Daily Prompt with Archer and Iris - Daily AI News

Daily Prompt with Archer & Iris — Friday, May 29, 2026

Anthropic had a two-headline day: it released Claude Opus 4.8 — with a new "dynamic workflows" tool that runs hundreds of parallel subagents and can carry a codebase-scale migration from kickoff to merge — and closed a $65 billion Series H at a $965 billion post-money valuation, overtaking OpenAI as the most valuable AI startup. Apple is building a standalone Siri app to take on ChatGPT, with a rebuilt assistant running on Google's Gemini under the hood, shipping with iOS 27 at WWDC in June. Plus: CNN sues Perplexity over alleged verbatim copying, Visa invests in Replit for agentic payments, Asana buys no-code agent-builder StackAI for $75M, and exchanges start designing futures markets for AI compute.

Stories

- Anthropic releases Claude Opus 4.8 — new "Dynamic Workflows" tool (research preview in Claude Code) runs hundreds of parallel subagents and can take codebase-scale migrations across hundreds of thousands of lines from kickoff to merge against an existing test suite. ~4x less likely than Opus 4.7 to let code flaws pass unremarked; more likely to flag uncertainty and avoid unsupported claims (Bridgewater cited). New Effort Control slider; pricing unchanged at $5/M input, $25/M output. Released 41 days after Opus 4.7.

https://www.anthropic.com/news/claude-opus-4-8

https://techcrunch.com/2026/05/28/anthropic-releases-opus-4-8-with-new-dynamic-workflow-tool/

- Apple builds a standalone Siri app to rival ChatGPT, powered by a rebuilt model using Google's Gemini under the hood. Type or talk, upload documents and photos, persistent chat history; invoked from the Dynamic Island and Spotlight. Ships with iOS 27 at WWDC in June 2026. Apple's ~2.5B active devices vs. ChatGPT's ~900M weekly users.

https://techcrunch.com/2026/05/28/sneak-peek-at-new-siri-app-reveals-apples-plans-to-take-on-chatgpt-and-more/

- Anthropic raises $65B Series H at a $965B post-money valuation, overtaking OpenAI as the most valuable AI startup (OpenAI ~$730B per Axios). Co-led by Altimeter, Dragoneer, Greenoaks, Sequoia, Capital Group, Coatue, and D1; strategic partners Samsung, SK Hynix, Micron; includes $15B of committed hyperscaler money ($5B from Amazon, April 2026). Run-rate revenue crossed $47B in May; framed as a likely final private round before an IPO.

https://techcrunch.com/2026/05/28/anthropic-raises-65-billion-nears-1t-valuation-ahead-of-ipo/

https://www.axios.com/2026/05/28/anthropic-ai-fundraising-openai

- CNN sues Perplexity in New York federal court, alleging its tools produce "verbatim" copies of CNN journalism and serve paywalled content for free; says Perplexity copied thousands of stories, videos, and images. First AI copyright suit by a TV network; CNN tried and failed to license last year. Perplexity also faces suits from the NYT, Reddit, and Dow Jones.

https://www.theverge.com/ai-artificial-intelligence/938893/cnn-perplexity-ai-copyright-lawsuit

- Visa invests in Replit to power agentic payments — developers and their AI agents can accept payments inside the platform via Visa Intelligent Commerce and the Trusted Agent Protocol, which lets agents prove their identity and verify payments. Over 1,000 Visa employees already build on Replit.

https://techcrunch.com/2026/05/28/visa-invests-in-replit-to-power-agentic-payments-for-developers/

- Asana acquires no-code agent-builder StackAI for $75M to become "the operating system for human-agent teams." StackAI (YC W23, ~$20M raised) builds no-code agents that pull from Salesforce, Slack, and Google Suite to automate workflows.

https://techcrunch.com/2026/05/28/asana-acquires-no-code-agent-builder-stack-ai/

- Exchanges design futures markets for AI compute: the Shanghai Futures Exchange is building derivatives on AI tokens while CME and ICE launch GPU compute futures, letting data-center operators hedge compute cost like gold or oil — treating inference tokens as a raw-material input.

https://techcrunch.com/2026/05/28/just-like-gold-and-oil-well-soon-be-able-to-trade-ai-token-futures/

Daily Prompt with Archer & Iris — Thursday, May 28, 2026

The Model Context Protocol locked a release candidate for its 2026-07-28 spec — its biggest revision since launch. Session management is gone, so MCP servers can run behind a plain load balancer; extensions become first-class (including server-rendered UI in a sandbox), three older features are deprecated (with a 12-month removal window), sign-in is aligned to OAuth/OpenID Connect, and there are breaking changes with a 10-week window for SDKs to catch up. China is increasingly keeping its best AI talent at home — researchers and founders now need government approval to travel abroad, Manus AI's co-founders were barred from leaving during a review of Meta's $2B bid, and labs like Moonshot, StepFun, and ByteDance reportedly need sign-off to take U.S. capital — as the top U.S.–China model gap narrows to 2.7% (from ~31% in 2023). And OpenRouter raised a $113M Series B led by CapitalG at a $1.3B valuation (double a year ago), now serving 8M users and ~100 trillion tokens/month across 400+ models. Plus: Snowflake's $6B AWS Graviton CPU deal, Epoch's single-number Capabilities Index, ClickHouse tripling revenue to $250M on an IPO track, and researchers hijacking voice assistants with sounds humans can't hear.

