Braid

Braid

By Lenar Kess · Damra VolNewsTechnology
Download on the App Store

All Braid episodes

  • Twenty Billion Parameters, One Big Harness

    A twenty-billion-parameter model claiming frontier-level search, a recipe that says to train the harness as hard as the weights, and a week of releases where the interesting part keeps living in the scaffolding around the model rather than in the model itself. Lenar and Damra follow that thread from agent architecture down to the hardware you can own — and up to the courts and committees that decide where any of it is allowed to touch the record.

    • Patrick Jiang's Harness-1 post — a 20B search agent trained with a "state-externalizing harness" that he claims rivals Opus-4.6; the architecture, not the parameter count, is the claim worth examining.
    • Viv's "agent = model + harness" recipe — train both components together; the same specialization logic shows up everywhere this week.
    • Nate on one-shotting a full-stack app and Jon Shulkin on Grok Build — orchestration as the product, with the model treated as a commodity.
    • CRUX's agent publishing an iOS app — "a few human interventions" is the detail that decides whether open-world evals beat pass/fail scores.
    • Sem — code-understanding entities built on Git history, not a language server; the structured store a harness would actually lean on.
    • Universal Memory Protocol vs Databricks' end-to-end Instructed Retriever — standardize memory, or specialize retrieval for a 3x win? The incentives point opposite ways.
    • NVIDIA's RTX Spark at Korea's PC Bangs and the GLM Air/GGUF thread — the local crowd wants the smallest good-enough model on hardware they own.
    • UK police told to stop using AI for court statements and the AGI-economics conversation — when intelligence gets cheap, trust is the scarce resource nobody can manufacture.
    17 min
  • When the Harness Carries the Model

    An open-weights model that fumbles tool calls on its own can go toe to toe with a frontier closed model — once you wrap the right error-handling around it. That gap, between what a model scores and what it does inside your repo, runs through everything we covered today.

    • Ahmad Awais on Latent Space describes "tool confusion" — open models repeating the same invalid tool call roughly fifty-six times per billion tokens — and Command Code's deterministic repair layer that patches malformed output instead of arguing with the model. The claim that reframes the day: the harness, not the weights, decides whether a cheap model is usable.
    • DeepSeek V4 Flash support in llama.cpp (PR #24162) makes the same model runnable locally — but the repair layer that makes it pleasant stays behind Command Code's API. Access to weights isn't access to the experience.
    • Knowledge Activation (Bakal et al.) argues AI skills should be the institutional-knowledge unit for agentic development; Mutation Without Variation warns that repeated LLM edits converge rather than diverge — together a hint that skill files plus a converging model could homogenize a codebase.
    • Agents' Last Exam, SentinelBench, and Stability vs. Manipulability in LLM judges all poke at the same wound: our scores have drifted from the work, especially for long-running and judge-graded evaluation.
    • Anthropic's "When AI builds itself" (via a thin Reddit summary) claims AI is accelerating its own development; a zero-knowledge verification paper offers a cryptographic path to actually check claims like that — and the pause proposals that depend on verification.
    • The Washington Post (Elizabeth Dwoskin), via Techmeme, reports an FDA fast track for digital health tech including AI chatbots — the same model behavior that costs a retry in coding costs a patient in a clinic.
    18 min
  • What the Mug Lets You Do

    A strange Friday: no launch, no valuation, just a wall of version-one arXiv preprints. Read together, they rhyme — robots reasoning about what objects let you do instead of what they look like, policies fighting the latency tax of diffusion, and agents that change themselves mid-run. Lenar and Damra hold all of it at preprint altitude: these are claims from serious groups, graded on their own benchmarks.

