:29:36] companies. Am I, am I gonna continue with the onions and such?

[00:29:38] Allen Park: Yeah. [00:29:39] Or so are the eggs cooked?

[00:29:40] Dylan Patel: Yeah.

[00:29:40] Allen Park: Okay. You could put the [00:29:42] eggs on the plate.

[00:29:43] Dylan Patel: Yeah. Already done.

[00:29:44] Allen Park: Okay. And [00:29:45] then now add some oil and then Yeah. Add all the veggies.

[00:29:47] Dylan Patel: Yeah.

[00:29:47] Allen Park: Yeah. [00:29:48] But make sure not to add too many onions. ‘cause there’ll probably be [00:29:51]more onions and carrot’s. Another thing just try to have like an [00:29:54] even balance of vegetables.[00:29:57]

[00:29:57] Dylan Patel: Yeah.

[00:29:58] Allen Park: Yeah. And [00:30:00] then, yeah. So just get some color [00:30:03] on the veg.

[00:30:04] Dylan Patel: Um, do you, do you, uh, tend [00:30:06] to, um. Do you tend to do the [00:30:09] carrots and onions at the exact same time?

[00:30:10] Allen Park: Yeah. Like doing ‘em [00:30:12] all together is probably just easiest. So [00:30:15] carrots, onions in the white part of the scallions. [00:30:18]

[00:30:18] Dylan Patel: Okay. Um,[00:30:21] [00:30:24]

so, [00:30:27] so the hyperscalers have been the most profitable companies ever?

[00:30:29] Allen Park: Yeah. [00:30:30]

[00:30:30] Dylan Patel: Um, and now they’re, they’re about to face like sort of this like. [00:30:33] Interesting conundrum, right? Yeah. There’s a huge innovator’s dilemma here, [00:30:36] whether it’s, you know, meta, meta doesn’t own the [00:30:39] platform. Um, and, uh, wherever [00:30:42] people’s eyeballs are is where people are gonna like, [00:30:45] uh, spend their cash.

Um, or in, [00:30:48] in, in the case of Google, right? Hey, AI can disrupt search. Or in the [00:30:51] case of Microsoft, right? Productivity, uh, suite [00:30:54] is where they make all the money, right? Mm-hmm. Um, office 3, 6, 5, [00:30:57]windows, et cetera. But things like cloud code, quad bot, [00:31:00] and, and future iterations of it, just take a little imagination, we’ll [00:31:03] displace those immediately.

Mm-hmm. Right? Um, you know, you’ve [00:31:06] got like, you know, same with Amazon, right? AWS is, is. [00:31:09]Is, is a general purpose, [00:31:12] uh, you know, AI infrastructure or sort of infrastructure [00:31:15] play. Uh, but there’s a lot of risk with everything [00:31:18] else, um, with everyone getting disrupted quite heavily. Mm-hmm. And [00:31:21] so they’ve got this, this, this dilemma where they could get, um, [00:31:24] disrupted heavily.

Um, at the same time, they’ve also got, [00:31:27] uh, they’ve also got this challenge [00:31:30] with regards to. Um, potentially, you [00:31:33] know, being beaten, right? So they, they have to invest hugely in ai. They have [00:31:36] to try and win ai.

[00:31:37] swyx: Yeah.

[00:31:37] Dylan Patel: Um, and if they don’t, [00:31:39] then they’re really, really screwed. Um, but right [00:31:42]now the demand for AI is insatiable.

[00:31:44] swyx: Mm-hmm.

[00:31:44] Dylan Patel: And they [00:31:45] can get pretty good returns just by building infrastructure and renting it out [00:31:48] to the labs. Uh, but they’ll obviously get way better returns [00:31:51] if they, uh, if they have AI models in house. So they need to [00:31:54] spend like crazy to do this. Um, and [00:31:57] at the same time, um, if everyone else is like, sort of, it’s [00:32:00] like, um, it’s Pascal’s wager, right?

[00:32:03] Mm-hmm. Um, if I don’t spend like crazy and others do I [00:32:06] lose, right? If I don’t believe in God, right, digital God coming [00:32:09] then and others do and it happens, then I’m, I’m a loser, [00:32:12] right? Yeah. Um, and so they both got this [00:32:15] dilemma and the only solution is I have to spend more and [00:32:18] more, you know, until, uh, uh, to, to keep up in the [00:32:21] race.

And so, you know, this year’s announcements of [00:32:24] $180 billion of CapEx from Google and 200 from [00:32:27]Amazon, right. Which is, you know, four x what they were just [00:32:30] doing a few years ago.

[00:32:31] swyx: Mm-hmm.

[00:32:31] Dylan Patel: Um, is, is [00:32:33] quite incent intense. Uh, but in addition to that, [00:32:36] we’re looking at um, we’re looking at this like [00:32:39] skyrocketing in the next few years.

Yeah. Right. There’s no [00:32:42] reason why Google will have any profit, uh, in [00:32:45] 27 at all. Right? In terms of cash flow. They will just spend [00:32:48] every dollar they make on, on AI infrastructure. [00:32:51] Um, and I think that’s at least my belief, um, and, [00:32:54] and AI models and so on and so forth, because that, that’s basically my [00:32:57] belief and the market hasn’t fully woken up to this [00:33:00] realization.

Mm-hmm. Um, we’ve been saying it for a couple years. Um, in fact, we [00:33:03] even did a piece, uh, last year, which was like, how [00:33:06] much debt can the hyperscalers borrow? Right? Because at some point, [00:33:09] you know, they, they have to lever up on building [00:33:12] capacity, right? So an example of this is meta, right? Meta is not [00:33:15] as large as Google and Amazon, but they wanna be in the race.

[00:33:17] Allen Park: Yeah.

[00:33:17] Dylan Patel: [00:33:18] Um, and so they’ve already started taking some debt on to build their data center. [00:33:21] Now, obviously they have a tremendously profitable business that could pay it [00:33:24] off. Uh, they just have to stop spending the money, [00:33:27] um, on CapEx, that it’s not necessary. Uh, but, but Zuckerberg has [00:33:30]woken up and fully realizes this.

How much do I wanna cook the onions [00:33:33] and carrot carrots?

[00:33:33] Allen Park: Just get some color on it and then just put it [00:33:36] in the plate with the eggs. [00:33:39]

[00:33:39] Dylan Patel: Okay? Um. And so, so you’ve [00:33:42] seen some hyperscalers, such as, uh, meta, [00:33:45]they’ve, they’re already taking debt on for their largest AI cluster in [00:33:48] Louisiana.

[00:33:48] Allen Park: Mm-hmm.

[00:33:49] Dylan Patel: Um, you know, they’re taking like $40 billion of debt [00:33:51] on for that.

Uh, but they’re in the market to take on much, much [00:33:54] more. Um, Google and Amazon haven’t taken on debt yet for AI [00:33:57] infrastructure, but they will. Right. Mm-hmm.

[00:33:58] Allen Park: Yeah.

[00:33:59] Dylan Patel: And, and so [00:34:00] I think, I think people really realize and [00:34:03] panic and, and, and probably this year when [00:34:06] they see all of these companies doing exactly this, right?

Yeah. Are they [00:34:09] gonna, you know, what ends up happening, uh, when the most profitable [00:34:12] companies that have ever existed, which have compounded at double [00:34:15] digits for, uh, over a like a decade and a half [00:34:18]

[00:34:18] Allen Park: Yeah.

[00:34:18] Dylan Patel: Now all of a sudden say, we’re not gonna, we don’t care about profit [00:34:21] anymore. We’re just building pixie dust, right?

Mm-hmm. We’re building digital God. [00:34:24] Um, and if you believe in it, great. If you don’t, then you know, [00:34:27] tough luck. Um, and, and all this CapEx [00:34:30] predates infra revenue, right? ‘cause you need to have spent the CapEx [00:34:33] brought on the clusters and all that before you can have the, [00:34:36] um, before you can ever have the, [00:34:39] um, the revenue come online.

[00:34:40] Allen Park: Mm-hmm.

[00:34:41] Dylan Patel: Right? And then the [00:34:42] revenue starts off at lower margin, right?

[00:34:44] Allen Park: Yeah. [00:34:45] Also, it’s time to cook the chicken. So just add some oil and then we’ll cook [00:34:48] chicken.

[00:34:48] Dylan Patel: Okay. Sounds

[00:34:48] Allen Park: great. We’re done. Yeah. [00:34:51] Yeah. So saying investment before the revenue [00:34:54]comes online.

[00:34:54] Dylan Patel: Yeah. Yeah. So, so there’s like a timeline lag when [00:34:57] revenue comes online.

Uh, there’s a timeline lag in terms of like [00:35:00] when you rent the infrastructure versus like, Hey, you have to train the [00:35:03] model before you can ever start to, um, [00:35:06] um, before you can ever start [00:35:09] to[00:35:12]

actually get the, like, uh, AI [00:35:15] service revenue, right? Mm-hmm. Um, and so you’ve [00:35:18] seen this like with like. All the vendors, right? Like, you know, there, [00:35:21] there’s huge spend for open air anthropic and Google on [00:35:24] training models. Um, and, and the others are doing it too, like, uh, [00:35:27] Amazon and such, uh, before they ever end up with [00:35:30] enough revenue, uh, [00:35:33] generating revenue from the models, from the services that they sell on [00:35:36] top.

And so the market is just gonna really hate this.

[00:35:38] Allen Park: Yeah.

[00:35:38] Dylan Patel: [00:35:39] And I feel like that’s gonna lead to, despite the fact everyone [00:35:42] in San Francisco is gonna see revenue skyrocketing, they’re gonna see all the [00:35:45] amazing capabilities. Uh, but they, we live in a bubble, right? [00:35:48] If you told, we, we put out some research that was like, Hey, 4% [00:35:51] of commits on GitHub are quad code, and [00:35:54] everyone SF is like, that’s too low, right?

Like, it’s like a hundred [00:35:57] percent of mine. Yeah. Maybe it’s 50% for all the boomers, [00:36:00] right? And it’s like, no, no, no. Like, you know, we’ve got a lot of [00:36:03] adoption to go. Um, and so people are gonna [00:36:06]like, sort of like see all this amazing model progress and revenue [00:36:09] growth and adoption in sf, but then like in New York and [00:36:12] in London and like.

Hong Kong and other financial [00:36:15] capitals of the world, Singapore, et cetera, people are gonna see the [00:36:18]exact opposite, right? Mm-hmm. They’re gonna see, they’re gonna see the most [00:36:21] profitable companies ever are destroying their business [00:36:24] model to build, um, capacity in [00:36:27] something that maybe doesn’t necessarily, doesn’t have returns.

[00:36:29] Allen Park: Yeah. A big [00:36:30] bet

[00:36:30] Dylan Patel: towards. And, and so that, I think is gonna, and then, and [00:36:33] likewise, they’re gonna see the general public fucking hates ai. [00:36:36] And you’re gonna, you likely see like a real backlash to [00:36:39] AI from both the financial class and the normal people [00:36:42] of the world. Um,

[00:36:43] Allen Park: does it matter though, if the public hates [00:36:45] ai, if it provides a lot of value towards enterprise and [00:36:48] companies?

Like, isn’t the main value and profit coming [00:36:51] from enterprises more so than the general public? [00:36:54]

[00:36:54] Dylan Patel: Exactly. And I think, I think that, that like is like the [00:36:57] big fear, right? Mm-hmm. You know, we’ve already had like decades of like [00:37:00] people being like. Hey, income inequality is [00:37:03] bad. Um, and the value of labor has been, has been [00:37:06] falling, right?

Yeah. Uh, the value of labor used to be way, way higher, [00:37:09] right? Mm-hmm. Um, you know, as a percentage of the economy, [00:37:12] but as we’ve recognized as capital has become [00:37:15] more and more important as machinery has grown, um, we’ve [00:37:18] sort of had this major change, which is that, um, [00:37:21] you know, capital is, is taking more [00:37:24] and more of the, uh, uh, earnings, [00:37:27] uh, from of, of, of, uh.

[00:37:30] Capital’s taking more and more of their earnings. Mm-hmm. And, and, [00:37:33] and so people are really mad about that. And now we’re gonna start [00:37:36] seeing huge, uh, job loss too, right? Like, [00:37:39] hey, like, turns out there’s shit loads of software [00:37:42] developers just outta school, uh, who can’t get [00:37:45] jobs. Okay, fine. But what about like the 2 million people who [00:37:48] drive cars for a living?

Okay. Well, Waymo works well. [00:37:51] Tesla robo taxis starting to be deployed. Zoox is starting to be deployed. Yeah. [00:37:54] Um, you know, we’re, we’re starting to see really [00:37:57] the beginnings of all that. Yeah. Um, we’re gonna see, [00:38:00] you know, the stock market maybe does well or the eco, the [00:38:03] GDP d is gonna look good, but then normal people aren’t gonna be [00:38:06] accruing much value from it.

And so more and more, you know, [00:38:09] and eventually, like the financial markets will not do too well either because software’s [00:38:12] imploding. Because hyperscalers are gonna invest all their capital. [00:38:15] And you’re gonna end up with this like, major, major, [00:38:18] uh, weird fear and worry for everyone in the industry. [00:38:21] Um, or sorry, everyone in the world.[00:38:24]

[00:38:24] swyx: Mm-hmm.