Stories

- Model Context Protocol locks its 2026-07-28 release candidate — the biggest revision since launch. Stateless core (session handshake and session header removed; servers run behind ordinary round-robin load balancers). Extensions become first-class, with two official ones: MCP Apps (server-rendered HTML in sandboxed iframes) and a reworked Tasks extension. Authorization hardened to align with OAuth 2.0 / OpenID Connect. Three features deprecated (Roots, Sampling, Logging) under a new lifecycle policy with a 12-month minimum removal window. Breaking changes; RC locked May 21, 2026, final spec July 28, 2026, with a 10-week SDK validation window.

https://blog.modelcontextprotocol.io/posts/2026-07-28-release-candidate/

- China is increasingly keeping its best AI talent — and capital — to itself. Top researchers and founders now need government approval to travel abroad; in March 2025 authorities advised leading founders to avoid the U.S. Manus AI's co-founders were barred from leaving during a regulatory review of Meta's $2B acquisition attempt. Moonshot AI, StepFun, and ByteDance reportedly need sign-off before accepting American capital. Backed by 2025 rare-earth export controls and a ban on foreign AI chips in state-funded data centers. The top U.S.–China model performance gap narrowed to 2.7% as of March 2026, down from roughly 31% in 2023.

https://techcrunch.com/2026/05/27/china-is-increasingly-keeping-its-best-ai-talent-to-itself/

- OpenRouter more than doubles to a $1.3B valuation on a $113M Series B led by CapitalG (Alphabet's growth fund); prior post-money ~$547M in June 2025. The AI gateway serves 8 million users and ~100 trillion tokens/month (~25T/week, 5x in six months) across 400+ models from providers including Anthropic, Google, OpenAI, xAI, and DeepSeek. Series A was $40M (June 2025, a16z and Menlo, with Sequoia). Founded 2023.

https://techcrunch.com/2026/05/26/openrouter-more-than-doubles-valuation-to-1-3b-in-a-year/

- Snowflake signs a $6B, five-year AWS deal for Graviton CPUs (not GPUs) — roughly all the revenue Snowflake has made via AWS Marketplace since 2012 ($7B total). CPU demand rises as AI shifts from training to deployment and agents. AWS's Andy Jassy cites better price-performance; Nvidia's Jensen Huang counters that his new Vera CPU is a $200B market with $20B already sold. Meta also signed a Graviton deal.

https://techcrunch.com/2026/05/27/in-more-good-news-for-amazon-snowflake-signs-6b-deal-with-aws-for-ai-cpu-chips/

- Epoch's Capabilities Index (ECI) folds 40+ benchmarks into a single, saturation-resistant capability scale. Calibrated so Claude 3.5 Sonnet = 130 and GPT-5 = 150, with harder benchmarks weighted more and a minimum of 4 evaluations per model. Domain-specific variants exist for software engineering and math. Funded by Google DeepMind researchers but an independent Epoch AI product.

https://epoch.ai/eci

- ClickHouse triples annualized revenue to $250M (high-nine-digits expected by end of 2026), at a $15B valuation set in January 2026 after a $400M Series D led by Dragoneer (60x+ revenue multiple). The open-source database for AI-scale data has 4,000+ customers including Anthropic, Meta, Capital One, and Decagon. Hired ex-Snowflake exec Jimmy Sexton as CFO; IPO expected within a few years. Tech originated at Yandex 17 years ago; spun out 2021.

https://techcrunch.com/2026/05/27/clickhouse-triples-annualized-revenue-to-250m-charting-a-path-toward-an-ipo/

- Researchers hijack AI voice assistants with inaudible "adversarial audio." A Zhejiang University / Singapore team led by Meng Chen embedded sounds imperceptible to humans into podcasts, videos, and songs that trigger voice assistants to run unauthorized commands (potentially reaching photos and bank data). ~30 minutes to train an attack signal, then reusable at will. Requires model weights, so it targets open-weight systems and commercial products built on them (Microsoft, Mistral named). Presented at the IEEE Symposium on Security and Privacy, May 2026.

https://futurism.com/artificial-intelligence/hackers-inaudible-recordings-hijack-ai-voice-chatbots

Daily Prompt with Archer & Iris — Wednesday, May 27, 2026

Cognition raised a $1B round at a $25B pre-money valuation — Devin's annualized revenue is now $492M and enterprise usage has been growing 50% month-over-month for six straight months. Robinhood opened a beta that lets AI agents trade stocks through a Model Context Protocol service, with their own dedicated wallets and a virtual agentic credit card. OpenAI and Anthropic split publicly on the AI jobs apocalypse — Anthropic's Chris Olah warned at the Vatican of "very large scale" labor displacement, while Sam Altman told Australia's Commonwealth Bank CEO he was "delighted to be wrong" about white-collar layoffs. Plus: YouTube starts auto-labeling photorealistic AI videos, ElevenLabs Music v2 can switch genres mid-track, DuckDuckGo installs jump 28-30% as users opt out of Google AI Mode, and Box CEO Aaron Levie coins "AI psychosis" for executives who confuse a clean demo with a clean rollout.