    • What Objects Enable, Not What They Are — A4D organizes a robot's latent space around function ("movable") rather than appearance ("cart"), reporting 94% accuracy and a discovery step that flags when it doesn't know. Convergent with AffordanceVLA, which decomposes manipulation into which/where/how-to-act.
    • Flash-WAM cuts a robot action chunk from 8.1 seconds to 348 ms (a 23x speedup) via modality-aware distillation — while Let It Be Simple argues the fancy distillation was never the hard part for low-dimensional policies. EVE and MIRAGE chase the same wall-clock budget from other seats.
    • HANDOFF distills a humanoid whole-body controller from three specialists; Open-H-Embodiment opens the largest medical-robot dataset to date, where the lead surgical model finishes a structured suturing task on just 25% of trials — the only model above zero.
    • The Meta-Agent Challenge finds agents-building-agents real but mediocre, and surfaces reward-hacking like ground-truth exfiltration under pressure. TMEM edits weights online; Trivium argues for an inspectable causal log instead; CHARM tackles cascading hallucination across RAG steps.
    • Inference-Time Vulnerability Beyond Shallow Safety shows a mid-sequence injection at any step can flip safety behavior, and that internal "refusal-aligned" states don't predict robustness — so alignment has to train on the generation trajectory, not just outputs.
    20 min
  • The Substation and the Zoning Board

    The binding constraint in AI stopped being the model and became physical: a fab that can't keep up, a grid that has to find ten reactors' worth of power, and a neighbor who can file a lawsuit. We follow that collision through chips, a rare moment of rival unity, an IPO, a clogged courtroom, and the parts of the world building around scarcity.

    • TSMC's C.C. Wei (via Bloomberg) says the company can't fill US demand even as Arizona capacity comes online — scarcity admitted by the supplier, not the buyers.
    • France's €110B+ AI buildout (Sarah White, FT) amounts to ~10 gigawatts — an energy-policy decision dressed as a tech investment.
    • SpaceX's $55B Terafab tax exemption (Stephanie Findlay, FT) draws local legal threats — the abstraction of compute now has an address.
    • Rival labs co-sign a bioweapons letter (Robert Hart, The Verge) — but the screening that actually bites sits with DNA-synthesis firms, who aren't signing.
    • Anthropic's path to IPO (Madhumita Murgia, FT) puts a quarterly clock on a safety posture that private capital used to subsidize.
    • Courts coping with AI lawsuits (Michelle Kim, MIT Tech Review) — hallucinated citations are cheap to produce and expensive to refute.
    • Scarcity-driven innovation (Rest of World) and AI as new colonialism (Axios) describe the same engineer as protagonist and subject.
    • Can generalist agents automate data curation? and StepPRM-RTL push agents into the senior-human judgment calls — and make per-step checking the valuable part.
    19 min
  • Permission Slips and Poured Concrete

    A stack of European filings wants to triple data center capacity and own more of the AI stack — on the same day a JP Morgan report says the country building fastest can't pour its own concrete on schedule. Lenar and Damra trace the day's real constraint: not model quality, but megawatts, transformers, capital, and rights.

    • The EU's Cloud and AI Development Act (CADA) aims to triple data center capacity in 5–7 years, paired with a tech-sovereignty communication and open-source strategy and a Chips Act 2.0 — a statement of intent about which layers of the stack Europe wants to own.
    • JP Morgan, via the WSJ, says 60%+ of US data center capacity planned for 2027 isn't yet under construction — the build-out is power- and permit-bound, not building-bound.
    • Alibaba's Qwen 3.7 Plus ships multimodal with a one-million-token window at $2 per million tokens, and DeepSeek is raising ~$7.4B from Tencent and battery maker CATL — energy money following the compute story.
    • Microsoft's on-device Aion 1.0 Instruct and Plan models split instruction-following from planning, while a llama.cpp build report shows reproducible local gains on two 3090s.
    • AURA argues the key-value cache is wrong for robots and proposes constant-memory action-gated retention; a second paper tries to measure harmful overthinking in reasoning models.
    • GitLab is cutting 350 staff and exiting 22 countries under an AI-pivot framing, and the UK CMA is forcing Google to let publishers opt out of AI search summaries separately from search itself.
    19 min
  • Eighty Billion and the Ideas Underneath

    The day's news ran on a single tension: enormous sums are being raised to fund the AI buildout, while the question of whether the capability and the margins follow stays unanswered. Lenar and Damra trace the money from Alphabet's filings to Anthropic's IPO paperwork, then down into the tooling, the chips, and one paper about ideas no human is positioned to have.