[00:38:24] Dylan Patel: Um, and, and there’s like an AI backlash, right? [00:38:27] Um, and, and I think that’s gonna be like the hottest button issue of like [00:38:30] the next election, right. Um, if not [00:38:33] the midterms. Right? Yeah. Um, and it [00:38:36] seems obvious to me that like any party that wants to win should just become the [00:38:39] anti AI party, um, because [00:38:42] life as we know it is changing.

Mm-hmm.

[00:38:44] Allen Park: [00:38:45] Taking a little detour. Where do you think the alpha [00:38:48] is, or like the bet is because you said that NVIDIA’s kind [00:38:51] of covering their bases with the Ruben [00:38:54] CPX, uh, the GR chips standard GPUs, [00:38:57] um, and there’s a lot more startups out there that are very [00:39:00] specialized and even like a lot of YC companies right, are like popping [00:39:03] up and kind of tackling this industry.

Do you [00:39:06] think it’ll kind of turn out to be a play where like [00:39:09] Nvidia at the end of the day still rains and, um, like. [00:39:12] Crushes every other company that tries to take away the [00:39:15] market share? Or do you think there will actually be a lot of value [00:39:18] crew to these more specialized, smaller companies? [00:39:21]

[00:39:21] Dylan Patel: Um, in, in the, in the, in the chip space [00:39:24] specifically?

[00:39:24] Allen Park: Yeah.

[00:39:26] Dylan Patel: Um, [00:39:27] yes. I think that’s a really strong debate that, uh, [00:39:30] people are having.

[00:39:30] swyx: Mm-hmm.

[00:39:31] Dylan Patel: Um, is how much, how much [00:39:33] value accrues to Nvidia, how much value accrues to the [00:39:36] model companies? Um, how much do they start to, [00:39:39] um, really. Um, you know, do, [00:39:42] do, do smaller chip companies take [00:39:45] charge and, uh, win. Um, and it is [00:39:48] really a innovator’s dilemma in the sense that like, you know, [00:39:51] Hey, why did, why did Intel and a MD not win in a [00:39:54] IUs?

Well, it’s because they were, they were making money off of CPUs and [00:39:57] Nvidia was focused on parallel computing.

[00:39:59] swyx: Mm-hmm.

[00:39:59] Dylan Patel: [00:40:00] Um, and now you’ve got, got sort of the same question, which is, [00:40:03] um, you know, will, will Nvidia be able [00:40:06] to innovate on all the things that needs to be innovated? [00:40:09] Um, or, [00:40:12] or, we’ll,[00:40:15]

there’s a lot of scrape stuff on the bottom. Yeah, [00:40:18] yeah, yeah. Well, it’s like there’s stuff [00:40:21] stuck in the pan, you know? So I’m trying to scrape it off.

[00:40:23] Allen Park: Okay.

[00:40:23] Dylan Patel: We [00:40:24] could also

[00:40:24] Allen Park: get a new pan. I have two new pans,

[00:40:26] Dylan Patel: [00:40:27] actually. That would be amazing.

[00:40:28] Allen Park: The final thing, we just gotta [00:40:30] now add everything together. So first, add some oil to the pan [00:40:33] once it’s dry, and then add the garlic ginger, [00:40:36] but have it kind of lower.

Um,

[00:40:38] Dylan Patel: have it, [00:40:39] what

[00:40:39] Allen Park: have like the temperature be lower, like not too high. [00:40:42]

[00:40:42] Dylan Patel: Oh, really? Okay.

[00:40:42] Allen Park: Yeah.

[00:40:44] Dylan Patel: [00:40:45] Nvidia is, um, they’ve kind of got this like [00:40:48] innovator’s dilemma. The nice thing is they embody Silicon Valley spirit [00:40:51] more than maybe any other company.

[00:40:52] Allen Park: Mm-hmm.

[00:40:53] Dylan Patel: Uh, which is Andy [00:40:54] Grove. Right. Andy Grove, uh, from Intel.

Yeah. Um, only the

[00:40:56] Allen Park: paranoid [00:40:57] survive. Um,

[00:40:57] Dylan Patel: only the paranoid survive. Right. And, and I [00:41:00] think Jensen Long is like one of the most paranoid people in [00:41:03] the industry. Right. Yeah. Um, he’s, [00:41:06] he’s constantly like. Freaking out changing internal [00:41:09] things like, you know, in a good way though, right? Yeah. Like truly founder mode.[00:41:12]

Um, all right. Uh, the aromatics are [00:41:15] very aromatic.

[00:41:15] Allen Park: Okay. Then just add, um, all [00:41:18] the veg, don’t add everything, um, ‘cause like [00:41:21]proportions, but add like the onions, [00:41:24] carrots, add eggs, peas, and then chicken.[00:41:27] [00:41:30] [00:41:33]

And then once you have that, then add the rice. And [00:41:36] like should mix everything together and then [00:41:39] at the very end you’re gonna add [00:41:42] soy sauce and then some sugar and [00:41:45] salt to adjust.[00:41:48]

Yeah. But [00:41:51] Jensen,

[00:41:51] Dylan Patel: Jen very, okay. Jensen’s very paranoid. [00:41:54] Um, and that makes him like an amazing founder.

[00:41:56] Allen Park: Mm-hmm. [00:41:57]

[00:41:57] Dylan Patel: Um, and CEO. Um, and [00:42:00] so you, you have all these people freaking out, but it’s like the moment [00:42:03] he sniffed wind of the open Ai cereus deal. [00:42:06] Yeah. He immediately went out and was like, okay. I don’t actually, I didn’t [00:42:09] actually wasn’t building this technology because I didn’t believe in it.

But now [00:42:12] I do because open air trying to acquire, uh, is trying to use [00:42:15] gro uh, CEUs. So I’m just gonna go acquire rock, right? Like, you know, [00:42:18] it’s like, that’s like why he did it, right? Mm-hmm. So it’s like, you know, there’s, there’s a [00:42:21] bit of like, um, you know, the moment he sniffs [00:42:24] anything, he changes course in tune updates, his [00:42:27] priors.

Um, and I think that’s like [00:42:30] really impressive. And so as you step forward to like, Hey, [00:42:33] what about, um, you step forward to like, [00:42:36] okay, well what does that mean for his hardware roadmap? Well, yeah. Before he was [00:42:39] like, making one, one, you know, just a few kinds of architectures and [00:42:42] chips. Uh, but primarily it was all like very [00:42:45] similar, right?

Mm-hmm. It was a large G-P-G-P-U, [00:42:48] um, and it was like having, it was like the best memory, the best [00:42:51] networking, everything, sort of the best as possible. Um, [00:42:54] and sort of like one size fits all, uh, with the [00:42:57] main line of like a 100 H, 100 B 200, right? [00:43:00] Um, but as we look to Ruben and beyond mm-hmm. [00:43:03] Right? Jensen is really like starting to fully [00:43:06] embrace heterogeneity.

Mm-hmm. Right? Um, you know, [00:43:09] much like this fried rice, right? There’s no one individual ingredient that [00:43:12] shines above all right? You’ve kind of gotta have a [00:43:15] little bit of everything. Um, [00:43:18] and so, so this is like, this is like sort of like what [00:43:21] Jenssen’s believing here. So he is got, you know, he’s got this CPX chip.

[00:43:23] swyx: [00:43:24] Yeah.

[00:43:24] Dylan Patel: Right. Which is made for, um, [00:43:27] context processing pre-fill. It’s pretty good at video and image [00:43:30] gen as well. Mm-hmm. Um, but it’s not really good at latency sensitive [00:43:33] applications. You know, they’ve of course got their main line of GPUs [00:43:36] and now they’ve got this, these GR chips. Right. So, so, [00:43:39] uh, you know, they’re, they’ve, NVIDIA’s sort of [00:43:42] got every single, uh, aspect, uh, or [00:43:45] type of chip possible now within his company.

Um, and [00:43:48] he’s, he’s continuing to try and like ex uh, innovate and, [00:43:51] and move as fast as possible and all these things.

[00:43:52] Allen Park: Yeah.

[00:43:53] Dylan Patel: Um, [00:43:54] and so when we think about like, Hey, what [00:43:57] ends up happening? Um. [00:44:00] With Nvidia in, in this case, [00:44:03] it’s um, it’s, you know, JJ [00:44:06] Nvidia knows they will lose because they have a business, uh, [00:44:09] model deficit, right?

[00:44:10] Allen Park: Mm-hmm.

[00:44:10] Dylan Patel: Google, Amazon, [00:44:12] they get to vertically intergra integrate and vertical [00:44:15]integration always saves tons of money. Um, so he has to be [00:44:18] better than everyone.

[00:44:18] Allen Park: Yeah.

[00:44:19] Dylan Patel: Um, by not just like a little [00:44:21] bit by, by a ton to justify his margins [00:44:24]otherwise. The vertical integration of, of its competitors will win out.[00:44:27]

[00:44:27] Allen Park: Mm-hmm.

[00:44:28] Dylan Patel: Um, and, and, and so this is sort of like, I think the [00:44:30] big challenge, um, and I think, I think the story is not [00:44:33] finished right. Nvidia will remain on top this year and next year, [00:44:36] um, based on what we see, but others will gain some ground. [00:44:39] Uh, and the question is, what happens? In the long term. [00:44:42] Um, and, and honestly, [00:44:45] you know, the, the, the, the cards are up in the air, right?

[00:44:48] Um, no one has the the right to win. No one has a destiny to win. [00:44:51] Uh, things are moving so fast. Whoever, whoever does [00:44:54] the, you know, and innovates the hardest will win. Not, not [00:44:57] necessarily like. Oh. You know, and, and, and I think moats [00:45:00] are as shallow as they’ve ever been. Mm. Right. Uh, [00:45:03] because how fast things are moving the, the, the size of the [00:45:06] numbers that are being thrown around now, right.

It’s hundreds of billions of [00:45:09] dollars for each indi, uh, major hyperscalers. Yeah. [00:45:12] The size of the numbers are so large that you can just go and justify [00:45:15] hiring anyone, any talent. The mo’s become [00:45:18]much smaller. Um. And this, [00:45:21] this is sort of like pretty, pretty big [00:45:24] deal with regards to, you know, doesn’t video win or [00:45:27] not, right?

[00:45:27] Allen Park: Yeah. I guess what do you think’s the biggest [00:45:30] bottleneck for speed to keep us from [00:45:33] going as fast ball? Is it memory? What do you think is the main bottleneck? [00:45:36]

[00:45:36] Dylan Patel: Yeah, I think I have zero walkee, by the way. [00:45:39]

[00:45:39] Allen Park: Yeah.

[00:45:39] Dylan Patel: Um,

[00:45:40] Allen Park: I don’t think I’ll have any w with dude. [00:45:42]

[00:45:42] Dylan Patel: I swear to God, if Uncle Roger finds this video, I’m gonna cry.[00:45:45]

I’m like, he’s like, he’s gonna be like, no [00:45:48] MS Gya. No w hiya [00:45:51] induction furnace. What are you doing? [00:45:54] Um, no, but I, I, I, [00:45:57] but anyway, sorry. Um, you know, what’s the biggest bottleneck to [00:46:00] speed? You know, I think you cook fried rice much faster if you have a [00:46:03] walk. Um, so, okay. All right. I’m done. I’m done. [00:46:06] I’m not good enough to justify that there’s like a hundred other [00:46:09] mistakes I made, you know?

Um, but like, I think the biggest bottleneck to [00:46:12] like, Hey. Why, why only $200 billion this year [00:46:15] for Amazon? Why not 500? Right?

[00:46:16] Allen Park: Yeah.

[00:46:17] Dylan Patel: Um, I think, [00:46:18] I think there’s like a number of limiting factors and it’s sort of like year by year it’s [00:46:21] been different, right?

[00:46:21] Allen Park: Yeah.

[00:46:22] Dylan Patel: In 2023, it was [00:46:24] definitely all, um, related to [00:46:27] chips, right?

Semiconductors, COOs. Uh, [00:46:30] which is chip on wafer, on substrate. Um, driving up [00:46:33] production of this was very, very difficult. And then as we step forward [00:46:36] to 2020, um, [00:46:39] four, as you step forward to 2020, oh, I didn’t throw any sugar [00:46:42] or soy sauce. No wonder,

[00:46:43] Allen Park: oh, did you serve it already?

[00:46:44] Dylan Patel: [00:46:45] No, I didn’t. I almost did.

[00:46:45] Allen Park: Okay.

Yeah. Final step, just [00:46:48] sugar, soy sauce to taste. So taste it as the, [00:46:51]

[00:46:51] Dylan Patel: I dunno how much sugar to do, but. [00:46:54] Dude, this shit’s about to be sweeter than Panda [00:46:57] Express. Oh, [00:47:00] shit. I completely forgot about the soy sauce. Um, [00:47:03] anyways, [00:47:06]um, 2023, you know, in 2023 it was COOs, [00:47:09] it was co semiconductor supply chains. As we stepped forward to 24, [00:47:12] 25, it started to become data centers.