Stories

- Cognition closes $1B at $25B pre-money — led by Lux Capital and General Catalyst. Post-money $26B (up from $10.2B in September 2025). Annualized run-rate $492M. Enterprise usage growing 50% month-over-month for six months. Customers include Mercedes-Benz, NASA, Goldman Sachs, and Santander. Founders Fund and 8VC re-upped; Ribbit, Atreides, and Layer Global new on the cap table. Competitive set named: Claude Code, Codex, Jules.

https://techcrunch.com/2026/05/27/ai-coding-startup-cognition-raises-1b-at-25b-pre-money-valuation/

- Robinhood opens stock trading to AI agents (beta). Users create a separate Robinhood account for the agent, fund a dedicated wallet, and expose portfolio data, analyst notes, sector exposure, and concentration risk via an MCP service. Some trades require user approval; stocks at launch with options, crypto, event contracts, futures, and prediction markets on the roadmap. Companion agentic credit card available to Gold Card holders in beta. VP of product Abhishek Fatehpuria: "We've heard a lot of demand from our customers to bring their own tools, LLMs, and agents, and connect them to Robinhood."

https://techcrunch.com/2026/05/27/robinhood-now-lets-your-ai-agents-trade-stocks/

- OpenAI and Anthropic dig in against each other on the AI jobs apocalypse. Anthropic co-founder Chris Olah at the Vatican's AI ethics conference: "There is a real possibility that AI will displace human labor at very large scale." Sam Altman to Commonwealth Bank of Australia CEO Matt Comyn: "I'm delighted to be wrong about this, I thought there would have been more impact on entry-level white-collar jobs being eliminated by now than has actually happened." Displacement-side data: ~8,000 Meta layoffs, restructurings at Coinbase, Block, Pinterest, Shopify; AI walk-backs at Uber and Microsoft. Non-displacement data: Stanford research finds losses concentrated in low-AI-exposure sectors, SWE openings up 18% YoY, LinkedIn counts ~1.3M new postings attributed to AI.

https://www.axios.com/2026/05/27/ai-hype-doom-openai-anthropic

- YouTube automatically labels photorealistic AI videos. Triggered by C2PA metadata or YouTube's own internal signals. Labels under the player for long-form, overlay on Shorts. Will not affect recommendations or monetization. Creators can dispute disclosure status but cannot remove labels on content made with Veo, Dream Screen, or content carrying C2PA metadata. Rollout began May 2026.

https://techcrunch.com/2026/05/27/youtube-will-now-automatically-label-ai-videos/

- ElevenLabs ships Music v2 — mid-track genre switching, section-by-section editing, prompt-driven region regeneration. Built on licensed training data (cleared for commercial use), an explicit contrast with Suno and Udio, both in copyright litigation. Live on ElevenCreative and ElevenMusic; API access "coming soon."

https://techcrunch.com/2026/05/27/elevenlabss-new-music-generation-model-can-switch-genres-mid-track/

- DuckDuckGo installs jump 28-30% as users reject Google AI Mode. U.S. app installs averaged 18.1% week-over-week, peaked at 30.5% on May 25. iOS averaged 33% (peak 69.9%). noai.duckduckgo.com averaged 22.7% growth. Apptopia independently confirmed ~29% lift in daily U.S. downloads. Spike survived Memorial Day weekend. CEO Gabriel Weinberg: "Google is force-feeding AI with no way to opt out... their results are getting worse, not better."

https://techcrunch.com/2026/05/26/duckduckgo-installs-are-up-30-as-users-reject-being-force-fed-googles-ai-search/

- Box CEO Aaron Levie coins "AI psychosis" — the executive version. "CEOs are uniquely prone to AI psychosis because they're sufficiently distant from the last mile of work." Case study: ClickUp's Zeb Evans deployed ~3,000 internal AI agents, cut 22% of staff, called the result a "100x org." Berkeley, MIT, NBER, and HBR studies all show measured productivity gains lag perceived gains. Tech layoffs: 115,430 across 152 companies in the first five months of 2026, with AI most cited as the justification.

https://techcrunch.com/2026/05/27/tech-ceos-are-apparently-suffering-from-ai-psychosis/

Daily Prompt with Archer & Iris — Tuesday, May 26, 2026

Pope Leo XIV released his first encyclical, Magnifica Humanitas, on May 15 — naming five ways AI could warp humanity and presenting it Monday at the Vatican alongside Anthropic co-founder Chris Olah. Microsoft Research shipped Fara 1.5, a family of open-weight computer-use agents (4B/9B/27B) where the 27B variant scores 72.0% on Online-Mind2Web — beating OpenAI Operator (58.3%) and Google Gemini Computer Use (57.3%) on the same benchmark. And Axios reports CISA — the U.S. civilian cyber agency — has lost about a third of its staff since 2025, has no chief AI officer, and was excluded from early access to Anthropic's specialized cybersecurity model Mythos. Plus: Apple credits Claude on a macOS kernel CVE, ClickUp lays off 22% with million-dollar salary bands for survivors, and TestingCatalog spots Anthropic's Mythos-1 preview build in Claude Code and Claude Security.

