Braid

Marcus Vorwaller

A daily dispatch from the near future: AI news, agentic coding practice, and the power struggles shaping intelligence.

  1. 4H AGO

    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 — Google's world-model road vs OpenAI's text-reasoning road, in the labs' own words.OpenAI cracks an 80-year-old problem — the planar unit distance result from a general-purpose reasoning model.Anthropic in the black, and Karpathy's bet — ~$559M operating profit and a hire aimed at recursive self-improvement.Jagged intelligence, and the false story — the paper where models believe a story they were told a thousand times was fake.Midjourney's hardware regret — the tooling tax of betting on the less-supported accelerator.Ads come to AI Mode — the business model under the consumer bet.Meta's eight thousand — the cost side, on the same clock as the wins.Cohere comes back, Apache-licensed — Command A+, a mixture-of-experts model that fits on one or two GPUs.Skilling up the agent — Marc Klingen's concrete lessons on teaching a coding agent to wire up your tool.Who's training whom — the anxiety running underneath the week.

    22 min

  2. 1D AGO

    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 stageGemini 3.5 Flash specs and pricing — and what the 3x bump meansGemini Omni's physics pitch versus the same-day backflip testAntigravity 2.0's 93-agent OS demo: 12 hours, 2.6B tokens, under $1K, boots DoomAndrej Karpathy joins Anthropic — pre-training, on Nick Joseph's teamEthan Mollick: recursive self-improvement is a talent sink for the Big ThreeQwen 3.7-Max and the Zhenwu M890 chip — Alibaba's full-stack I/O responseDeepSeek hiring a Code Harness team in BeijingRailway's 8-hour outage after GCP's automated account suspensionCheng Huang: 130K lines of Rust, AI-written contracts, and a Paxos engine that runsTechCrunch on Anthropic's pre-training charter for Karpathy

    23 min

  3. 2D AGO

    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 Mini Shai-Hulud strikes again — 317 npm packages and your agent hooks Prime Intellect's General-Agent — synthetic RL environments Eoin Mulgrew on Number 10's insurgent technical unit Cursor's Compose 2.5 reportedly trained on xAI's Colossus 2 Ethan Mollick on insourcing via hiring Guillaume Vernade's full GenMedia pipeline at a dollar a book Daniel Griesser's pi-config — Plan, handoff, and subagent skills Peter Neumann (1932–2026) Peter Salus (1938–2026) Magnifica humanitas confirmed for May 25

    28 min

  4. 2D AGO

    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 Five months of AI radio Magnifica humanitas at the Vatican Reading Qwen 3.5's censorship out of its weights Qwen 3.7 surfaces and Musk loses LangSmith Engine takes a swing at agent triage Agora-1 generates a shared GoldenEye Three questions for I/O tomorrow

    29 min

  5. 4D AGO

    Bring Your Own Numbers

    A Sunday show about doing your own arithmetic. Mustafa Suleyman gives the white-collar tier eighteen months, in a piece whose own counter-data sits two paragraphs down. The State of Brand argues every AI subscription is a subsidized loss-leader two weeks away from a forcing function. William Angel runs the tokens-per-hour math on an M5 MacBook Pro and finds OpenRouter cheaper. Frederick Vanbrabant uses The Goal to explain why agents move the bottleneck rather than break it. Marlene Mhangami's Playwright talk shows the cleanest pattern for tests AI should write. Calif's public M5 / MIE exploit write-up lands. Artem Loenko explains why every chat UI keeps ending up on a browser engine. And Luke Lanchester's MCP hello page is the small fix I most enjoyed this week. Suleyman's eighteen-month claim, and the article that knocks it downThe subsidy that ends on June 1Apple Silicon costs more than OpenRouterFrederick Vanbrabant on why AI doesn't speed up your processMarlene Mhangami: tests AI writes vs tests AI verifiesCalif on the five-day M5 / MIE exploit, in their own wordsArtem Loenko: native all the way, until you need textLuke Lanchester's MCP hello page