Mm-hmm. Um, [00:47:15] energy is a bigger deal in 25, 26. Um, but [00:47:18] as we step forward, right. You know, supply chains are fast and they [00:47:21] react quickly. Right. So this, this current like whole [00:47:24] thing of like, oh, data centers are the shortage. Yes. Data centers are a [00:47:27] shortage. Mm-hmm. Yes. Power is a shortage. Um, [00:47:30] at the end of the day, actually, there’s a [00:47:33] lot of other shortages around too.

Right. [00:47:36] Um, you know, and, and, and, and when you think about power, [00:47:39] okay, well, like, if you were not creative, right? And you [00:47:42] just relied on grid power, well, there’s only three companies that make [00:47:45]dual combine cycle reacts. Mm-hmm. But if you step forward to like, oh, [00:47:48] okay, well what if I make, uh, what if I want, um, [00:47:51] what if I, what if I take something else, right?

What if I [00:47:54] take, um, aero engines, okay, there’s a [00:47:57] few more vendors. What if I take industrial gas turbines, there’s a few more vendors. [00:48:00] What if I take, um.[00:48:03] [00:48:06]

What if I take a medium speed reciprocating [00:48:09] engines? Right. Um, which are these like, sort of like any, any [00:48:12] company that makes diesel engines, there’s dozens of them. Mm-hmm. Uh, [00:48:15] they can make medium speed reciprocating engines and I can use, I could [00:48:18] connect those up to make, um, power for the data center.

Right. So [00:48:21] when I look at, when I look at like, hey, what did, who, who, [00:48:24] who sort of broke these bounds, right? Elon was the first one. Just sort of say, [00:48:27] well, well, no, I don’t care about the actual rules. Let me just like, [00:48:30] lemme just put power generation on site with low [00:48:33] quality mobile turbines, right? Mm-hmm. Not turbines even, right?

[00:48:36] Industrial gas, uh, engines, um, you know, [00:48:39] reciprocating engines, et cetera, et cetera. Um, so, so Elon [00:48:42] broke all these rules and now the, the whole industry’s reacted [00:48:45] fast enough because there’s so many suppliers, right? Mm-hmm. And the lead time to [00:48:48] like ramp up production of these things is [00:48:51]ultimately not nearly as long as it is in the semiconductor [00:48:54] supply chain.

So going back to your question of like, Hey, what’s the big bottleneck? [00:48:57] Um, well, there’s, there, the, the big bottleneck is now [00:49:00] back again to semiconductors, right? Yeah. Semiconductors are extremely [00:49:03]cyclical. Uh, the, the, the buildings that chips are made in are the [00:49:06] most complicated buildings people make. Um, [00:49:09] you know, they are, they are, um, [00:49:12] you know, they have multi-year ti timelines.

Uh, they require [00:49:15] not, not just all the complexity of like electricians and [00:49:18] plumbers that da, uh, that data centers do, but they actually [00:49:21] require a lot more complex. Um, [00:49:24] because there’s, there’s all sorts of, uh, chemicals and, [00:49:27] and precursors and so on and so forth that are going through, uh, the [00:49:30] data center, right?

Or, or through the fab. And so, you know, people have just [00:49:33] not built enough fabs. And then, you know, that’s, that’s ignoring all the complicated [00:49:36] tools, right? These tools cost hundreds of millions of dollars in some [00:49:39] cases. Yeah. Um, they’re the most complicated thing people make. [00:49:42] Um, and so you end up with, wow.

So the, the [00:49:45] most, the, the, the challenge here is not just, [00:49:48] um, it’s, it’s, it’s ramping up production of semiconductors. [00:49:51] And, and so now we’ve entered an age in, especially in [00:49:54] 26, but as we go into 27, 28, um, you know, and, and [00:49:57] when we look in 2026, Google would buy a lot more tuss, but they [00:50:00]can’t ramp production fast enough.

Right. And so they have to buy tons of [00:50:03] GPUs and we go to 27, it applies again, right? Google simply [00:50:06] cannot buy enough Tuss and they have to buy tons of GPUs. [00:50:09] Um, and, and when you look across, um, the entire [00:50:12] supply chain, no one is getting enough capacity of semiconductors. Mm-hmm. [00:50:15] Um. Yes, they can put them in data centers.

Yes, they can get the [00:50:18] energy, uh, through maybe ghetto methods like putting [00:50:21] reciprocating engines, right? Uh, diesel reciprocating engines or [00:50:24] gas engines. But like, you know, may, maybe not the most [00:50:27]cleaner, uh, efficient thing, but they can do it. And so you end up [00:50:30] with, oh, okay, semiconductors are the shortage, but what’s the bottleneck to building more [00:50:33] fabs, uh, or to, to building more chips is more fabs.[00:50:36]

And people just have not built these fabs yet. Right. Yeah. And that’s, I think, the [00:50:39] big bottleneck now.

[00:50:39] Allen Park: Yeah,

[00:50:40] Dylan Patel: that makes sense. And that’s gonna persist through the end of the [00:50:42] decade or until ai, you know, sort of slows down.

[00:50:44] Allen Park: How are you [00:50:45] feeling, Dylan? Are you good?

[00:50:45] Dylan Patel: Uh, I, I turned around and I saw yours briefly, [00:50:48] so now I’m trying to like, wipe the edges of my bowl so it looks [00:50:51] beautiful, you know?

[00:50:51] Allen Park: Yeah. No needle worry too much.

[00:50:53] Dylan Patel: No, no, [00:50:54] no, no. People are gonna judge me hardcore. [00:50:57]

[00:50:58] Allen Park: How was that, Dylan? Was that fun?

[00:50:59] Dylan Patel: Uh, [00:51:00] it, it was quite fun. I think there was the right amount of stress in, uh, that’s

[00:51:02] Allen Park: [00:51:03] stressful

[00:51:03] Dylan Patel: involved. You know, I think I maybe didn’t share my thoughts as [00:51:06] well as normal, but

[00:51:06] Allen Park: Okay. That’s

[00:51:07] Dylan Patel: good.

Maybe that’s more natural and fun.

[00:51:08] Allen Park: Yeah. Like [00:51:09] a mix of things. Great. Yeah. And this is yours. Okay. Let’s try yours first [00:51:12] and then we’ll, well, let’s try the restaurant actually first, and then

[00:51:14] Dylan Patel: Control.

[00:51:15] Control.

[00:51:15] Allen Park: Yeah. Control. Okay. Cheers. Maybe a little cold. [00:51:18]

[00:51:18] Dylan Patel: I always feel like you gotta cheers the US food. [00:51:21]

[00:51:22] Allen Park: Mm.

That was, that was good.

[00:51:23] Dylan Patel: Mm-hmm. [00:51:24]

[00:51:24] Allen Park: Okay. And we’ll try yours now.

[00:51:26] Dylan Patel: Is that [00:51:27] okay?

[00:51:27] Allen Park: A lot of meat. I love it [00:51:30] as well. Cheers.[00:51:33]

Mm. Yeah. [00:51:36] I like my

[00:51:36] Dylan Patel: more.

[00:51:37] Allen Park: Yeah. [00:51:39] No, it’s very, it’s like a lot stronger. Like it’s [00:51:42] very deep. Mine’s probably gonna be [00:51:45] bland compared to that,

[00:51:46] Dylan Patel: so I, I, I think, um, [00:51:48] you know, there’s this like debate in the world, [00:51:51] right? French people, they don’t season [00:51:54] their food so much. Mm-hmm.

[00:51:55] Allen Park: They’re

[00:51:55] Dylan Patel: all about the ingredients shining.[00:51:57]

Then you got like, you know, equator, uh, [00:52:00] equatorial, people write slop, like Indian food and like [00:52:03]Caribbean food and like Southeast Asian food. Just [00:52:06] like throw on the spice, throw on the sugar, throw on the, like, [00:52:09] everything. And it’s like, okay, well then is this, [00:52:12] and, and sort of like the, the elitist French would say, that’s ‘cause [00:52:15] your ingredients suck.

You have to throw all this slop on there. Right. Gotta

[00:52:17] Allen Park: overdo it.

[00:52:17] Dylan Patel: Yeah. [00:52:18] But I don’t know. I’m, I’m, I’m a, I’m a slop son, you know?

[00:52:19] Allen Park: Yeah. I mean, it is very [00:52:21] tasty. So let’s try this. The [00:52:24] palettes are.[00:52:27]

Yeah. I think compared to yours, it’s like a lot [00:52:30] bland. It’s very [00:52:33] light.

[00:52:35] Dylan Patel: Not gonna [00:52:36] lie.

[00:52:37] Allen Park: You like your yours, the back,

[00:52:38] Dylan Patel: right? No, [00:52:39] I can tell

[00:52:40] Allen Park: honestly. Yours is like, [00:52:42]

[00:52:42] Dylan Patel: there’s a, there’s a mess up of technique, right? The bottom of [00:52:45] my, uh, pan was getting burnt.

[00:52:46] Allen Park: Mm-hmm.

[00:52:46] Dylan Patel: And you know, we can’t get a [00:52:48] walkee because this is induction cooktop, but because the bottom was getting [00:52:51] burnt, the smoke flavor I think was importing into the rice [00:52:54] and I got a walkee.

It’s intentional.

[00:52:56] Allen Park: [00:52:57] Yeah. Yeah. You could kind of tasted it here, [00:53:00] but it’s like hard. Very good. Okay. Are there any [00:53:03] callouts things that you all people don’t know? Are you [00:53:06] hiring, uh, clients? [00:53:09]

[00:53:09] Dylan Patel: Yeah, I am. I am hiring, um, we’re [00:53:12] 60 people now. Um, we work with all the top [00:53:15] companies in the world, major AI labs, major hyperscalers.

Mm-hmm. Um, [00:53:18] semiconductor companies, data center companies, et cetera. We cover the entire [00:53:21]swath from, um, AI infrastructure, AI [00:53:24] models, uh, token omics, right. Usage of AI models, [00:53:27] um, where, who, who’s using them, what are they using them for, what’s the [00:53:30] cost of running it? All these sorts of things. Yeah. So that’s the area we’re really expanding [00:53:33] into this year, um, and last year.

And so I think that’s the audience [00:53:36] that also matches sort of this. This audience that you have [00:53:39] here. Mm-hmm. So if anyone wants to track those things, right. Usage of [00:53:42] ai, the person who was working on this before, unfortunately [00:53:45] for them, but also a call out, they got hired by Anthropic. Right. You know, so [00:53:48]it’s sort of like, you know, the person who was working on this got, got hired by [00:53:51] Anthropic.

Yeah. As a, as a sort of like, you know, please don’t poach anymore by people. [00:53:54] Um, but I think it’s, it’s a good pathway, right? [00:53:57] Yeah. It’s like it’s showing your work is public. You know, if you’re gonna [00:54:00] kill it, you’re gonna kill it. You have all the resources behind you of knowing and understanding [00:54:03] infrastructure.

Yeah. I think that’s the big call out is like, we’re hiring for that role. We pay, [00:54:06] well, we have healthcare, we hire globally, we’re in eight or [00:54:09] 10 different countries like us, Japan, Taiwan, Singapore, France, [00:54:12]Germany. Wow. Israel. Yeah. Yeah. Canada, um, uk. Right. [00:54:15] So we’re everywhere. Um, yeah. I think that’s, that’s, [00:54:18] that’s the allure.

We get to work with all the coolest people.

[00:54:19] Allen Park: Great. That’s amazing. [00:54:21] Well, thank you so much, Dylan. I hope it was a great time.

[00:54:22] Dylan Patel: Well, thank you for having me. [00:54:24] Yes. Yes. [00:54:27]

[00:54:27] swyx: I really care about the quality of the chicken. Okay. [00:54:30] This

[00:54:30] Lydia: is like chicken egg fried rice, right?

[00:54:32] swyx: Yeah.

[00:54:32] Lydia: Wow. [00:54:33]

[00:54:33] swyx: Yeah. What’s your, how do you assess? [00:54:36]

[00:54:37] Dylan Patel: He, mines a little intense.[00:54:39]

[00:54:39] swyx: The isn’t intense. There’s probably too much soy sauce.

[00:54:41] Dylan Patel: [00:54:42] That’s fair. I forgot to throw it in, and then I just threw it

[00:54:43] swyx: all in [00:54:45] yo load. How much does you have? See, he, [00:54:48] he left like he

[00:54:48] Dylan Patel: little left a little

[00:54:49] swyx: bit. Yeah. Yeah. Yeah. [00:54:51] I would say, you know, like I, I would take this, I [00:54:54] would take Evans, I would take, I, I would also take this, you said this sauce for me.

I don’t know, but. [00:54:57]

[00:54:57] Lydia: I like Dylan’s.

[00:54:58] swyx: Okay. [00:55:00]

[00:55:00] Lydia: Of salty.

[00:55:00] swyx: There’s one each. Brandon, you’re the tiebreaker. Come on the [00:55:03] side of the slot. [00:55:06]Wait, what do you vote? What do you vote? You worked in a restaurant, bro.