Stories

- Pope Leo XIV's first encyclical, Magnifica Humanitas: On safeguarding the human person in the time of artificial intelligence, released May 15, 2026 — exactly 135 years after Leo XIII's Rerum Novarum. Names five concerns: eroding judgment, simulating care without relationship, deepening inequality, destabilizing democracy, facilitating warfare. Anthropic co-founder Chris Olah spoke alongside the Pope at the May 25 Vatican presentation.

https://www.anthropic.com/news/chris-olah-pope-leo-encyclical

https://www.axios.com/2026/05/25/pope-leo-xiv-ai-humanity-war-jobs-warning

https://techcrunch.com/2026/05/25/the-popes-ai-encyclical-isnt-really-about-ai/

- Microsoft Research ships Fara 1.5 — open-weight browser computer-use agents in 4B/9B/27B parameter variants, all built on Qwen3.5. Fara 1.5-27B scores 72.0% on Online-Mind2Web (300 tasks, 136 sites) vs. OpenAI Operator at 58.3% and Google Gemini 2.5 Computer Use at 57.3%. The 9B variant is live today on Microsoft Foundry; inference harness on GitHub at microsoft/fara.

https://www.microsoft.com/en-us/research/articles/fara1-5-computer-use-agent/

- Axios: CISA sidelined as the White House scrambles on AI cyber threats. Agency has lost ~1/3 of its workforce since early 2025; Trump's proposed budget cuts another $707M and 766 FTE positions. Acting Director Nick Andersen leads a weakened organization with no chief AI officer. CISA never received initial access to Anthropic's Mythos cybersecurity model despite other agencies getting early access. Quotes from former DHS undersecretary Suzanne Spaulding and Senator Gary Peters.

https://www.axios.com/2026/05/26/cisa-white-house-cybersecurity-ai

- Apple credits Claude on CVE-2026-28952 — a kernel integer overflow patched in macOS Tahoe 26.5 (released May 11, 2026). Discoverers listed as "Calif.io in collaboration with Claude and Anthropic Research." First-party Apple security advisory crediting an AI model as co-discoverer of a kernel CVE.

https://support.apple.com/en-us/127115

- ClickUp lays off 22% of its workforce (May 23, 2026). CEO Zeb Evans frames it as AI transformation, says "most savings... will flow directly back into the people who stay" with new "million-dollar salary bands" and a plan to "gamify value created and time saved." ~3,000 internal AI agents absorbing work. Gartner: 80% of companies using autonomous technology have already cut jobs.

https://techcrunch.com/2026/05/25/what-clickups-mass-layoff-tells-us-about-the-future-of-work/

- TestingCatalog spots "claude-mythos-1-preview" in Anthropic UI strings and source code — a dedicated cybersecurity model class distinct from the main Claude family, attached to Claude Code and Claude Security. Project Glasswing reportedly used Mythos-class models to find more than 10,000 high- or critical-severity vulnerabilities in essential software.

https://www.testingcatalog.com/anthropic-prepares-mythos-1-for-claude-code-and-claude-security/

Daily Prompt with Archer & Iris — Monday, May 25, 2026

An autonomous LLM agent resolved nine open Erdős problems and forty-four OEIS conjectures with machine-checkable Lean proofs. Google's API key revocation was found to lag up to twenty-three minutes — two developers ate four- and five-figure surprise Gemini bills on keys originally deployed for Google Maps. Nvidia's Nemotron Labs shipped open-weight diffusion language models at 3B/8B/14B, with ~4× the throughput of autoregressive baselines on a B200. Plus: Anthropic ships 31 Skills for small businesses, Cursor hits $3B annualized ahead of a SpaceX-led round, California signs a first-in-the-nation AI worker EO, memory becomes 63% of AI chip component cost, and Anthropic's finance team reports 10–20 hours per week reclaimed with Claude.

Stories

- Autonomous LLM agent resolves 9 of 353 open Erdős problems and 44 of 492 OEIS conjectures via formal Lean proofs; deployed across combinatorics, optimization, graph theory, algebraic geometry, and quantum optics; multi-author preprint led by George Tsoukalos posted Thursday

https://arxiv.org/abs/2605.22763

- Google API key revocation lags up to 23 minutes after deletion; Aikido Security's Joseph Leon measured >90% attacker success rate inside the window; two developers (Rod Danan of Prentus, Isuru Fonseka in Sydney) hit with $10,138 and ~AUD $17,000 surprise Gemini bills on Maps-scoped keys; refunds only after press coverage

https://techcrunch.com/2026/05/24/everyone-is-navigating-ai-security-in-real-time-even-google/

