Ctrl Alt Debrief

Read the 2026-04-17 edition of the @CtrlAltDebriefX newsletter here: ctrlaltdebrief.com/archive/e10cdf…

Plus: Jane Street commits $7B to CoreWeave, Ukraine captures a Russian trench using only drones and robots, Google ships three Gemini releases, and a transformer runs in HyperCard on a 1989 Mac.

OpenAI's Codex now clicks and types on your Mac alongside you, with an in-app browser and persistent memory. The White House queued Anthropic's Mythos for six federal agencies. Physical Intelligence's π0.7 robots started nailing tasks they were never trained on. All in one day.

Read the 2026-04-02 edition of @CtrlAltDebriefX here: ctrlaltdebrief.com/archive/6f1260…

Plus: Anthropic quietly walked back its safety commitments, Datadog rebuilt its replication architecture from scratch, and a 15KB library that skips the DOM entirely.

Claude wrote a full FreeBSD kernel exploit in 4 hours, root shell and all. Q1 startup funding hit $297B (2.5x last quarter). Slack turned Slackbot into an enterprise AI agent with 30 features.

There's a Bloomberg piece from Friday about what's been happening in competitive chess over the last couple of years. So chess had this problem where engine prep got so good that top level tournament games were just ending in draws constantly. Both players had Stockfish, both memorised whatever it said was the best line, and neither could break through. The 2018 World Championship had all 12 games end in draws which was a first in 130+ years. Then this new generation of grandmasters who literally grew up training with engines started doing something the older players thought was stupid. At the 2024 Candidates Tournament, an 18 year old Indian GM called Praggnanandhaa played a move that a former world championship challenger said he hadn't seen in 25 years. Stockfish rates it as weaker than the alternatives. His opponent had absolutely zero preparation for it and lost the game. A German GM has a name for what the old guard was doing, he calls it "hitting spacebar", just letting the engine tell you the best move over and over. You feel like you're good at chess but the moment someone goes off-book you're completely screwed because you never understood why you were playing those moves in the first place. I run Claude Code across basically all my projects now and this is exactly what happens when you just accept whatever it suggests. It'll propose something that looks clean, passes tests, and is completely wrong for how the rest of your codebase actually works. Like it'll refactor a function to be "better" but break an assumption that four other files depend on because it doesn't have the context you have. A chess author put it well, if there are four moves that are basically equal, you have to be good enough to see that the third of the four is going to be the most annoying choice for your specific situation. Same thing with coding agents. The top suggestion and the right suggestion for your codebase are often different things and knowing when to override the model is kind of the whole job now. Rest of today's edition has Anthropic winning a court injunction against the Pentagon (judge called it "First Amendment retaliation"), H100 GPU rentals crashing to $2/hr, Apple getting the rights to distil Google's Gemini into on-device models, GitLab's co-founder building 10+ personalised cancer drugs and founding 10 companies after standard treatment ran out, and my favourite, someone running a multi-agent system on a $7/month VPS using IRC.

15 stories in the 2026-03-30 edition. Read here → ctrlaltdebrief.com/archive/82979e…

Plus: GitLab's founder built 10+ personalised cancer drugs after standard treatment ran out, RotorQuant cuts KV cache params 44x, and grandmasters beat chess AI with deliberately bad moves.

Federal judge blocks Pentagon's Anthropic blacklist, calling it "First Amendment retaliation." H100 GPU rentals crashed from $8/hr to under $2. Eli Lilly signs $2.75B deal for AI-discovered drugs. Apple gets rights to distil Google's Gemini for on-device AI.

Read the 2026-03-26 edition here: ctrlaltdebrief.com/archive/12062f…

Plus: Fox runs local LLMs 2x faster than Ollama. Open source funding is broken. Figma rebuilt its cache for six nines. And 9 more.

Anthropic built a microscope for Claude's brain. It doesn't reason the way it claims. Apple plans a standalone Siri app for iOS 27. ARC-AGI-3 drops 1,000+ game worlds to test if AI can actually learn.

15 stories in today's edition: ctrlaltdebrief.com/archive/8c0b26…

Arm shipped its first in-house chip. Claude Code got auto mode. Ed Zitron says AI data centre capacity is inflated. And a piece arguing the AI wealth window has 5-10 years left.

OpenAI raised another $10B (round now past $120B) and killed Sora on the same day. Disney pulled its $1B deal

a16z says the comfortable middle is over for software. 96% of enterprise permissions go unused, which is fine until an agent inherits them. Plus DSPy's 4.7M downloads vs LangChain's 222M. Read the 2026-03-24 edition of @CtrlAltDebriefX here: ctrlaltdebrief.com/archive/d33143…

OpenAI's chief scientist says they're building an AI research intern by September and a full automated system by 2028 (MIT Tech Review exclusive). Claude shipped computer use on Mac. OpenAI offering PE firms 17.5% guaranteed returns.

Read the 20 March 2026 edition here: ctrlaltdebrief.com/archive/78dec0…

Also today: Karpathy's autoresearch hits 9x speedup with a GPU cluster, CNBC on why Jensen Huang needs a moat not a chip, and why perfect AI retrieval still gives wrong answers.

OpenAI just acquired Astral (uv, ruff) - 126M downloads/month of Python tooling now inside Codex. Bezos is raising $100B to buy and automate manufacturing with AI. Uber ordered 50,000 Rivian robotaxis for $1.25B.