    • Alphabet's $80bn equity raise — a profitable company choosing to dilute shareholders rather than borrow, with $10bn going to Berkshire Hathaway, signals how hard the compute commitment is to walk back.
    • Anthropic's confidential IPO filing lands as corporate America hits "AI sticker shock" — and Anthropic's biggest customers are the companies tightening those budgets.
    • Knowledge workers are now ~1/5 of OpenAI Codex users, growing three times faster than developers — moving code generation to people who can't always read the output.
    • Cloudflare's Agents SDK v0.14.0 ships durable workflows, schedules, and skills — the difference between an agent you operate and a worker you delegate to.
    • China adds data and algorithms to its trade-secret rules while military-linked universities seek Nvidia H200 chips and Arm names Oracle and ByteDance as data-center CPU customers.
    • "Alien Science" samples research directions that are coherent but cognitively unavailable — logical ideas no community is positioned to propose.
    18 min
  • Cheaper From Both Ends

    A Chinese lab cut the price of a frontier-class coding model to a fraction of Opus, Nvidia tried to own every layer from the laptop to the data center, and one developer ran the new Gemma 4 on a decade-old Xeon. The cost of running intelligence got attacked from both ends on the same morning — and the question underneath all of it is who gets to set that cost.

    • MiniMax M3 claims parity with Opus 4.7 at roughly twelve cents per million input tokens versus five dollars — but the weights are promised in about ten days, so "open-weights" is still a countdown.
    • Nvidia's DGX Station puts a GB300 chip and up to 748GB of memory on a desktop, enough to run a one-trillion-parameter model locally; the RTX Spark chip pushes the same idea into laptops, while the Vera CPUs — with Anthropic, OpenAI, and SpaceX as early customers — signal a move off x86.
    • A 10-year-old Xeon is all you need: cafkafk runs a 26B mixture-of-experts model at reading speed on a 2016 CPU with no GPU, arguing mainstream tools hide the performance levers.
    • Cosmos 3 is Nvidia's open physical-AI world model, backed by a Cosmos Coalition with Runway as a founding member.
    • Cadence and Nvidia claim a "Level 5" autonomous chip-verification agent that turns months into a day — a large autonomy claim in a domain where mistakes ship in silicon.
    • Anthropic will let the EU's ENISA join Project Glasswing for access to a model called Mythos, even as a Wirescreen analysis documents 500+ PLA attempts to procure Nvidia chips and governments from India and the UAE to France move to own their compute.
    20 min
  • Who Holds the Dial

    A frontier model gets called a step toward God in one window and a judgmental token-burner in the next. We spend the morning on the gap between the marketing altitude and the desk, and find the same thread running through everything: every layer now has a control surface someone's reaching for.

    • Dylan Field on Opus 4.8 calls it "a very strange model" — honesty up, curiosity down, personality judgmental — a reminder that a tuning dial has costs you can feel.
    • scaling01 on DeepSWE says GPT-5.5 "score-, time- and token-mogged" Opus 4.8, putting the efficiency column — the one that pays your bill — back in the conversation.
    • Ben Kunkle on Zed's Zeta 2 shows how a ten-second editing pause becomes a training label, and how a million frontier-model calls got replaced by a self-grading student model.
    • Philipp Schmid (DeepMind) on the five assumptions that trip up senior engineers building agents — errors as inputs, evals not unit tests, and "build to delete."
    • Komi-learn and a year on knowledge-graph memory share one missing thing: a controlled before-and-after proving the memory layer, not the model, made the agent better.
    • A Lancet correspondence finds 4,046 fabricated references across 2,810 published articles — model honesty rising while the literature's integrity falls.
    • Quick hits: AMD's Lisa Su vs Nvidia's Jensen Huang on China, IBM's Sovereign Core, and a court ordering Circle to freeze a $12.6M contract.
    19 min
  • The number nobody optimized for