- Nvidia Nemotron Labs releases open-weight diffusion language models — 3B/8B/14B text + 8B vision; 2.6× tokens/forward-pass in diffusion mode, 6–6.4× in self-speculation; ~865 tok/s on B200; commercial-friendly license; targets code, math, summarization, fill-in-the-middle, and single-query latency-sensitive workloads

https://huggingface.co/blog/nvidia/nemotron-labs-diffusion

- Anthropic releases 31 "Skills for small businesses"; community reports ~382,000 downloads on day one; third-party setup workflow claims ~10-minute deploy of the full pack

https://www.reddit.com/r/ClaudeAI/comments/1tktssn/skills_for_small_businesses_officially_released/

- Cursor reaches $3B annualized revenue ahead of a SpaceX-led financing round (not yet closed)

https://www.bloomberg.com/news/articles/2026-05-21/cursor-hits-3-billion-annual-sales-rate-ahead-of-spacex-deal

- California Governor Gavin Newsom signs first-in-the-nation executive order on AI's labor impacts; mandates a sectoral AI impact dashboard, a WARN Act revision within 180 days, training and unified benefits portal modernization, and a public-good incentive framework

https://www.gov.ca.gov/2026/05/21/governor-newsom-signs-first-of-its-kind-executive-order-to-prepare-workers-and-businesses-for-potential-ai-disruption/

- Memory reaches 63% of AI chip component cost (up from 52% in Q1 2024); HBM spend ~$12B (2024) → ~$32B (2025); Microsoft adds ~$25B to FY26 capex, Meta adds $10B, both citing higher component prices

https://epoch.ai/data-insights/ai-chip-component-cost-shares

- Anthropic finance team profile: Alice Fong reports 10–20 hours per week reclaimed using Claude for variance commentary, board deck reconciliation, and Excel reference tracing

https://claude.com/blog/how-anthropics-finance-team-uses-claude-to-shape-the-narrative-behind-the-numbers

Ep. 59 — Deep Dive: Then Someone Had to Teach It How to Talk to Us

Daily Prompt with Archer & Iris — 2026-05-24

Part two of a weekend pair. Saturday was the cast list — Mira Murati's path from Vlore to OpenAI to Thinking Machines Lab, and the recorded deposition that played in a Bay Area courtroom on May 6. Sunday is the technical half. What the lab actually shipped, what it argues OpenAI is doing wrong, and what any of it means if you spend your Monday building a voice product.

On May 11, 2026, Thinking Machines Lab released the limited research preview of TML-Interaction-Small — a 276-billion-parameter mixture-of-experts with 12 billion active parameters, trained from scratch around a single thesis: "For interactivity to scale with intelligence, it must be part of the model itself."

The technical claim is that today's voice products are wrapped in a "harness" — speech recognition feeds an LLM, the LLM feeds text-to-speech, and external timing logic decides when each runs. The model in the middle never sees the conversation as a stream; it sees it as turns. TML-Interaction-Small ingests audio, video, and text as continuous streams in 200-millisecond micro-turn chunks, with input and output on the same clock. It listens while it speaks. It sees while it listens. A second model — slow, asynchronous, sharing context — handles long thinking in the background.

The lab cites Rich Sutton's bitter lesson to argue hand-crafted external interaction systems (voice-activity detectors, turn-end classifiers) will be outpaced by general scaling.

The numbers, as the lab published them:

- FD-bench v1.5 (interaction quality — interruption, backchanneling, talking to others, background speech): TML 77.8 vs GPT-Realtime-2 (highest reasoning setting) 47.8 — a 30-point gap on the metric the lab is optimizing for.

- Turn-taking latency: TML 0.40 s vs GPT-Realtime-2 1.18–1.63 s.

- Audio MultiChallenge (audio reasoning, outside benchmark): TML 43.4% vs GPT-Realtime-2 (xhigh) 48.5%. OpenAI's model is stronger on raw audio reasoning.

- FD-bench v3 tool-use: TML (with background model) 68.0% pass@1 vs GPT-Realtime-2 58.0%.

The lab is calling it a research preview. Wider release is planned for later in 2026. The small foreground model alone cannot do long reasoning or web search.

Live stakes thread: the same week the model dropped, Murati's recorded deposition in Musk v. Altman became public. The jury rejected the Musk lawsuit. The transcript stayed. Co-founders Barret Zoph (CTO) and Luke Metz returned to OpenAI on January 14, 2026, with Sam Schoenholz. Andrew Tulloch left for Meta in October 2025. The lab is thinner. The lab is shipping. The thesis is intact.

— Key facts —

- TML-Interaction-Small released May 11, 2026 (research preview)

- 276B-parameter MoE, 12B active parameters

- Processes audio + video + text in 200 ms micro-turn chunks; encoder-free early fusion

- Foreground interaction model + asynchronous background reasoning model with shared context

- FD-bench v1.5: 77.8 (TML) vs 47.8 (GPT-Realtime-2 highest reasoning)

- Turn-taking latency: 0.40 s vs 1.18–1.63 s

- Audio MultiChallenge: 43.4% (TML) vs 48.5% (GPT-Realtime-2 xhigh)

- FD-bench v3 tool-use pass@1: 68.0% vs 58.0%

- Deposition video played in Musk v. Altman: May 6, 2026 (Musk lawsuit ultimately rejected)