This Substack is reader-supported. To receive new posts and support my work, consider becoming a free or paid subscriber.

AIE Europe CFP and AIE World’s Fair paper submissions for CAIS peer review are due TODAY - do not delay! Last call ever.

We’re excited to welcome METR for their first LS Pod, hopefully the first of many:

METR are keepers of currently the single most infamous chart in AI:

But every Latent Space reader should be sophisticated enough to know that the details matter and that hype and hyperbole go hand in hand in AI social media, because the millions of impressions that got, by people who don’t understand or care about the nuances, disclaimers, and error bars, far outreaches the 69k views on the corrections by the people who actually made the chart:

There’s a lot of nuance both in making benchmarks (as we discovered with OpenAI on our SWE-Bench Verified podcast) and in extrapolating results from them, especially where exponentials and sigmoids are concerned. METR’s Long Horizons work itself has known biases that the authors have responsibly disclosed, but go far too underappreciated in the pursuit of doomer chart porn.

If you’re interested in a short, sharable TED talk version of this pod, over at AIE CODE we were blessed to feature Joel twice, as a stage talk and with a longer form small workshop with Q&A:

We also make sure cover some of METR’s lesser known work on Threat Evaluation but also Developer Productivity, where 2x friend of the pod and now Zyphra founder Quentin Anthony was the ONLY productive participant!

Finally, if you’re the sort to read these show notes to the end, then you definitely deserve some pictures of Joel shredding the guitar at Love Band Karaoke which we mention at the end:

00:00 What METR Means
00:39 Podcast Intro With Joel
01:39 ME vs TR
03:33 Time Horizon Origin Story
04:56 Picking Tasks And Biases
09:13 Time Horizon Misconceptions
11:37 Opus 4.5 And Trendlines
14:27 Productivity Studies And Explosions
29:50 Compute Slows Progress
30:47 Algorithms Need Compute
32:45 Industry Spend and Data
34:57 Clusters and Shipping Timelines
36:44 Prediction Markets for Models
38:10 Manifold Alpha Story
43:04 Beyond Benchmarks Evals
51:39 METR Roadmap and Farewell

The Best Image Model is back!

AI News for 2/25/2026-2/26/2026. We checked 12 subreddits, 544 Twitters and 24 Discords (263 channels, and 12920 messages) for you. Estimated reading time saved (at 200wpm): 1283 minutes. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies!

Congrats to Perplexity on Computer and for replacing Bixby as default AI on hundreds of millions of Samsung phones going forward, but those are more consumery news.

News that AI Engineers can use from today is Nano Banana 2, which is more formally called 3.1 Flash Image. The big story is the pricing: it is rated the #1 image model in the world per Arena and ArtificialAnalysis, and yet costs half the price (At $67/1k images, vs Nano Banana Pro ($134/1k) and GPT Image 1.5 ($133/1k) for generation, and FLUX.2 [max] at $140/1k images for editing).

It comes with 6 character consistency and search grounding, and great text rendering. Of course every generation is stamped by SynthID.

Of course, there’s no research detail or paper whatsoever, so our coverage ends here.

Google’s Nano Banana 2 (Gemini 3.1 Flash Image Preview) and the new image-eval meta

• Nano Banana 2 launch + rollout footprint: Google and DeepMind shipped Nano Banana 2 (aka Gemini 3.1 Flash Image Preview) as a “Flash-tier” image generation/editing model rolling out across Gemini App, Search (AI Mode/Lens), Flow, Google Ads, and in preview via AI Studio / Gemini API / Vertex AI (Google, GoogleDeepMind, GeminiApp, sundarpichai, demishassabis). Product claims emphasize world knowledge, improved i18n text rendering, aspect ratio control, upscaling up to 4K, and multi-subject consistency (e.g., “up to 5 characters & 14 objects”) (Google, joshwoodward).

• Arena/Artificial Analysis results + pricing signal: Multiple evaluators report Nano Banana 2 taking #1 Text-to-Image and strong editing placements, while undercutting “Pro” pricing—e.g., $67 / 1k images cited by Artificial Analysis vs ~$133–134 for GPT Image 1.5 and Nano Banana Pro (ArtificialAnlys, arena, kimmonismus). Arena added image subcategories and highlighted largest gains in text rendering and 3D imaging/modeling (arena). This is also a reminder that leaderboards are becoming product levers: “day-0” integrations (e.g., fal) and prompt packs/templates ship alongside eval wins (GeminiApp templates, GoogleAI prompts).

• Real-time search-conditioned generation: Google repeatedly frames NB2 as powered by real-time information and images from web search (not just static pretraining), positioning it as “more accurate views from any window in the world” style demos (sundarpichai).

• Downstream availability: Nano Banana 2 quickly appeared in third-party products, including Perplexity Computer (AravSrinivas).

Agentic coding + productized “tasks”, memory, and evals (and the backlash against complexity)

• Agents “just work” more often now, but still fail off-distribution: Several practitioners describe a step change in reliability/utility for coding agents across recent frontier models (e.g., GPT-5.2 / GPT-5.3 Codex, Opus 4.6, Gemini 3.1), shifting from “proof of concept” to something like delegating CLI work to juniors (teortaxesTex, paul_cal). Others caution that advanced ML/data engineering remains brittle off-distribution (michalwols, MParakhin).

• “Tasks” as the new packaging layer: Microsoft’s Copilot Tasks pitches “talk less, do more” delegation with user-visible plans and control, in “research preview” (mustafasuleyman, yusuf_i_mehdi).

• Persistent memory becomes table stakes—and introduces interoperability pain: A widely shared update claims Claude rolled out auto-memory (“remembers what it learns across sessions”) (trq212), echoed in the Claude ecosystem (omarsar0). Developers immediately hit workflow friction when memory/state lives in tool-specific hidden directories (hurting “multi-agent, multi-tool” continuity) (borisdayma).

• Tooling ships fast: PR bug-fixing bots, code↔design loops, and editor-level improvements:

  • Cursor Bugbot Autofix automates fixing issues found in PRs (cursor_ai, aye_aye_kaplan).

  • OpenAI’s Codex “code → design → code” roundtrip with Figma aims to make UI iteration less lossy (OpenAIDevs, figma).

  • VS Code’s long-distance Next Edit Suggestions focuses on predicting where not to edit and supporting “flow” (code, pierceboggan, alexdima123).

• Eval inflation + benchmark gaming concerns: Threads call out that high leaderboard scores can mask token-inefficient reasoning and failures on “bullshit tests” (e.g., repeated-token “strawberry” variants), warning against over-trusting HLE/GPQA-style metrics without cost accounting (scaling01). Arena responds by adding more granular test regimes like Multi-File React for code models (arena).

• Complexity is the real tax: A recurring engineering concern is that “10k LOC/day” bragging creates long-term complexity debt—agents make it easier to ship, not easier to maintain (Yuchenj_UW). Another angle: coding agents can create implicit lock-in if they “sloppify” your codebase such that working without them becomes painful (typedfemale).

Perplexity’s distribution + retrieval stack: Samsung integration and new embedding models

• Samsung S26 system-level Perplexity (“Hey Plex”): Perplexity says every Galaxy S26 will ship with Perplexity built in, including a wake word and deep OS integration; Bixby routes web/research/generative queries to Perplexity while handling on-device actions (perplexity_ai, perplexity_ai, AravSrinivas). This is framed as part of a broader partnership that also targets Samsung Internet and optional default search positioning (perplexity_ai).

• pplx-embed / pplx-embed-context released (MIT): Perplexity launched two embedding model families at 0.6B and 4B, including a “context” variant intended for doc chunk embeddings in RAG; both are MIT licensed and available via HF + Perplexity API, with a paper (arXiv:2602.11151) (perplexity_ai, perplexity_ai, alvarobartt). They also disclose internal benchmarks like PPLXQuery2Query / PPLXQuery2Doc with 115k real queries over 30M docs from 1B+ pages (perplexity_ai). Arav claims the embedding models are “industry leading” (AravSrinivas).

• Strategic read: The pair of moves—OS distribution + retrieval primitives—suggests Perplexity is trying to own both front door (assistant entry point) and core search stack (embeddings + evals), rather than depending on third-party platforms.

Inference, kernels, and infra: MoE support, heterogeneous hardware, and KV movement

• MoE becomes “first-class” in 🤗 Transformers: Hugging Face shipped deeper MoE plumbing (loading, expert backends, expert parallelism, hub support) and highlights collaboration on faster MoE training (including with Unsloth) (ariG23498, mervenoyann).

• DeepSeek and multi-hardware inference seriousness: Early in the batch, DeepSeek is called out as “serious about inference support on diverse hardware” (teortaxesTex). Separately, a DeepSeek DualPath detail describes staging KV cache in decode-server DRAM then moving it to prefill GPUs via GDRDMA to avoid local PCIe bottlenecks (JordanNanos). This reflects a broader shift: inference is increasingly a systems architecture problem, not just kernel-level optimization.

• Kernel coverage and GPU generations: vikhyatk describes building inference kernels across NVIDIA architectures (sm80→sm110) and notes edge-device ISA issues like Orin CPU lacking SVE (vikhyatk, vikhyatk).

• Quantization isn’t uniformly safe: Evaluations show MiniMax M2.5 GGUF quantizations degrade much more than expected vs Qwen3.5, arguing “just take Q4” doesn’t generalize across model families (bnjmn_marie).

World models, agents in simulators, and “multiplayer” environments

• Solaris: multiplayer Minecraft world modeling stack: A major research drop proposes that world modeling should focus on shared global state rather than pixel rendering, releasing (1) a multiplayer data collection engine, (2) a multiplayer DiT with a “memory efficient self forcing design” trained on 12.6M frames, and (3) a VLM-judge evaluation suite for multi-agent consistency (sainingxie, georgysavva). The pitch: multi-agent capability requires a shared representation beneath individual views.

• LLMs as embodied controllers (toy but telling): A CARLA→OpenEnv port shows a small Qwen 0.6B learning to brake/swerve to avoid pedestrians in ~50 steps using TRL + HF Spaces (SergioPaniego). This exemplifies a trend toward “LLM+env” loops where reversibility is limited and mistakes persist.

Governance flashpoint: Anthropic vs the Pentagon on surveillance and autonomous weapons

• Pentagon pressure campaign reported, then Anthropic responds publicly: A widely shared claim says the DoD issued a “final offer” to Anthropic, including threats to label it a “supply chain risk” and demands for unrestricted military use (KobeissiLetter). Anthropic then published a CEO statement drawing explicit red lines: no mass domestic surveillance and no fully autonomous weapons (given current reliability), also alleging threats involving the Defense Production Act (AnthropicAI). A widely quoted excerpt is reposted with detail (AndrewCurran_).

• Industry reaction + solidarity mechanics: The stance triggered strong support from prominent researchers/engineers, framing it as values-under-pressure rather than “policy theater” (fchollet, TrentonBricken, awnihannun). A petition aiming to coordinate “shared understanding” reportedly gathered signatures from OpenAI/Google staff (jasminewsun, sammcallister, maxsloef). This is notable as an explicit attempt to prevent a race-to-the-bottom dynamic via transparency about where each lab stands.

• Why this matters technically: The core dispute is about capability vs. reliability and “lawful use” language being misaligned with what frontier models can safely do today. Reliability concerns show up elsewhere in the dataset too (e.g., minimal security test cases where models leak confidential info even when instructed not to) (jonasgeiping, random_walker).

• Anthropic CEO statement on DoD demands (surveillance + autonomous weapons red lines) — @AnthropicAI

• Google launches Nano Banana 2 / Gemini 3.1 Flash Image Preview (broad rollout + “pro at flash speed”) — @GeminiApp, @sundarpichai, @GoogleDeepMind

• Perplexity + Samsung S26 system-level integration (“Hey Plex”) — @perplexity_ai

• Claude connectors available on free plan (150+ connectors) — @claudeai

• Pentagon vs Anthropic “final offer” reporting thread — @KobeissiLetter

• Claude Code auto-memory is huge (developer reaction) — @trq212

Swyx joined SAIL! Thank you , , , , , and many others for tuning into SAIL Live #6 with and . Sharing here for the LS paid subscribers.

We covered:

The problem with wanting to cover AI without giving in to hype is, sometimes, there REALLY IS something warranting the hype every few weeks. It is hard to tell when many influencers make their living telling you {MODEL THAT JUST RELEASED} is AGI and {MODEL THEY PREVIOUSLY ENDORSED} is now absolute trash. The simple solution is to try to form your own opinions… and rely on credible voices that won’t sell out to the attention game.

There’s been an increasing number of the latter (and also the noise as well).

You can see Andrej trying to articulate his feelings about it in real time, from the original vibe coding tweet to agentic engineering. Today he joins a chorus of people calling out December 2025 (where Opus 4.5, Gemini 3, and GPT 5.2 on the OpenAI 10 year anniversary were all released in quick succession) as a major turning point:

We were already planning a mini essay around Greg Brockman’s similar observation:

and Boris’ observation that Claude Code is now self-hosting:

and SemiAnalysis’ observation that Claude Code now writes about 5% of GitHub commits:

We’ve made a microsite for this:

https://wtfhappened2025.com/

We’ll maintain this as a live updating resource as more observations and chart inflections come in. The data is of course sparse now and lacks time series charts because it is hard to estimate a sigmoid while you are living through it, but we will do our best to curate datapoints the way the old WTF 1971 did. We like to joke that AINews is the “only newsletter that tells you not to read it” because of the prevalence of “nothing days”, but this is increasingly not a laughing matter.