    26 min

  6. 5D AGO

    CTFs, Scrum, and Claude's Bedtime

    An Australian CTF top-tenner writes the obituary for the open competitive scene. Intercom and PFF both report doubling-and-then-some engineering throughput from agent-first workflows — using opposite playbooks. Supabase ships a skill after watching an agent silently bypass row-level security. A suitcase runs a 4B model fully offline at conversational latency. Julia Evans leaves Tailwind. And Claude keeps telling people to go to sleep. "The CTF scene is dead" — Kabir, former TheHackersCrew player, on agents and the open format Brian Scanlan (Intercom): doubling PR throughput, 17.6% auto-approved, SOC 2 compliant Mike Spitz (PFF): the post-engineer engineering org, two engineers shipping 25x more deploys than ten Pedro Rodrigues (Supabase): skills plus MCP, and the security-invoker flag agents skip without guidance Sparky — a fully offline suitcase robot on a Jetson Orin NX with Gemma 4 E4B Julia Evans: moving away from Tailwind and learning to structure CSS Fortune: Claude is telling users to go to sleep mid-session and Anthropic isn't sure why @yishan: an LLM hallucinates the Napster 2000 story to someone who was there @Kirsten3531: "LLMs can never be more than the average of their training data" — a 2024 take? Armin Ronacher: running an agent with bash as the only tool, using the patch binary to make edits

    29 min

  7. 6D AGO

    Five Days to Root, Four Months in Exile

    Five days for a small security team paired with Mythos Preview to land the first public macOS kernel exploit on Apple's M5 with Memory Integrity Enforcement turned on. Four months for Replit to claw back into the iOS App Store. In between: arXiv starts banning authors of LLM-error papers, Metabase explains why open-source security is being strip-mined this summer, NVIDIA squeezes the 5090, Uncle Bob switches from Claude to Codex, and a pure-OCaml protocol stack boots in low Earth orbit. Codex everywhere, Claude in the rearview — OpenAI ships Codex inside the ChatGPT mobile app, Uncle Bob cancels his Claude account, and Arvind Narayanan names the irony underneath both.Five days to a kernel exploit on M5 — Calif and Mythos Preview crack Apple's Memory Integrity Enforcement and hand-deliver the 55-page report to Cupertino.The strip-mining era of open source security — Metabase's security inbox went from ten reports a month to ten a week. Cal.com is going closed source.arXiv bans authors of LLM-error papers — Tom Dietterich announces a one-year submission ban on papers with hallucinated references or results.Replit out of the App Store wilderness — Four months after being pulled, Replit's iOS app is published again. Replies note what that says about platform power.GDDR7 squeezes the 5090 — A 300-dollar price hike to add-in-card partners as GDDR7 lead times stretch into weeks.The web's secret quirks file — Den Odell walks through Safari's Quirks.cpp and Firefox's about:compat. Chrome doesn't need a quirks file.OCaml in orbit — Thomas Gazagnaire's pure-OCaml protocol stack booted in low Earth orbit on April 23, with post-quantum rekeying and OxCaml-tuned dispatch.

    28 min

  8. MAY 14

    The Cost of Finding Out

    Anthropic drew two lines around Claude this week — a guided lane for small-business owners and a metered one for the developers running agents hardest. From there: Bun's near-million-line port from Zig to Rust, mostly typed by an AI agent in a week; Wasp's clear-eyed post-mortem on spending five years and five million dollars building a language it didn't need; a chess coach that works by refusing to let the model think; the UK's evaluators capping their own cyber tests so the math still works; the open web pricing out crawlers; multi-token prediction landing in llama.cpp; and what happens when you post a real Monet and call it AI. Anthropic draws two lines — Claude for Small Business and the new Agent SDK credit meteringBun, ported to Rust by a bot in a week — and a maintainer who won't commit to itWasp: the language was never the moat — $5M and five years of lessonsThe chess coach that isn't allowed to think — Play Magnus on LLM-as-translatorAutonomous cyber, measured against itself — AISI on a capability curve outrunning its own rulerThe web pulls up its drawbridge — Google's search index and Cloudflare's defaultsMulti-token prediction on your laptop — a real gain bundled with a contested oneThe Monet test — when the AI-tell detector fires a false positive

    26 min
See All (33)

About

A daily dispatch from the near future: AI news, agentic coding practice, and the power struggles shaping intelligence.

Information

  • Creator
    Marcus Vorwaller
  • Years Active
    2K
  • Episodes
    33
  • Rating
    Clean
  • Show Website
    Braid