- Founders gone: Tulloch (Meta, Oct 2025); Zoph + Metz + Schoenholz (back to OpenAI, Jan 14, 2026)

- Seed round: $2B at $12B valuation, July 2025 (Andreessen Horowitz lead; Accel, Nvidia, AMD, Jane Street, government of Albania)

— Sources —

- Thinking Machines Lab — Interaction Models (research blog): https://thinkingmachines.ai/blog/interaction-models/

- the-decoder — Thinking Machines Lab ships its first model and argues interactivity is what OpenAI gets wrong about voice: https://the-decoder.com/thinking-machines-lab-ships-its-first-model-and-argues-interactivity-is-what-openai-gets-wrong-about-voice/

- DataCamp — Interaction Models: What TML-Interaction-Small Gets Right: https://www.datacamp.com/blog/thinking-machines-interaction-models

- StartupHub — AI's Next Leap: Interaction Models: https://www.startuphub.ai/ai-news/technology/2026/ai-s-next-leap-interaction-models

- StartupHub — Thinking Machines Lab Wants to Replace OpenAI Realtime With a Model That Listens While It Speaks: https://www.startuphub.ai/ai-news/artificial-intelligence/2026/thinking-machines-lab-interaction-models-mira-murati-2026

- Rich Sutton — The Bitter Lesson (2019): http://www.incompleteideas.net/IncIdeas/BitterLesson.html

— Format —

Deep Dive (Sunday). Three-angle debate. Part two of a weekend pair.

Tomorrow: the daily briefing returns.

Ep. 58 — Deep Dive: Everyone Wanted a Mind in the Machine

Daily Prompt with Archer & Iris — 2026-05-23

On the sixth of May, a recorded deposition played in a Bay Area courtroom. Mira Murati — former OpenAI CTO, now founder of Thinking Machines Lab — testified that Sam Altman did not tell her the truth about whether the company's legal team had cleared a model to bypass the internal safety board. The lawsuit she was testifying in went on to be rejected by the jury. The deposition is still in the record.

Saturday is the backstory. Sunday is the live stakes.

This episode traces how we got here. Vlore in 1988. Albania switching economies before Murati finished middle school. Math at Colby. Engineering at Dartmouth — the same campus where John McCarthy and Marvin Minsky named "artificial intelligence" in 1956. Goldman Sachs in Tokyo. Tesla Model X. Leap Motion (the gesture-control startup whose obsession with making screens listen to fingers foreshadows everything she's building today). OpenAI in 2018. CTO by May 2022. ChatGPT, DALL-E, Codex, Sora, GPT-4 — all shipped under her.

Then the November 2023 board crisis. Three days as interim CEO. Emmett Shear on Sunday. Altman back on Wednesday. She stayed ten more months, then left.

Thinking Machines Lab files paperwork in February 2025. A $2B seed round at a $12B valuation in July, led by Andreessen Horowitz, with Accel, Nvidia, AMD, Jane Street — and the government of Albania. The name itself is a deliberate reach back to Danny Hillis's 1983 Thinking Machines Corporation, the Cambridge company that built the Connection Machine and went bankrupt in 1994.

The thesis: build AI systems that are interpretable, customizable, collaborative.

The thesis came under pressure fast. Andrew Tulloch left for Meta in October 2025. On January 14, 2026, co-founders Barret Zoph (CTO) and Luke Metz announced their return to OpenAI; Sam Schoenholz went with them. OpenAI CEO Fidji Simo confirmed the move 58 minutes after Murati's own X post.

This month, the lab published a paper arguing the next leap in AI is not raw model intelligence — it is interactivity. The line worth quoting: "For interactivity to scale with intelligence, it must be part of the model itself."

That is the bridge to tomorrow.

— Key facts —

- Mira Murati: born December 16, 1988 in Vlore, Albania

- Colby (BA Mathematics, 2011); Dartmouth Thayer (BE Mechanical Engineering, 2012)

- Tesla Model X PM (2013); Leap Motion (2016–2018)

- OpenAI 2018–2024; CTO from May 2022

- November 2023: interim CEO of OpenAI during the "Blip"

- Stepped down September 2024

- Thinking Machines Lab founded February 2025

- $2B seed at $12B valuation, July 2025 (Andreessen Horowitz lead; Accel, Nvidia, AMD, Jane Street, government of Albania)

- January 14, 2026: co-founders Barret Zoph and Luke Metz return to OpenAI; Sam Schoenholz also returns

- May 6, 2026: video deposition played in Musk v. Altman

- Musk's lawsuit ultimately rejected by a California federal jury

— Sources —

- Wikipedia — Mira Murati: https://en.wikipedia.org/wiki/Mira_Murati

- Fortune — Meet Mira Murati: https://fortune.com/2025/10/03/mira-murati-career-ai-thinking-machines-goldman-sachs-tesla-leap-openai/

- The Vertical — Who is Mira Murati: https://thevertical.la/entrepreneurs/who-is-mira-murati-the-albania-born-founder-of-openais-new-rival-thinking-machines-lab/

- Futurism — Under threat of perjury, OpenAI's former CTO is admitting some very interesting stuff about Sam Altman: https://futurism.com/artificial-intelligence/perjury-openai-cto-murati-altman