AI News for 2/24/2026-2/25/2026. We checked 12 subreddits, 544 Twitters and 24 Discords (262 channels, and 10751 messages) for you. Estimated reading time saved (at 200wpm): 1086 minutes. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies!

Perplexity “Computer”: an orchestration-first agent product (multi-model, tool+env, usage-based pricing)

• Perplexity Computer launch: Perplexity introduced Computer, positioned as an end-to-end system that can “research, design, code, deploy, and manage” projects by orchestrating files, tools, memory, and models in one interface (launch tweet, Arav Srinivas). Key product signals:

  • Access + pricing: available on web for Max subscribers first, then Pro/Enterprise; usage-based pricing with sub-agent model selection, spending caps, and credits included for Max (10k/mo) plus a time-limited bonus credit grant (pricing details, availability, Arav on rollout).

  • Architecture emphasis: multiple tweets stress that the “breakthrough” is parallel, asynchronous sub-agents with a coordinator model assigning tasks to specialist models (research vs coding vs media), rather than a single monolithic agent loop (Lior’s breakdown, Denis Yarats).

  • “Everything is computer” narrative: Perplexity staff amplified Computer as a platform built by a small team with extensive use of coding agents and automated eval/debug loops (Arav, Denis).

• Why it matters to engineers: Computer is a concrete push toward systems-level agent UX: multi-model routing, isolation/sandboxes, persistent memory, and cost controls—i.e., treating “agentic work” as a distributed workflow rather than a single chat session (Arav, Computer site).

Coding agents: “it started working in December” + new model/tooling drops (GPT‑5.3‑Codex, Claude Code ecosystem, Copilot CLI GA)

• Karpathy’s “phase change” claim: Andrej Karpathy argues that coding agents crossed a qualitative threshold since December—from brittle demos to sustained, long-horizon task completion with coherence and tenacity. He gives a detailed example of delegating an end-to-end local deployment (SSH keys → vLLM → model download/bench → server endpoint → UI → systemd → report) with minimal intervention (Karpathy). This aligns with broader “software is changing” sentiment from devtool builders and users (Cursor, snowmaker).

• OpenAI GPT‑5.3‑Codex release + early eval chatter:

  • OpenAI shipped GPT‑5.3‑Codex in the API (snsf) and Cline announced support with claimed gains: ~25% faster vs 5.2, fewer tokens/task, and strong SWE-Bench Pro performance (Cline).

  • Community benchmark reactions were sharp (and noisy): e.g., “86% on IBench” surprise (tweet) and “first benchmarks incoming” (kimmonismus). Treat these as directional until methodology is clear.

• Claude Code: product maturity + observability + integrations:

  • Claude Code’s “first birthday” framing and retrospectives emphasize it as a foundational coding agent product, plus concerns about context length scaling hitting memory constraints (swyx).

  • Practical ecosystem bits: Slack plugin integration for Claude Code (catwu); LangSmith tracing for Claude Code to debug “nerfing”/routing issues (hwchase17, observability complaint).

• GitHub Copilot CLI goes GA + “/research”:

  • Copilot CLI reached GA (Evan Boyle) and added /research for repo-wide deep research using GitHub code search + MCP-based dynamic fetching, exporting reports to gists for sharing (feature).

  • Smaller UX note: Copilot CLI in terminal updates titles in real time (tweet).

Open models & local inference: Qwen3.5 “Medium” wave (MoE + long context + FP8/quant), and the local-agent tipping point

• Qwen3.5 Medium series distribution blitz: Alibaba pushed day-0 tooling support across vLLM, GGUF, LM Studio, Ollama, and Jan, highlighting how fast the deployment stack is now for major open releases (vLLM thanks, GGUF, LM Studio, Ollama, Jan).

• Key technical claims from Qwen (as posted, not independently verified here):

  • Quantization robustness: “near-lossless” accuracy under 4-bit weight + KV-cache quantization.

  • Long-context: Qwen3.5‑27B supports 800K+, 35B‑A3B >1M context on 32GB VRAM consumer GPUs, 122B‑A10B 1M+ on 80GB GPUs.

  • Open base: Qwen open-sourced Qwen3.5‑35B‑A3B‑Base to support research (Alibaba_Qwen).

  • FP8 weights open with native vLLM/SGLang support (FP8 announcement).

• Local agents “before/after”: A notable practitioner claim is that Qwen3.5‑35B‑A3B makes local agent loops feel meaningfully more reliable (tool calling, stability) while activating only ~3B params/token—explicitly positioning local as viable alongside Claude Code/Codex for many workflows (victormustar).

• Eval discourse warning: benchmaxxing & MoE vs dense confusion:

  • Multiple threads caution against over-reading leaderboards (“please stop falling for benchmaxxing”) (scaling01) and highlight surprising parity across Qwen sizes on some benchmarks, suggesting either tooling effects or benchmark artifacts (eliebakouch, teortaxesTex on HLE/MoE interpretation).

  • Arena added Qwen3.5 Medium to Text/Vision/Code Arena for head-to-head comparisons (Arena).

Agents, reliability, and “building for agents”: minimal benchmarks, tool-interface optimization, and failure modes

• Reliability hasn’t improved like capability: A reliability-focused line of work argues that despite rapid model progress, reliability gains are modest, decomposing reliability into many dimensions and warning against reducing agent performance to a single “success rate” number (IEthics, Justin Bullock quote).

• Agent failures are often reliability, not capability: A summary of an “agent failure” paper claims agents frequently fail by compounding small off-path tool calls, where one mistake increases the likelihood of the next, especially in long-horizon settings (omarsar0).

• Minimal “safe & helpful” benchmark idea: Instead of harder tasks, one proposal is to measure whether models can reliably do trivially specified safe behaviors (e.g., “send email only if asked”), including under irrelevant/distracting context; the claim is frontier models still miss cases (jonasgeiping).

• Tool descriptions as an optimization target (Trace‑Free+): Intuit AI Research work suggests agent success depends heavily on tool-interface text, and introduces a curriculum that teaches models to rewrite tool descriptions into agent-usable forms without requiring traces at inference time; reported gains on StableToolBench/RestBench and robustness with >100 tools (omarsar0).

• GUI/web agents: planning vs reactive: ActionEngine reframes GUI agents as graph traversal with offline exploration producing a state-machine; runtime generates a full program with ~1 LLM call, claiming big success/cost/latency improvements over step-by-step vision loops (dair_ai).

Compute, memory, and inference-speed frontiers: chip memory hierarchies, diffusion LLMs, and infra for scaling

• Karpathy on the “tokens tsunami” and memory orchestration: A high-engagement thread frames the core constraint as two distinct memory pools—fast, tiny on-chip SRAM vs large, slow off-chip DRAM—and argues the biggest puzzle is orchestrating memory+compute for LLM workflows (prefill/decode/training) with best throughput/latency/$, especially decode under long context + tight agentic loops, which is hard for both “HBM-first” (NVIDIA-like) and “SRAM-first” (Cerebras-like) camps (Karpathy).

• Diffusion LLMs as a speed alternative:

  • Andrew Ng highlighted impressive inference speed from Inception Labs’ diffusion LLMs (AndrewYNg).

  • Separate discussion claims diffusion approaches can hit ~1000 tok/s and shift the speed game via architecture, not chips (interpret cautiously; marketing often outpaces reproducible evals) (kimmonismus).

  • Research thread: “Diffusion Duality (Ch.2) Ψ-Samplers” for inference-time scaling in uniform diffusion-LLMs (ssahoo_).

• Interpretability at scale: Goodfire described infra work enabling trillion-parameter-scale interpretability with minimal inference overhead, harvesting billions of activations and enabling real-time steering of chain-of-thought in at least one case study (GoodfireAI).

Major announcements & policy/safety pressure points: Anthropic acquisitions + RSP shift, surveillance concerns, and market/power constraints

• Anthropic acquires Vercept to advance Claude’s “computer use” capabilities (AnthropicAI); Vercept’s founder thread frames the mission as moving from “telling users what to do” to acting for users, especially for non-technical tasks (ehsanik).

• Anthropic “RSP v3” shift (Responsible Scaling Policy): Commentary indicates a move away from rigid, unilateral “stop training past thresholds unless mitigations are guaranteed” toward more frequent transparency artifacts (roadmaps + risk reports), plus updated threat models and external review commitments (MaskedTorah). A more sensationalized summary claims this reflects competitive pressure and uncertainty in risk science (kimmonismus).

• Surveillance and civil liberties: Jeff Dean explicitly agreed that mass surveillance chills speech, invites misuse, and violates constitutional protections (JeffDean). Related tweets raised concerns about autonomous policing/surveillance agents that can’t refuse illegal orders (BlackHC).

• Energy as a binding constraint: One report claims U.S. political leadership is pushing major AI/data-center firms to self-provision electricity to avoid ratepayer backlash as demand strains the grid (kimmonismus)—an example of AI scaling becoming as much infrastructure/policy as algorithms.

• Grok 4.20 Beta leaderboard movement: Arena reports Grok‑4.20‑Beta1 at #1 on Search Arena and #4 on Text Arena (arena). Treat as one signal among many; Arena rankings can shift with sampling policies and model variants.

• Karpathy on the “phase change” in coding agents since December

• Perplexity launches “Computer”

• Arav Srinivas: what Perplexity has been building + “Computer”

• Karpathy on compute: SRAM vs DRAM orchestration for token-heavy LLM workloads

• Anthropic acquires Vercept for computer-use capabilities

• Qwen3.5 long-context + quantization + base model details

• Local agents tipping point: run Qwen3.5‑35B‑A3B locally with 32GB RAM

• Goodfire: infra for interp at trillion-parameter scale

• ActionEngine: offline GUI exploration → O(1) LLM-call execution programs

Editor’s note: CuspAI raised a $100m Series A in September and is rumored to have reached a unicorn valuation. They have all-star advisors from Geoff Hinton to Yann Lecun and team of deep domain experts to tackle this next frontier in AI applications.

In this episode, Max Welling traces the thread connecting quantum gravity, equivariant neural networks, diffusion models, and climate-focused materials discovery (yes, there is one!!!).

We begin with a provocative framing: experiments as computation. Welling describes the idea of a “physics processing unit”—a world in which digital models and physical experiments work together, with nature itself acting as a kind of processor. It’s a grounded but ambitious vision of AI for science: not replacing chemists, but accelerating them.

Along the way, we discuss:

• Why symmetry and equivariance matter in deep learning

• The tradeoff between scale and inductive bias

• The deep mathematical links between diffusion models and stochastic thermodynamics

• Why materials—not software—may be the real bottleneck for AI and the energy transition

• What it actually takes to build an AI-driven materials platform
  

Max reflects on moving from curiosity-driven theoretical physics (including work with Gerard ‘t Hooft) toward impact-driven research in climate and energy. The result is a conversation about convergence: physics and machine learning, digital models and laboratory experiments, long-term ambition and incremental progress.

• 00:00:00 – The Physics Processing Unit (PPU): Nature as the Ultimate Computer

• Max introduces the idea of a Physics Processing Unit — using real-world experiments as computation.

• 00:00:44 – From Quantum Gravity to AI for Materials

• Brandon frames Max’s career arc: VAE pioneer → equivariant GNNs → materials startup founder.

• 00:01:34 – Curiosity vs Impact: How His Motivation Evolved

• Max explains the shift from pure theoretical curiosity to climate-driven impact.

• 00:02:43 – Why CaspAI Exists: Technology as Climate Strategy

• Politics struggles; technology scales. Why materials innovation became the focus.

• 00:03:39 – The Thread: Physics → Symmetry → Machine Learning

• How gauge symmetry, group theory, and relativity informed equivariant neural networks.

• 00:06:52 – AI for Science Is Exploding (Not Emerging)

• The funding surge and why AI-for-Science feels like a new industrial era.

• 00:07:53 – Why Now? The Two Catalysts Behind AI for Science

• Protein folding, ML force fields, and the tipping point moment.

• 00:10:12 – How Engineers Can Enter AI for Science

• Practical pathways: curriculum, workshops, cross-disciplinary training.

• 00:11:28 – Why Materials Matter More Than Software

• The argument that everything—LLMs included—rests on materials innovation.

• 00:13:02 – Materials as a Search Engine

• The vision: automated exploration of chemical space like querying Google.

• 01:14:48 – Inside CuspAI: The Platform Architecture

• Generative models + multi-scale digital twin + experiment loop.

• 00:21:17 – Automating Chemistry: Human-in-the-Loop First

• Start manual → modular tools → agents → increasing autonomy.

• 00:25:04 – Moonshots vs Incremental Wins

• Balancing lighthouse materials with paid partnerships.

• 00:26:22 – Why Breakthroughs Will Still Require Humans

• Automation is vertical-specific and iterative.

