Braid

The recant, the runtime, and a Pantheon built in code

A corporate takedown answered with a recant letter and a mirror in Germany, the protocols and computers agents actually run on, six tools trying to build the Pantheon in code, and a paper where the model writes its own GPU kernel. Plus Codex learning to keep going, a security tool hardened against the real world, and a graduation room that cheered for human intelligence. * Meta emails Heretic; Heretic recants [https://www.reddit.com/r/LocalLLaMA/comments/1tjmvx6/heretic_has_been_served_a_legal_notice_by_meta_inc/] — a takedown of abliterated Llama derivatives answered with a Galileo joke and a Codeberg mirror in Germany. * Five hundred PRs a day, and the harness that triages them [https://www.youtube.com/watch?v=VaS2h-dY1-4] — Onur Solmaz on OpenClaw, acpx, and the Agent Client Protocol. * The computer the agent runs on [https://www.youtube.com/watch?v=kaX43RRRUKY] — Ivan Burazin of Daytona on stateful, composable machines for agents and 74% month-over-month growth. * Building the Pantheon, in code [https://modelrift.com/blog/openscad-llm-benchmark/] — six coding tools tackle parametric CAD, and the gap between a good preview and a clean export. * When the model writes its own kernel [https://arxiv.org/abs/2605.19269] — CODA folds memory-bound ops into the matrix multiply, and model-authored kernels keep up with human ones. * Codex learns to keep going [https://www.youtube.com/watch?v=rgh0hMYPcd0] — goal mode graduates, plus Appshots and shared plugins. * Hardening the thing that reads your CI config [https://x.com/trailofbits/status/2057782296527208709] — Trail of Bits stress-tests zizmor against forty-one thousand real workflows. * The headcount bet [https://libertas.software/en/knowledge-hub/19/the-companies-cutting-headcount-for-ai-will-lose-to-the-ones-who-didnt] — and a graduation room that cheered for actual intelligence [https://www.businessinsider.com/steve-wozniak-apple-ai-graduation-speech-2026-5].

Two bets on AGI, an 80-year-old problem, and Anthropic in the black

Google's I/O keynote is a day behind us, and the week it kicked off turned into a referendum on two very different bets on artificial general intelligence — plus a pile of counter-programming from everyone else. Today: OpenAI cracking an 80-year-old math problem with a general-purpose model, Anthropic's first profitable quarter and what Karpathy was actually hired to do, a 70-page paper on why frontier models still can't tell a fact from a labeled lie, Midjourney's hardware regret, ads arriving inside Google's AI answers, Meta's layoffs, Cohere's open-weights comeback, and a field guide to skilling up coding agents. * Two bets on the same finish line [https://www.youtube.com/watch?v=o_av1b9rs2g] — Google's world-model road vs OpenAI's text-reasoning road, in the labs' own words. * OpenAI cracks an 80-year-old problem [https://openai.com/index/model-disproves-discrete-geometry-conjecture/] — the planar unit distance result from a general-purpose reasoning model. * Anthropic in the black, and Karpathy's bet [https://www.cnbc.com/2026/05/20/anthropic-revenue-explosive-growth-ipo-profitable-quarter.html] — ~$559M operating profit and a hire aimed at recursive self-improvement. * Jagged intelligence, and the false story [https://www.youtube.com/watch?v=o_av1b9rs2g] — the paper where models believe a story they were told a thousand times was fake. * Midjourney's hardware regret [https://www.reddit.com/r/singularity/comments/1tiut2d/midjourney_says_their_research_was_set_back_by_a/] — the tooling tax of betting on the less-supported accelerator. * Ads come to AI Mode [https://blog.google/products/ads-commerce/google-marketing-live-search-ads/] — the business model under the consumer bet. * Meta's eight thousand [https://nypost.com/2026/05/20/business/meta-kicks-off-bloodbath-with-8000-layoffs-in-shift-to-ai/] — the cost side, on the same clock as the wins. * Cohere comes back, Apache-licensed [https://www.reddit.com/r/LocalLLaMA/comments/1tizmar/re_what_ever_happened_to_coheres_commanda_series/] — Command A+, a mixture-of-experts model that fits on one or two GPUs. * Skilling up the agent [https://www.youtube.com/watch?v=vNCY9kXXyDQ] — Marc Klingen's concrete lessons on teaching a coding agent to wire up your tool. * Who's training whom [https://www.reddit.com/r/singularity/comments/1tjgm3x/every_office_employee_is_training_their_own/] — the anxiety running underneath the week.