- Newsbytes — Sam Altman created distrust, ex-OpenAI CTO Mira Murati tells court: https://www.newsbytesapp.com/news/science/sam-altman-created-distrust-ex-openai-cto-mira-murati-tells-court/story

- Let's Data Science — Mira Murati testifies Sam Altman misled her: https://letsdatascience.com/news/mira-murati-testifies-sam-altman-misled-her-82ce8a15

- AI Certs — Murati's testimony intensifies OpenAI lawsuit turmoil: https://www.aicerts.ai/news/muratis-testimony-intensifies-openai-lawsuit-turmoil/

- TechCrunch — Thinking Machines Lab is losing two co-founders to OpenAI: https://techcrunch.com/2026/01/14/mira-muratis-startup-thinking-machines-lab-is-losing-two-of-its-co-founders-to-openai/

- StartupHub — TML wants to replace OpenAI Realtime with a model that listens while it speaks: https://www.startuphub.ai/ai-news/artificial-intelligence/2026/thinking-machines-lab-interaction-models-mira-murati-2026

- MindStudio — What is Thinking Machines Labs: https://www.mindstudio.ai/blog/what-is-thinking-machines-labs-mira-murati

- Wikipedia — Thinking Machines Corporation: https://en.wikipedia.org/wiki/Thinking_Machines_Corporation

- Wikipedia — Dartmouth workshop (1956): https://en.wikipedia.org/wiki/Dartmouth_workshop

- Wikipedia — Computing Machinery and Intelligence (Turing, 1950): https://en.wikipedia.org/wiki/Computing_Machinery_and_Intelligence

— Format —

Deep Dive (Saturday). Narrative mode. Part one of a weekend pair.

Tomorrow: Then Someone Had to Teach It How to Talk to Us — the technical half of the same story. What changed when models started listening while they spoke, and what the live stakes from this week's testimony actually mean for builders.

Daily Prompt with Archer & Iris — Friday, May 22, 2026

A poisoned VS Code extension breached GitHub from a single employee laptop. The White House postponed an AI security executive order at the door after Mark Zuckerberg, Elon Musk, and David Sacks called the president. Google launched Gemini Omni Flash — multimodal inputs, video output, free on YouTube before the API ships. Plus: Cohere drops Command A+ open weights, Anthropic opens a Claude Compliance API with 28 security partners, Spotify and Universal Music open licensed AI covers, and Hark closes a $700M Series A with no product.

Stories

- GitHub: poisoned NX Console VS Code extension compromised an employee device on May 18; attackers claim ~3,800 internal repos exfiltrated; no customer repos affected; critical secrets rotated Monday into Tuesday

https://github.blog/security/investigating-unauthorized-access-to-githubs-internal-repositories/

- Trump postpones AI security executive order at signing; would have required a 90-day pre-release review window; Zuckerberg, Musk, and David Sacks spoke with the president Wed–Thu; Sacks called it "just something doomers wanted"; order motivated by Anthropic's Mythos and OpenAI's GPT-5.5-Cyber

https://www.axios.com/2026/05/21/trump-ai-executive-order-postponed-why

https://techcrunch.com/2026/05/21/trump-delays-ai-security-executive-order-i-dont-want-to-get-in-the-way-of-that-leading/

- Google launches Gemini Omni Flash — first model in the Omni family; multimodal inputs (image, audio, video, text), video output today, image and audio outputs planned; multi-turn editing, reference blending, avatar creation, physics handling; live now for AI Plus/Pro/Ultra in Gemini app and Flow; free on YouTube Shorts/Create; API "coming in the coming weeks"

https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-omni/

- Cohere ships Command A+ — open-weights MoE under Apache 2.0; 218B total / 25B active; 128K input context; 48 languages; Tau-2 Telecom 37%→85%, Terminal-Bench Hard 3%→25%; runs on 2× H100 or 1× B200 at W4A4

https://cohere.com/blog/command-a-plus

- Anthropic opens a Claude Compliance API with 28 security/DLP partners — Cloudflare, CrowdStrike, Wiz, Datadog, Microsoft Purview, Okta, Palo Alto Networks, Snyk, Tenable, Zscaler and more; surfaces Claude Enterprise chats/files/projects plus an activity event stream

https://claude.com/blog/compliance-api-security-partners

- Spotify and Universal Music strike a licensed deal for fan-made AI covers and remixes; Premium add-on; revenue share with opted-in artists; no artist list, no launch date, no price; Spotify co-CEO Alex Norström: "consent, credit, and compensation"

https://techcrunch.com/2026/05/21/spotify-and-universal-music-strike-deal-allowing-fan-made-ai-covers-and-remixes/

- Hark raises $700M Series A at $6B; Parkway leads; Nvidia, AMD, Intel, Qualcomm, Salesforce all participate; founder Brett Adcock (Figure.AI, Archer Aviation); 70 employees, Nvidia B200 data center; first multimodal models this summer; custom hardware after; no product, no demo, no customers

https://techcrunch.com/2026/05/21/hark-raises-700m-series-a-for-its-secretive-universal-ai-interface/

Daily Prompt with Archer & Iris — Thursday, May 21, 2026

Anthropic projects its first profitable quarter at $10.9B in revenue, while signing a $1.25B/month, $40B+ compute deal with xAI through 2029. Nvidia posts a record $81.6B quarter and discloses $43B in private startup holdings. Meta and Intuit cut a combined 11,000 jobs the same day. Plus: OpenAI's reasoning model disproves a 1946 Erdős conjecture (this time with named mathematicians backing it up), a September IPO target, Figma adds AI agents to its canvas, Stability Audio 3.0 ships 6-minute songs, and Palantir sues over the DIA's eight-year in-house build.