• 00:29:01 – What Is Equivariance (In Plain English)?

• Symmetry in neural networks explained with the bottle example.

• 00:30:01 – Why Not Just Use Data Augmentation?

• The optimization trade-off between inductive bias and data scale.

• 00:31:55 – Generative AI Meets Stochastic Thermodynamics

• His upcoming book and the unification of diffusion models and physics.

• 00:33:44 – When the Book Drops (ICLR?)

Max: I want to think of it as what I would call a physics processing unit, like a PPU, right? Which is you have digital processing units and then you have physics processing units. So it’s basically nature doing computations for you. It’s the fastest computer known, as possible even. It’s a bit hard to program because you have to do all these experiments. Those are quite bulky, it’s like a very large thing you have to do. But in a way it is a computation and that’s the way I want to see it. You can do computations in a data center and then you can ask nature to do some computations. Your interface with nature is a bit more complicated. But then these things will have to seamlessly work together to get to a new material that you’re interested in.

[01:00:44:14 - 01:01:34:08]

Brandon: Yeah, it’s a pleasure to have Max Woehling as a guest today. Max has done so much over his career that I’ve been so excited about. If you’re in the deep learning community, you probably know Max for his work on variational autocoders, which has literally stood the test of prime or officially stood the test of prime. If you are a scientist, you probably know him for his like, binary work on graph neural networks on equivariance. And if you’re a material science, you probably know him about his new startup, CASPAI. Max has a long history doing lots of cool problems. You started in quantum gravity, which is I think very different than all of these other things you worked on. The first question for AI engineers and for scientists, what is the thread in how you think about problems? What is the thread in the type of things which excite you? And how do you decide what is the next big thing you want to work on?

[01:01:34:08 - 01:02:41:13]

Max: So it has actually evolved a lot. In my young days, let’s breathe, I would just follow what I would find super interesting. I have kind of this sensor. I think many people have, but maybe not really sort of use very much, which is like, you get this feeling about getting very excited about some problem. Like it could be, what’s inside of a black hole or what’s at the boundary of the universe or what are quantum mechanics actually all about. And so I follow that basically throughout my career. But I have to say that as you get older, this changes a little bit in the sense that there’s a new dimension coming to it and there’s this impact. Going in two-dimensional quantum gravity, you pretty much guaranteed there’s going to be no impact on what you do relative, maybe a few papers, but not in this world, this energy scale. As I get closer to retirement, which is fortunately still 10 years away or so, I do want to kind of make a positive impact in the world. And I got pretty worried about climate change.

[01:02:43:15 - 01:03:19:11]

Max: I think politics seems to have a hard time solving it, especially these days. And so I thought better work on it from the technology side. And that’s why we started CaspAI. But there’s also a lot of really interesting science problems in material science. And so it’s kind of combining both the impact you can make with it as well as the interesting science. So it’s sort of these two dimensions, like working on things which you feel there’s like, well, there’s something very deep going on here. And on the other hand, trying to build tools that can actually make a real impact in the world.

[01:03:19:11 - 01:03:39:23]

RJ: So the thread that when I look back, look at the different things that you worked out, some of them seem pretty connected, like the physics to equivariance and, yeah, and, uh, gravitational networks, maybe. And that seems to be somewhat related to Casp. Do you have a thread through there?

[01:03:39:23 - 01:06:52:16]

Max: Yeah. So physics is the thread. So having done, you know, spent a lot of time in theoretical physics, I think there is first very fundamental and exciting questions, like things that haven’t actually been figured out in quantum gravity. So that is really the frontier. There’s also a lot of mathematical tools that you can use, right? In, for instance, in particle physics, but also in general relativity, sort of symmetry space to play an enormously important role. And this goes all the way to gauge symmetries as well. And so applying these kinds of symmetries to, uh, machine learning was actually, you know, I thought of it as a very deep and interesting mathematical problem. I did this with Taco Cohen and Taco was the main driver behind this, went all the way from just simple, like rotational symmetries all the way to gauge symmetries on spheres and stuff like that. So, and, uh, Maurice Weiler, who’s also here, um, when he was a PhD student, he was a very good student with me, you know, he wrote an entire book, which I can really recommend about the role of symmetries in AI and machine learning. So I find this a very deep and interesting problem. So more recently, so I’ve taken a sort of different path, which is the relationship between diffusion models and that field called stochastic thermodynamics. This is basically the thermodynamics, which is a theory of equilibrium. So but then formulated for out of equilibrium systems. And it turns out that the mathematics that we use for diffusion models, but even for reinforcement learning for Schrodinger bridges for MCMC sampling has the same mathematics as this theoretical, this physical theory of non-equilibrium systems. And that got me very excited. And actually, uh, when I taught a course in, um, Mauschenberg, uh, it is South Africa, close to Cape Town at the African Institute for Mathematical Sciences Ames. And I turned that into a book site. Two years later, the book was finished. I’ve sent it to the publisher. And this is about the deep relationship between free energy, diffusion models, basically generative AI and stochastic thermodynamics. So it’s always some kind of, I don’t know, I find physics very deep. I also think a lot about quantum mechanics and it’s, it’s, it’s a completely weird theory that actually nobody really understands. And there’s a very interesting story, which is maybe good to tell to connect sort of my PZ back to where I’m now. So I did my PZ with a Nobel Laureate, Gerard the toft. He says the most brilliant man I’ve ever met. He was never wrong about anything as long as I’ve seen him. And now he says quantum mechanics is wrong and he has a new theory of quantum mechanics. Nobody understands what he’s saying, even though what he’s writing down is not mathematically very complex, but he’s trying to address this understandability, let’s say of quantum mechanics head on. And I find it very courageous and I’m completely fascinated by it. So I’m also trying to think about, okay, can I actually understand quantum mechanics in a more mundane way? So that, you know, without all the weird multiverses and collapses and stuff like that. So the physics is always been the threat and I’m trying to apply the physics to the machine learning to build better algorithms.

[01:06:52:16 - 01:07:05:15]

Brandon: You are still very involved in understanding and understanding physics and the worlds. Yeah. And just like applications to machine learning or introducing no formalisms. That’s really cool.

[01:07:05:15 - 01:07:18:02]

Max: Yes, I would say I’m not contributing much to physics, but I’m contributing to the interface between physics and science. And that’s called AI for science or science or AI is kind of a super, it’s actually a new discipline that’s emerging.

[01:07:18:02 - 01:07:18:19]

Speaker 5: Yeah.

[01:07:18:19 - 01:07:45:14]

Max: And it’s not just emerging, it’s exploding, I would say. That’s the better term because I know you go from investments into like in the hundreds of millions now in the billions. So there’s now actually a startup by Jeff Bezos that is at 6.2 billion sheep round. Right. Insane. I guess it’s the largest startup ever, I think. And that’s in this field, AI for science. It tells you something that we are creating a new bubble here.

[01:07:46:15 - 01:07:53:28]

Brandon: So why do you think it is? What has changed that has motivated people to start working on AI for science type problems?

[01:07:53:28 - 01:08:49:17]

Max: So there’s two reasons actually. One is that people have been applying sort of the new tools from AI to the sciences, which is quite natural. And there’s of course, I think there’s two big examples, protein folding is a big one. And the other one is machine learning forest fields or something called machine learning inter-atomic potentials. Both of them have been actually very successful. Both also had something to do with symmetries, which is a little cool. And sort of people in the AI sciences saw an opportunity to apply the tools that they had developed beyond advertised placement, right, or multimedia applications into something that could actually make a very positive impact in society like health, drug development, materials for the energy transition, carbon capture. These are all really cool, impactful applications.

[01:08:50:19 - 01:09:42:14]

Max: Despite that, the science and the kind of the is also very interesting. I would say the fact that these sort of these two fields are coming together and that we’re now at the point that we can actually model these things effectively and move the needle on some of these sort of science sort of methodologies is also a very unique moment, I would say. People recognize that, okay, now we’re at the cusp of something new, where it results whether the company is called after. We’re at the cusp of something new. And of course that always creates a lot of energy. It’s like, okay, there’s something, it’s like sort of virgin field. It’s like nobody’s green field. Nobody’s been there. I can rush in and I can sort of start harvesting there, right? And I think that’s also what’s causing a lot of sort of enthusiasm in the fields.

[01:09:42:14 - 01:10:12:18]

RJ: If you’re an AI engineer, basically if the people that listen to this podcast will be in the field, then you maybe don’t have a strong science background. How does, but are excited. Most I would say most AI practitioners, BM engineers or scientists would consider themselves scientists and they have some background, a little bit of physics, a little bit of industry college, maybe even graduate school that have been working or are starting out. How does somebody who is not a scientist on a day-to-day basis, how do they get involved?

[01:10:12:18 - 01:10:14:28]

Max: Well, they can read my book once it’s out.

[01:10:16:07 - 01:11:05:24]

Max: This is basically saying that there is more, we should create curricula that are on this interface. So I’m not sure there is, also we already have some universities actual courses you can take, maybe online courses you can take. These workshops where we are now are actually very good as well. And we should probably have more tutorials before the workshop starts. Actually we’ve, I’ve kind of proposed this at some point. It’s like maybe first have an hour of a tutorial so that people can get new into the field. There’s a lot out there. Most of it is of course inaccessible, but I would say we will create much more books and other contents that is more accessible, including this podcast I would say. So I think it will come. And these days you can watch videos and things. There’s a huge amount of content you can go and see.

[01:11:05:24 - 01:11:28:28]

Brandon: So maybe a follow-up to that. How do people learn and get involved? But why should they get involved? I mean, we have a lot of people who are of our audience will be interested in AI engineering, but they may be looking for bigger impacts in the world. What opportunities does AI for science provide them to make an impact to change the world? That working in this the world of pure bits would not.

[01:11:28:28 - 01:11:40:06]

Max: So my view is that underlying almost everything is immaterial. So we are focusing a lot on LLMs now, which is kind of the software layer.

[01:11:41:06 - 01:11:56:05]

Max: I would say if you think very hard, underlying everything is immaterial. So underlying an LLM is a GPU, and underlying a GPU is a wafer on which we will have to deposit materials. Do we want to wait a little bit?

[01:12:02:25 - 01:12:11:06]

Max: Underlying everything is immaterial. So I was saying, you know, there’s the LLM underlying the LLM is a GPU on which it runs. In order to make that GPU,

[01:12:12:08 - 01:12:43:20]

Max: you have to put materials down on a wafer and sort of shine on it with sort of EUV light in order to etch kind of the structures in. But that’s now an actual material problem, because more or less we’ve reached the limits of scaling things down. And now we are trying to improve further by new materials. So that’s a fundamental materials problem. We need to get through the energy transition fast if we don’t want to kind of mess up this world. And so there is, for instance, batteries. That’s a complete materials problem. There’s fuel cells.

[01:12:44:23 - 01:13:01:16]

Max: There is solar panels. So that they can now make solar panels with new perovskite layers on top of the silicon layers that can capture, you know, theoretically up to 50% of the light, where now we’re at, I don’t know, maybe 22 or something. So these are huge changes all by material innovation.

[01:13:02:21 - 01:13:47:15]

Max: And yeah, I think wherever you go, you know, I can probably dig deep enough and then tell you, well, actually, the very foundation of what you’re doing is a material problem. And so I think it’s just very nice to work on this very, very foundation. And also because I think this is maybe also something that’s happening now is we can start to search through this material space. This has never been the case, right? It’s like scientists, the normal way of working is you read papers and then you come up with no hypothesis. You do an experiment and you learn, et cetera. So that’s a very slow process. Now we can treat this as a search engine. Like we search the internet, we now search the space of all possible molecules, not just the ones that people have made or that they’re in the universe, but all of them.

[01:13:48:21 - 01:14:42:01]

Max: And we can make this kind of fully automated. That’s the hope, right? We can just type, it becomes a tool where you type what you want and something starts spinning and some experiments get going. And then, you know, outcome list of materials and then you look at it and say, maybe not. And then you refine your query a little bit. And you kind of do research with this search engine where a huge amount of computation and experimentation is happening, you know, somewhere far away in some lab or some data center or something like this. I find this a very, very promising view of how we can sort of build a much better sort of materials layer underneath almost everything. And also more sustainable materials. Our plastics are polluting the planet. If you come up with a plastic that kind of destroys itself, you know, after, I don’t a few weeks, right? And actually becomes a fertilizer. These are things that are not impossible at all. These things can be done, right? And we should do it.

[01:14:42:01 - 01:14:47:23]

RJ: Can you tell us a little bit just generally about CUSBI and then I have a ton of questions.

[01:14:47:23 - 01:14:48:15]

Speaker 5: Yeah.