    Claude Opus 4.8 landed overnight with a math score that leapt and a business-ops score that fell — and reading the release honestly means distrusting the chart. Lenar and Damra work through the gap between the number that moved and the number that matters, then chase it into agent budgets, the protocol wars, local-inference tooling, Mistral's on-prem bet, and the power grid.

    • A scrape of 100+ Opus 4.8 evals shows USAMO 2026 jumping 69%→97% while Vending-Bench 2 nearly halved — a retune that helped some distributions and hurt others.
    • "AI benchmarks are useless" argues the record scores ride on elaborate prompt setups: change a few prompt words and results swing 10–20 points.
    • The BAGEN study finds frontier agents can't estimate their own remaining budget mid-task — which collides with enterprises trying to rein in "tokenmaxxing" (WSJ via Techmeme).
    • "MCP is dead?" gets a sharp rebuttal from OpenAI's Max Stoiber: nearly every company is building an MCP server, even ones with no CLI or external API.
    • Multi-token prediction benchmarks hit ~3.3x faster local inference; llama.cpp got a real website and antirez shipped distributed inference.
    • Notes from the Mistral AI Now Summit — on-prem KYC at BNP Paribas, against a comment that Mistral's 120B "small" model loses to models a quarter its size. xAI countered with a one-dollar coding model.
    • FERC's June grid-connection proposal is the duller, realer infrastructure story next to an unsourced TerraFab "one terawatt" claim.
    19 min
  • Locally coherent, globally not

    Friday's room sits between a hobbyist voice assistant running entirely on Mario Zechner's desk and a cluster of arXiv papers all saying the same thing from different angles: long-running agents now fall apart in ways the model can't fix. Lenar and Damra read four reliability papers side by side, then turn to the personal-memory question every shipping assistant is already getting wrong.

    • Mario Zechner on pibot — full local voice loop with Parakeet, Qwen 3 TTS, and Qwen 3.6 through llama.cpp, with the STT and TTS engines ported from Python into Rust on mlx-c. The runtime detail is the news, not the model lineup.
    • Ethan Mollick on token budgets — split spend between building and learning. Read against yesterday's Kirkland and Ellis platform story, the question becomes who controls the learning budget at internal AI orgs.
    • MMPO — Ziyan Liu and team train a policy that decides when memory in long-horizon agents should be rewritten and when it should be left alone. Belief drift comes from over-eager rewrites, not missing updates.
    • RedundancyBench — Minyang Hu's group benchmarks how many steps in a long agent trajectory are repeats. Stale duplicates of state crowd out the relevant signal in context.
    • Locally Coherent, Globally Incoherent — Anany Kotawala's single-author paper bounds compositional incoherence in multi-component agents. Defensible local outputs assemble into contradictory global ones.
    • Agent-Radar — Hongxiang Zhang's group steers attention toward context-relevant tokens in multi-agent communication, so the receiver isn't drowned in noise from the sender.
    • Selective QA over conflicting personal memory — Tiancheng Yang's testbed for what happens when your assistant's memories about you disagree. No single resolution strategy dominates.
    • BioRefusalAudit — Caleb DeLeeuw uses sparse autoencoders to ask whether a model's refusal is shallow pattern matching or whether the dangerous capability isn't there at all.
    • AutoformBot and Atlas — Ahmad Rammal's team at FAIR Paris and NYU on a multi-agent system that pulls textbook math into Lean 4 at scale. Lean is the verifier the agents can't argue with.
    23 min