Stories

- Anthropic projects first operating profit at ~$10.9B Q2 revenue; signs $1.25B/month, $40B+ compute deal with xAI through May 2029 (Colossus 1, 300 MW, near Memphis)

https://techcrunch.com/2026/05/20/anthropic-says-its-about-to-have-its-first-profitable-quarter/

https://techcrunch.com/2026/05/20/anthropic-will-pay-xai-1-25-billion-per-month-for-compute/

- Nvidia posts $81.6B revenue ($75.2B data center, record); guidance $91B next quarter; $80B buyback; $43B in private equity holdings (up from $22B in January); $30B committed to OpenAI; new Anthropic buildout

https://techcrunch.com/2026/05/20/nvidia-posts-another-record-quarter-reveals-43-billion-of-holdings-in-startups/

- Jensen Huang names Vera CPU as a "brand new $200 billion TAM"; $20B in Vera sales already

https://techcrunch.com/2026/05/20/jensen-huang-says-hes-found-a-brand-new-200b-market-for-nvidia/

- Meta begins layoffs of ~8,000 (10% of workforce) in three regional waves; Intuit cuts 3,000 (17% of 18,200) citing "simplifying" and AI focus

https://www.businessinsider.com/layoff-meta-severance-details-cobra-jobs-2026-5

https://techcrunch.com/2026/05/20/intuit-to-lay-off-over-3000-employees-to-refocus-on-ai/

- OpenAI reasoning model disproves a 1946 Erdős conjecture on unit distance graphs; verified by Noga Alon, Melanie Wood, and Thomas Bloom

https://techcrunch.com/2026/05/20/openai-claims-it-solved-an-80-year-old-math-problem-for-real-this-time/

- OpenAI targeting September IPO with Goldman Sachs and Morgan Stanley

https://techcrunch.com/2026/05/20/openai-barrels-toward-ipo-that-may-happen-in-september/

- Figma launches AI agents on its multiplayer canvas; Q1 revenue $333.4M, +46% YoY

https://techcrunch.com/2026/05/20/figma-adds-an-ai-assistant-to-its-collaborative-canvas/

- Stability Audio 3.0 — four models, up to 6:20 of generated music, open weights for small/medium, paid Large, commercial license above $1M revenue

https://techcrunch.com/2026/05/20/stability-ai-release-a-new-audio-model-that-can-create-six-minute-songs/

- Palantir protests its exclusion from a DIA contract after the agency spent eight years building MARS in-house

https://www.axios.com/2026/05/21/palantir-pentagon-dia-contract-suit

Daily Prompt with Archer & Iris — Ep. 55

Date: 2026-05-20

I/O Lands, Karpathy Crosses the Parking Lot, and the White House Drafts a Pre-Release Approval

In this episode:

- Google I/O 2026 ships Gemini 3.5 (76.2 on Terminal-Bench, ~4x faster output) into the Gemini app, AI Mode in Search, Antigravity 2.0, the Gemini API, and Gemini Enterprise — and redesigns the search box for the first time in 25 years.

Source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5/

- Andrej Karpathy joins Anthropic's pre-training team under Nick Joseph, leading a new team that uses Claude itself to accelerate pre-training research.

Source: https://techcrunch.com/2026/05/19/openai-co-founder-andrej-karpathy-joins-anthropics-pre-training-team/

- Trump preparing an AI executive order asking Anthropic, OpenAI, Google, and xAI to share frontier models with the government 90+ days before public release — triggered by Anthropic's Mythos and OpenAI's GPT-5.5-Cyber.

Source: https://www.axios.com/2026/05/20/ai-trump-executive-order-white-house-infighting

Quick hits:

- KPMG embeds Claude across 276,000+ employees in 138 countries — Claude Cowork and Managed Agents inside KPMG Digital Gateway. https://www.anthropic.com/news/anthropic-kpmg

- OpenAI adopts Google's SynthID watermarking and the C2PA standard for image provenance; verification tool in preview. https://techcrunch.com/2026/05/19/openai-is-making-it-easier-to-check-if-an-image-was-made-by-their-models/

- D.C. Circuit hears Anthropic's appeal of the Pentagon "supply chain risk" designation — Judge Henderson calls it "a spectacular overreach." https://www.axios.com/2026/05/19/anthropic-trump-administration-court-arguments

- Gemini 3.1 Flash TTS launches — 1211 Elo, 70+ languages, Audio Tags, native multi-speaker, SynthID-watermarked. https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-1-flash-tts