[01:14:48:15 - 01:17:49:10]

Max: So CUSBI started about 20 months ago and it was because I was worried about I’m still worried about climate change. And so I realized that in order to get, you know, to stay within two degrees, let’s say, we would not only have to reduce our emissions to zero by 2050, but then, you know, another half century or even a century of removing carbon dioxide from the atmosphere, not by reducing your emissions, but actually removing it at a rate that’s about half the rate that we now emit it. And that is a unsolved problem. But if we don’t solve it, two degrees is not going to happen, right? It’s going to be much more. And I don’t think people quite understand how bad that can be, like four degrees, like very bad. So this technology needs to be developed. And so this was my and my co-founder, Chet Edwards, motivation to start this startup. And also because, you know, we saw the technology was ready, which is also very good. So if you’re, you know, the time is right to do it. And yeah, so we now in the meanwhile, we’ve grown to about 40 people. We’ve kind of collected 130 million investment into the company, which is for a European company is quite a lot. I would say it’s interesting that right after that, you know, other startups got even more. So that’s kind of tells you how fast this is growing. But yeah, we are we are now at the we’ve built the platform, of course, but it’s for a series of material classes and it needs to be constantly expanded to new material classes. And it can be more automated because, you know, we know putting LLMs in as the whole thing gets more and more automated. And now we’re moving to sort of high throughput experimentation. So connecting the actual platform, which is computational, to the experiments so that you can get also get fast feedback from experiments. And I kind of think of experiments as something you do at the end, although that’s what we’ve been doing so far. I want to think of it as what I would call a sort of a physics processing unit, like a PPU, right, which is you have digital processing units and then you have physics processing units. So it’s basically nature doing computations for you. It’s the fastest computer known as possible, even. It’s a bit hard to program because you have to do all these experiments. Those are quite, quite bulky. It’s like a very large thing you have to do. But in a way, it is a computation. And that’s the way I want to see it. So I want to you can do computations in a data center and then you can ask nature to do some computations. Your interface with nature is a bit more complicated. But then these things will have to seamlessly work together to get to a new material that you’re interested in. And that’s the vision we have. We don’t say super intelligence because I don’t quite know what it means and I don’t want to oversell it. But I do want to automate this process and give a very powerful tool in the hands of the chemists and the material scientists.

[01:17:49:10 - 01:18:01:02]

Brandon: That actually brings up a question I wanted to ask you. First of all, can you talk about your platform to like whatever degree, like explain kind of how it works and like what you your thought processes was in developing it?

[01:18:01:02 - 01:20:47:22]

Max: Yeah, I think it’s been surprisingly, it’s not rocket science, I would say. It’s not rocket science in the sense of the design and basically the design that, you know, I wrote down at the very beginning. It’s still more or less the design, although you add things like I wasn’t thinking very much about multi-scale models and as the common are rated that actually multi-scale is very important. And the beginning, I wasn’t thinking very much about self-driving labs. But now I think, you know, we are now at the stage we should be adding that. And so there is sort of bits and details that we’re adding. But more or less, it’s what you see in the slide decks here as well, which is there is a generative component that you have to train to generate candidates. And then there is a digital twin, multi-scale, multi-fidelity digital twin, which you walk through the steps of the ladder, you know, they do the cheap things first, you weed out everything that’s obviously unuseful, and then you go to more and more expensive things later. And so you narrow things down to a small number. Those go into an experiment, you know, do the experiment, get feedback, etc. Now, things that also have been more recently added is sort of more agentic sort of parts. You know, we have agents that search the literature and come up with, you know, actually the chemical literature and come up with, you know, chemical suggestions for doing experiments. We have agents which sort of autonomously orchestrate all of the computations and the experiments that need to be done. You know, they’re in various stages of maturity and they can be continuously improved, I would say. And so that’s basically I don’t think that part. There’s rocket science, but, you know, the design of that thing is not like surprising. What is it’s surprising hard to actually build it. Right. So that’s that’s the thing that is where the moat is in the data that you can get your hands on and the and actually building the platform. And I would say there’s two people in particular I want to call out, which is Felix Hunker, who is actually, you know, building the scientific part of the platform and Sandra de Maria, who is building the sort of the skate that is kind of this the MLOps part of the platform. Yeah. And so and recently we also added sort of Aaron Walsh to our team, who is a very accomplished scientist from Imperial College. We’re very happy about that. He’s going to be a chief science officer. And we also have a partnerships team that sort of seeks out all the customers because I think this is one thing I find very important. In print, it’s so complex to do to actually bring a material to the real world that you must do this, you know, in collaboration with sort of the domain experts, which are the companies typically. So we always we only start to invest in the direction if we find a good industrial partner to go on that journey with us.

[01:20:47:22 - 01:20:55:12]

Brandon: Makes a lot of sense. Over the evolution of the platform, did you find that you that human intervention, human,

[01:20:56:18 - 01:21:17:01]

Brandon: I guess you could start out with a pure, you could imagine two directions when you start up making everything purely automatic, automated, agentic, so on. And then later on, you like find that you need to have more human input and feedback different steps. Or maybe did you start out with having human feedback? You have lots of steps and then like kind of, yeah, figure out ways to remove, you know,

[01:21:17:01 - 01:22:39:18]

Max: that is the second one. So you build tools for you. So it’s much more modular than you think. But it’s like, we need these tools for this application. We need these tools. So you build all these tools, and then you go through a workflow actually in the beginning just manually. So you put them in a first this tool, then run this to them or this with sithery. So you put them in a workflow and then you figure out, oh, actually, you know, this this porous material that we are trying to make actually collapses if you shake it a bit. Okay, then you add a new tool that says test for stability. Right. Yeah. And so there’s more and more tools. And then you build the agent, which could be a Bayesian optimizer, or it could be an actual other them, you know, maybe trained to be a good chemist that will then start to use all these tools in the right way in the right order. Yeah. Right. But in the beginning, it’s like you as a chemist are putting the workflow together. And then you think about, okay, how am I going to automate this? Right. For one very easy question you can ask yourself is, you know, every time somebody who is not a super expert in DFT, yeah, and he wants to do a calculation has to go to somebody who knows DFT. And so could you start to automate that away, which is like, okay, make it so user friendly, so that you actually do the right DFT for the right problem and for the right length of time, and you can actually assess whether it’s a good outcome, etc. So you start to automate smaller small pieces and bigger pieces, etc. And in the end, the whole thing is automated.

[01:22:39:18 - 01:22:53:25]

Brandon: So your philosophy is you want to provide a set of specific tools that make it so that the scientists making decisions are better informed and less so trying to create an automated process.

[01:22:53:25 - 01:23:22:01]

Max: I think it’s this is sort of the same where you’re saying because, yes, we want to automate, yeah, but we don’t see something very soon where the chemists and the domain expert is out of the loop. Yeah, but it but it’s a retreat, right? It’s like, okay, so first, you need an expert to tell you precisely how to set the parameters of the DFT calculation. Okay, maybe we can take that out. We can maybe automate that, right? And so increasingly, more of these things are going to be removed.

[01:23:22:01 - 01:23:22:19]

Speaker 5: Yeah.

[01:23:22:19 - 01:24:33:25]

Max: In the end, the vision is it will be a search engine where you where somebody, a chemist will type things and we’ll get candidates, but the chemist will still decide what is a good material and what is not a good material out of that list, right? And so the vision of a completely dark lab, where you can close the door and you just say, just, you know, find something interesting and then it will it will just figure out what’s interesting and we’ll figure out, you know, it’s like, oh, I found this new material to blah, blah, blah, blah, right? That’s not the vision I have. He’s not for, you know, a long time. So for me, it’s really empowering the domain experts that are sitting in the companies and in universities to be much faster in developing their materials. And I should say, it’s also good to be a little humble at times, because it is very complicated, you know, to bring it to make it and to bring it into the real world. And there are people that are doing this for the entire lives. Yeah. Right. And it’s like, I wonder if they scratch their head and say, well, you know, how are you going to completely automate that away, like in the next five years? I don’t think that’s going to happen at all.

[01:24:35:01 - 01:24:39:24]

Max: Yeah. So to me, it’s an increasingly powerful tool in the hands of the chemists.

[01:24:39:24 - 01:25:04:02]

RJ: I have a question. You’ve talked before about getting people interested based on having, you know, sort of a big breakthrough in materials, incremental change. I’m curious what you think about the platform you have now in are sort of stepping towards and how are you chasing the big change or is this like incremental or is there they’re not mutually exclusive, obviously, but what do you think about that?

[01:25:04:02 - 01:26:04:27]

Max: We follow a mixed strategy. So we are definitely going after a big material. Again, we do this with a partner. I’m not going to disclose precisely what it is, but we have our own kind of long term goal. You could call it lighthouse or, you know, sort of moonshot or whatever, but it is going to be a really impactful material that we want to develop as a proof point that it can be done and that it will make it into the into the real world and that AI was essential in actually making it happen. At the same time, we also are quite happy to work with companies that have more modest goals. Like I would say one is a very deep partnership where you go on a journey with a company and that’s a long term commitment together. And the other one is like somebody says, I knew I need a force field. Can you help me train this force field and then maybe analyze this particular problem for me? And I’ll pay you a bunch of money for that. And then maybe after that we’ll see. And that’s fine too. Right. But we prefer, you know, the deep partnerships where we can really change something for the good.

[01:26:04:27 - 01:26:22:02]

RJ: Yeah. And do you feel like from a platform standpoint you’re ready for that or what are the things that and again, not asking you to disclose proprietary secret sauce, but what are the things generally speaking that need to happen from where we are to where to get those big breakthroughs?

[01:26:22:02 - 01:28:40:01]

Max: What I find interesting about this field is that every time you build something, it’s actually immediately useful. Right. And so unlike quantum computing, which or nuclear fusion, so you work for 20, 30, 40 years and nothing, nothing, nothing, nothing. And then it has to happen. Right. And when it happens, it’s huge. So it’s quite different here because every time you introduce, so you go to a customer and you say, so what do you need? Right. So we work, let’s say, on a problem like a water filtration. We want to remove PFAS from water. Right. So we do this with a company, Camira. So they are a deep partner for us. Right. So we on a journey together. I think that the breakthrough will happen with a lot of human in the loop because there is the chemists who have a whole lot more knowledge of their field and it’s us who will help them with training, having a new message. And in that kind of interface, these interactions, something beautiful will happen and that will have to happen first before this field will really take off, I think. And so in the sense that it’s not a bubble, let’s put it that way. So that’s people see that as actual real what’s happening. So in the beginning, it will be very, you know, with a lot of humans in the loop, I would say, and I would I would hope we will have this new sort of breakthrough material before, you know, everything is completely automated because that will take a while. And also it is very vertical specific. So it’s like completely automating something for problem A, you know, you can probably achieve it, but then you’ll sort of have to start over again for problem B because, you know, your experimental setup looks very different in the machines that you characterize your materials look very different. Even the models in your platform will have to be retrained and fine tuned to the new class. So every time, you know, you have a lot of learnings to transfer, but also, you know, the problems are actually different. And so, yes, I would want that breakthrough material before it’s completely automated, which I think is kind of a long term vision. And I would say every time you move to something new, you’ll have to start retraining and humans will have to come in again and say, okay, so what does this problem look like? And now sort of, you know, point the the machine again, you know, in the new direction and then and then use it again.

[01:28:40:01 - 01:28:47:17]

RJ: For the non-scientists among us, me included a bit of a scientist. There’s a lot of terminology. You mentioned DFT,

[01:28:49:00 - 01:29:01:11]

RJ: you equivariance we’ve talked about. Can you sort of explain in engineering terms or the level of sophistication and engineering? Well, how what is equivariance?

[01:29:01:11 - 01:29:55:01]

Max: So equivariance is the infusion of symmetry in neural networks. So if I build a neural network, let’s say that needs to recognize this bottle, right, and then I rotate the bottle, it will then actually have to completely start again because it has no idea that the rotated bottle. Well, actually, the input that represents a rotated bottle is actually rotated bottle. It just doesn’t understand that. Right. If you build equivariance in basically once you’ve trained it in one orientation, it will understand it in any other orientation. So that means you need a lot less data to train these models. And these are constraints on the weights of the model. So so basically you have to constrain the way such data to understand it. And you can build it in, you can hard code it in. And yeah, this the symmetry groups can be, you know, translations, rotations, but also permutations. I can graph neural network, their permutations and then physics, of course, as many more of these groups.

[01:29:55:01 - 01:30:01:08]

RJ: To pray devil’s advocate, why not just use data augmentation by your bottle is in all the different orientations?

[01:30:01:08 - 01:30:58:23]

Max: As an option, it’s just not exact. It’s like, why would you go through the work of doing all that? Where you would really need an infinite number of augmentations to get it completely right. Where you can also hard code it in. Now, I have to say sometimes actually data augmentation works even better than hard coding the equivariance in. And this is something to do with the fact that if you constrain the optimization, the weights before the optimization starts, the optimization surface or objective becomes more complicated. And so it’s harder to find good minima. So there is also a complicated interplay, I think, between the optimization process and these constraints you put in your network. And so, yeah, you’ll hear kind of contradicting claims in this field. Like some people and for certain applications, it works just better than not doing it. And sometimes you hear other people, if you have a lot of data and you can do data augmentation, then actually it’s easier to optimize them and it actually works better than putting the equivariance in.

[01:30:58:23 - 01:31:07:16]

Brandon: Do you think there’s kind of a bitter lesson for mathematically founded models and strategies for doing deep learning?