Foothills, and the morning Karpathy moved

Google I/O 2026 landed yesterday — Gemini Omni, Gemini 3.5 Flash, Antigravity 2.0, Spark, and Demis Hassabis closing the keynote on the "foothills of the singularity." About forty minutes before he walked on stage, Andrej Karpathy tweeted that he'd joined Anthropic. The rest of the day was the labs sorting themselves around both events. Today's show works through the announcements, the pricing shifts, the keynote demo that boots Doom, the Railway outage that happened while Google was selling Spark, and a builder's 100K-line Rust postmortem that's a sharper picture of agentic coding than anything on the I/O stage. * Hassabis: "foothills of the singularity" — DeepMind's CEO compresses his AGI timeline on stage [https://www.prismnews.com/news/google-deepmind-chief-says-ai-marks-foothills-of-the] * Gemini 3.5 Flash specs and pricing — and what the 3x bump means [https://www.latent.space/p/ainews-google-io-2026-gemini-35-flash] * Gemini Omni's physics pitch versus the same-day backflip test [https://www.reddit.com/r/singularity/comments/1thohgl/gemini_omni_model_is_still_unable_to_make_someone/] * Antigravity 2.0's 93-agent OS demo: 12 hours, 2.6B tokens, under $1K, boots Doom [https://www.latent.space/p/ainews-google-io-2026-gemini-35-flash] * Andrej Karpathy joins Anthropic — pre-training, on Nick Joseph's team [https://x.com/karpathy/status/2056753169888334312] * Ethan Mollick: recursive self-improvement is a talent sink for the Big Three [https://x.com/emollick/status/2057074407177130096] * Qwen 3.7-Max and the Zhenwu M890 chip — Alibaba's full-stack I/O response [https://meyka.com/blog/alibaba-upgrades-ai-stack-with-qwen-3-7-max-executes-tasks-for-35-hours-supports-1000-tools/] * DeepSeek hiring a Code Harness team in Beijing [https://x.com/victor207755822/status/2057064415300841626] * Railway's 8-hour outage after GCP's automated account suspension [https://blog.railway.com/p/incident-report-may-19-2026-gcp-account-outage] * Cheng Huang: 130K lines of Rust, AI-written contracts, and a Paxos engine that runs [https://zfhuang99.github.io/rust/claude%20code/codex/contracts/spec-driven%20development/2025/12/01/rust-with-ai.html] * TechCrunch on Anthropic's pre-training charter for Karpathy [https://techcrunch.com/2026/05/19/openai-co-founder-andrej-karpathy-joins-anthropics-pre-training-team/]

Mostly-work, malicious npm, and one engineer replacing a law firm

A six-month overview from Simon Willison anchors the day: coding agents crossed from often-work to mostly-work in November, and laptop-class models started outrunning expectations. Then a fresh npm supply-chain attack — 637 malicious versions in 22 minutes — that for the first time specifically hijacks Claude Code and Codex agent hooks for persistence. Plus a Number 10 talk on replacing a one-and-a-half-million-pound law-firm contract with one embedded engineer, an editor-layer company renting xAI's Colossus 2, Ethan Mollick on insourcing, the full GenMedia pipeline running for a dollar a book, Daniel Griesser's pi-config skill repo, and two obituaries that hit the Unix world in the same week. * Simon Willison's last-six-months-in-LLMs PyCon lightning talk [https://simonwillison.net/2026/May/19/5-minute-llms/] * Mini Shai-Hulud strikes again — 317 npm packages and your agent hooks [https://safedep.io/mini-shai-hulud-strikes-again-314-npm-packages-compromised/] * Prime Intellect's General-Agent — synthetic RL environments [https://x.com/PrimeIntellect/status/2056569877167808966] * Eoin Mulgrew on Number 10's insurgent technical unit [https://www.youtube.com/watch?v=ObNKGf9YR0g] * Cursor's Compose 2.5 reportedly trained on xAI's Colossus 2 [https://x.com/techdevnotes/status/2056543940052910237] * Ethan Mollick on insourcing via hiring [https://x.com/emollick/status/2056578946813100173] * Guillaume Vernade's full GenMedia pipeline at a dollar a book [https://www.youtube.com/watch?v=BcWFc3H7Khg] * Daniel Griesser's pi-config — Plan, handoff, and subagent skills [https://x.com/DanielGri/status/2056676488183689620] * Peter Neumann (1932–2026) [https://www.tuhs.org/pipermail/tuhs/2026-May/033748.html] * Peter Salus (1938–2026) [https://www.tuhs.org/pipermail/tuhs/2026-May/033750.html] * Magnifica humanitas confirmed for May 25 [https://www.vaticannews.va/en/pope/news/2026-05/pope-leo-xiv-first-encyclical-magnifica-humanitas.html]

Cold starts, radio stations, and a circuit you can subtract

Monday's lineup: Modal publishes the full architecture behind a 40x reduction in serverless-GPU cold-start latency, Andon Labs releases the five-month results from letting four frontier models run real radio stations, and a researcher locates and turns off the political-censorship circuit inside Qwen 3.5 9B. Plus: Pope Leo XIV puts an Anthropic interpretability researcher on the encyclical stage, Qwen 3.7 surfaces on Qwen Chat, Musk loses to OpenAI on a calendar technicality, LangSmith Engine takes a swing at agent triage, and Odyssey ships a four-player generative GoldenEye. * Modal's 50-second cold start [https://modal.com/blog/truly-serverless-gpus] * Five months of AI radio [https://andonlabs.com/blog/andon-fm] * Magnifica humanitas at the Vatican [https://www.vaticannews.va/en/pope/news/2026-05/pope-leo-xiv-first-encyclical-magnifica-humanitas.html] * Reading Qwen 3.5's censorship out of its weights [https://vas-blog.pages.dev/qwen-censorship/] * Qwen 3.7 surfaces [https://www.reddit.com/r/LocalLLaMA/comments/1tgpabe/qwen_37_droped_on_qwen_chat/] and Musk loses [https://www.cnbc.com/2026/05/18/musk-altman-openai-trial-verdict.html] * LangSmith Engine takes a swing at agent triage [https://www.langchain.com/blog/introducing-langsmith-engine] * Agora-1 generates a shared GoldenEye [https://odyssey.ml/introducing-agora-1] * Three questions for I/O tomorrow [https://x.com/sundarpichai/status/2056524502746747048]