[01:31:07:16 - 01:31:46:06]

Max: Yeah, ultimately it’s a trade-off between data and inductive bias. So if your inductive bias is not perfectly correct, you have to be careful because you put a ceiling to what you can do. But if you know the symmetry is there, it’s hard to imagine there isn’t a way to actually leverage it. But yeah, so there is a bitter lesson. And one of the bitter lessons is you should always make sure your architecture is scale, unless you have a tiny data set, in which case it doesn’t matter. But if you, you know, the same bitter lessons or lessons that you can draw in LLM space are eventually going to be true in this space as well, I think.

[01:31:47:10 - 01:31:55:01]

RJ: Can you talk a little bit about your upcoming book and tell the listeners, like, what’s exciting about it? Yeah, I should read it.

[01:31:55:01 - 01:33:42:20]

Max: So this book is about, it’s called Generative AI and Stochastic Thermodynamics. It basically lays bare the fact that the mathematics that goes into both generative AI, which is the technology to generate images and videos, and this field of non-equilibrium statistical mechanics, which are systems of molecules that are just moving around and relaxing to the ground state, or that you can control to have certain, you know, be in a certain state, the mathematics of these two is actually identical. And so that’s fascinating. And in fact, what’s interesting is that Jeff Hinton and Radford Neal already wrote down the variational free energy for machine learning a long time ago. And there’s also Carl Friston’s work on free energy principle and active entrance. But now we’ve related it to this very new field in physics, which is called stochastic thermodynamics or non-equilibrium thermodynamics, which has its own very interesting theorems, like fluctuation theorems, which we don’t typically talk about, but we can learn a lot from. And I think it’s just it can sort of now start to cross fertilize. When we see that these things are actually the same, we can, like we did for symmetries, we can now look at this new theory that’s out there, developed by these very smart physicists, and say, okay, what can we take from here that will make our algorithms better? At the same time, we can use our models to now help the scientists do better science. And so it becomes a beautiful cross-fertilization between these two fields. The book is rather technical, I would say. And it takes all sorts of things that have been done as stochastic thermodynamics, and all sorts of models that have been done in the machine learning literature, and it basically equates them to each other. And I think hopefully that sense of unification will be revealing to people.

[01:33:42:20 - 01:33:44:05]

RJ: Wait, and when is it out?

[01:33:44:05 - 01:33:56:09]

Max: Well, it depends on the publisher now. But I hope in April, I’m going to give a keynote at ICLR. And it would be very nice if they have this book in my hand. But you know, it’s hard to control these kind of timelines.

[01:33:56:09 - 01:33:58:19]

RJ: Yeah, I’m looking forward to it. Great.

[01:33:58:19 - 01:33:59:25]

Max: Thank you very much.

Through the dozens of midsize launches today (see the rest of the recaps below), one theme that we’re seeing is something I’ve come to call “closing the loop”:

In the DevTools industry, we often talk about the concept of the “inner loop” (what happens in the IDE, “squashed” inside a git commit), vs the “outer loop” (what happens in the cloud, “between” git commits, say after every git push). For a long time in the 2010s, all the action was in the “outer loop” - GitLab and multiple CI/CD unicorns, GitHub, Netlify/Vercel, and more, whereas the “inner loop” was generally free and open source (often thanks to Microsoft, but increasingly Facebook and Atlassian/Vercel) - VSCode, Language Servers, TypeScript, Prettier, and so on.

With AI, things flipped dramatically - Copilot kicked off the wave of Inner Loop coding agent applications that eventually became vanguarded by Cursor, whereas Outer Loop tools did grow thanks to AI, but remained relatively unchanged (or AI SRE or AI Data Engineering as a category seemed to particularly struggle).

Today many of the leading coding agent companies took steps in the “Close the Outer Loop” direction. For a lead graphic the first thing that springs to mind is the og:image for the Devin Autofix launch from a few weeks ago:

The above only applies to review comments, but you can close the outer loop in many, many other ways. Cursor’s highly anticipated big ship today was a combination of full computer use and video output to give you high bandwidth review of the results rather than just looking at code diffs:

Claude Code, celebrating it’s first anniversary (as we discussed on today’s special podcast and short retrospective), launched Remote Control for Max users, which allows people to close the loop in their phones, away from the Inner Loop of the terminal:

Last but not least, Cognition launched Devin 2.2, where a headline feature was how Devin Review now automatically closes the loop back into Devin runs.

If you are so inclined you could view all these as generalizations of the “Ralph Loops” that were popular in social media last month, often still with humans in the loop, but increasingly less so over time.

Don’t miss the other midsize launches today that would otherwise have been title stories in their own right - Notion Custom Agents, Inception Mercury 2, and the smaller Qwen 3.5 variants.

Frontier model ecosystem: Qwen 3.5 “medium series” and open-weight momentum

• Qwen 3.5 Medium Model Series: Alibaba released a tightly scoped set of “more intelligence, less compute” models—Qwen3.5-Flash (hosted), Qwen3.5-35B-A3B (MoE), Qwen3.5-122B-A10B (MoE), and Qwen3.5-27B (dense)—arguing that architecture + data + RL can outperform sheer parameter scaling. Notable details include Flash defaulting to 1M context and built-in tools in the hosted offering. See the full announcement and links to Hugging Face/ModelScope/APIs from @Alibaba_Qwen.

  • Early practitioner reactions emphasize how strong 35B-A3B and 122B-A10B feel in practice (e.g., @andrew_n_carr, @JustinLin610), plus the “intelligence-per-watt” implication of a 35B model surpassing a 235B predecessor noted by @awnihannun.

  • Deployment/serving stack is moving fast: community tooling quickly followed—GGUF + sizing guidance from @UnslothAI and local-run enthusiasm like “35B-A3B is all you need” from @terryyuezhuo. Qwen also highlighted SGLang support (tweet).

  • Quant + “local frontier” trendline: INT4 variants appeared (duplicate posts) via @HaihaoShen, and users continue pushing aggressive quantization workflows (e.g., Unsloth praise for ultra-low-bit local Qwen by @0xSero).

  • Evaluation signals: Qwen’s flagship Qwen3.5-397B-A17B trended on HF (@Ali_TongyiLab) and showed up strongly on agentic webdev-style evaluation in Code Arena (Arena post). Arena also posted rank deltas vs Qwen 3.0 (comparison).

OpenAI + Anthropic “coding agents as product surface area” (APIs, remote control, web sockets, proof-of-work UX)

• OpenAI: GPT-5.3-Codex in the Responses API: OpenAI shipped GPT-5.3-Codex to all developers via the Responses API (announcement), with pricing cited by @scaling01 ($1.75 input / $14 output as tweeted). OpenAI also expanded file input types (docx/pptx/csv/xlsx/etc.) for agents ingesting “real-world files” directly (tweet).

  • Infra detail: web sockets show up as a meaningful lever for agent throughput—“30% faster rollouts” per @gdb. This matches broader chatter about why websockets took time and how state is stored upstream vs VRAM (thread, follow-up).

  • Benchmarks: third-party scoreboard posts claim strong placements for Codex 5.3 across TerminalBench/IOI/LiveCodeBench/VibeCodeBench (ValsAI).

• Anthropic: “Claude Code Remote Control” + enterprise workflow push: Anthropic introduced “Remote Control” for Claude Code—start a terminal session locally and continue from your phone—first via @noahzweben, then officialized by @claudeai, with rollout confirmation from @_catwu.

  • Separate enterprise positioning: “Cowork and plugin updates” for customizing Claude across teams landed with extremely high engagement (@claudeai).

• Cursor: “review is demo videos, not diffs”: Cursor announced a major UX pivot—agents can use the software they build, then send videos of their work (“demos, not diffs”) (launch, links). Multiple builders describe cloud agents as a practical step-change: async, VM-based testing, self-verification, and demo artifacts (example, another, “creative director over sims”).

Diffusion for language: Inception Labs Mercury 2 and “speed as the next battleground”

• Mercury 2 (“reasoning diffusion LLM”): Inception Labs released Mercury 2, positioning it as a production diffusion LLM hitting ~1,000 output tokens/s (Stefano Ermon). Artificial Analysis contextualizes it as not frontier-leading on intelligence, but unusually strong on output speed with decent agentic/coding evaluations, including comparisons on Terminal-Bench Hard and IFBench scoring claims (analysis thread).

• The deeper takeaway across these posts: teams are betting that architecture-level parallel token refinement (diffusion) can make multi-step agent loops and voice assistants feel “native” rather than “batchy” (see the architectural explanation from @LiorOnAI). This sits alongside broader sentiment that 2026 competition may be defined by latency + throughput, not just raw benchmark maxima.

Agents: reliability, safety failures, memory + context rot, and new multilingual evals

• Agent reliability is not keeping pace with capability: A Princeton-led effort formalizes and measures the capability–reliability gap, decomposing reliability into 12 dimensions and finding only modest reliability gains despite large capability gains (paper + dashboard; additional commentary from @random_walker). This aligns with recurring “long tail of failures” intuition from practitioners comparing agents to AVs (ahall_research).

• OpenClaw and “routine-step decomposition” safety bypass: A concrete agent failure mode: “split a dangerous command into a few routine steps → safety is gone,” with inbox-wiping behavior cited; authors claim an open-source fix (paper thread).

• AGENTS.md (and equivalents) can hurt: Two high-signal posts summarize research showing LLM-generated context files decrease success while increasing costs; developer-written minimal context helps slightly but still increases cost. See @omarsar0 for the paper summary and @_philschmid for a practical “how to write it” guide grounded in the same result set.

• New SWE-bench Multilingual leaderboard: A push to evaluate software engineering agents beyond English/Python. The leaderboard covers 300 tasks in 9 languages, none from SWE-bench Verified, with reported SOTA at 72% (launch; more stats from @KLieret). The implication: model rankings can invert across languages—important for global dev tooling and for data-collection strategies.

Data + benchmarks: OCR saturation, “new optimizer” skepticism, and adaptive/continual data pitches

• OCR/document parsing benchmarks saturating: Multiple posts argue OmniDocBench is hitting a ceiling (e.g., ~95% with failures on real documents) and that exact-match metrics penalize semantically correct parses. See @llama_index and @jerryjliu0. Related: confusion at why OCR remains hard despite cheap synthetic data (gabriberton) and a study suggesting text extraction beats image representations for PDF QA (cwolferesearch).

• “Nature MI optimizer” controversy: A highly technical critique calls out suspicious baselines and potential test-set hyperparameter selection in a new optimizer paper with dramatic plots, urging independent validation and better-tuned baselines (e.g., nanogpt speedrun) (giffmana; plus additional experimental context from @YouJiacheng).

• Adaption Labs: “Adaptive Data”: Several tweets pitch a shift from static datasets to a “living asset” loop, with claims of 82% average quality gains across 242 languages and an early access/community program (company; additional framing from @sarahookr; third-party paraphrase here). Treat as a directional thesis (data drift/feedback loops) rather than a validated standard until more methodology is public.

Compute, chips, and robotics: Meta–AMD megadeal, MatX’s “HBM+SRAM” bet, and scaling humanoid control

• Meta ↔ AMD infrastructure deal: Meta announced a multi-year agreement to integrate AMD Instinct GPUs with ~6GW planned data center capacity for the deployment (@AIatMeta). Commentary frames it as a major capex/compute signal on the eve of NVIDIA earnings (kimmonismus).

• MatX “One” accelerator: MatX announced a $500M Series B and pitched a chip architecture combining systolic-array efficiency with better utilization on smaller matrices, aiming for high throughput and low latency, explicitly addressing long-context workloads via HBM while preserving SRAM-first latency characteristics (reinerpope). Karpathy highlights the “two memory pools” constraint (SRAM vs DRAM/HBM) and frames memory+compute orchestration as a core puzzle for upcoming token demand (karpathy).

• Liquid AI LFM2-24B-A2B: Liquid AI released LFM2-24B-A2B, a 24B MoE with ~2.3B active/token, optimized for efficiency and edge inference in a 32GB footprint (launch). Distribution arrived quickly across Ollama (tweet) and LM Studio (tweet).

• Robotics scaling: NVIDIA SONIC (GEAR-SONIC): A standout robotics thread claims a 42M-parameter policy trained on 100M+ mocap frames and 500k+ parallel simulated robots, transferring zero-shot to a real humanoid with 100% success across 50 sequences; code/weights are open (Jim Fan thread, plus links here). The key “systems” claim is that dense supervision from motion tracking acts like a scalable analogue to next-token prediction for whole-body control.

• Claude Code Remote Control rollout: @claudeai

• Qwen 3.5 Medium Model Series release: @Alibaba_Qwen

• Cursor agents ship “demos not diffs”: @cursor_ai

• Karpathy on CLIs as agent-native interface: @karpathy

• Meta–AMD 6GW infrastructure deal: @AIatMeta

• Mercury 2 diffusion LLM launch: @StefanoErmon

• NVIDIA SONIC humanoid control (open source): @DrJimFan

• MatX chip + $500M Series B: @reinerpope

• AGENTS.md research summary (context can hurt): @omarsar0

• OpenAI GPT-5.3-Codex in Responses API: @OpenAIDevs