Stories, page 5
- Theo's judge panel ranks Astra first, Grok 4.7 far back
Then he ran the same ranking with Astra as judge, and Fable finished last.
- Anthropic says it made claude.ai 3x faster in two weeks
The write-up includes the prompts and methods, which is the part worth testing on your own codebase.
- Theo says no lab has both small and large models worth using
A three line post ranks three labs by what they are missing, and shows no work.
- Boris Cherny had two models port HAProxy from C to Rust
Both models passed nearly all of HAProxy's tests, but one got there faster and cheaper.
- Kenton Varda says testers mistook Opus 5.5 for a new Fable
A reviewer who says he still reads the code rates Opus 5.5 close to Fable on style, and keeps GPT models for bug hunting.
- Lauren says she shipped 2,500 PRs to production last month
The talk was meant for Cursor Compile in London, and she put it on X instead.
- Nat Friedman says Meta built Muse from scratch
Peter Steinberger, whose harness started the argument, backed Friedman's account and said kudos.
- Theo says Grok 4.7 is slower, worse and over 2x the cost of 4.6
The promised token efficiency gain went the other way, and Theo says he cannot find a single benchmark where it holds.
- Claude Code Projects turns a codebase into one conversation
One agent per project handles memory and spins off subagents, and the work keeps running after you close the laptop.
- ThePrimeagen ran four model passes and hated the result
Three review passes caught guardrails and dead code, and none of them caught the design.
- Open models hit 78.4% of tokens on Vercel AI Gateway
Spend still skews closed, but the cheap models are now doing most of the work.
- Two teams benchmark Jev, and it wins on speed and accuracy
Independent evals from Malte Ubl and the fx team both put Jev ahead of the small models they were already running.
- Theo lists six problems with Jev compaction, Armin Ronacher pushes back
A demo that scores every tool call and deletes the rest turned into the week's argument about what a fast classifier can actually decide.
- dax says matching DeepSeek inference is a $100M problem
He also says any provider advertising a 99% cache rate is wrapping DeepSeek, because DeepSeek is the only one hitting that number.
- Theo says ranking reasoning outputs is Jev's worst use
Malte Ubl counters with the Lean proof argument, that checking work is easier than doing it.
- Claude Code now falls back to AGENTS.md
Thariq says the fallback is live and toggleable, and the replies moved straight on to skills.
- Mario Zechner warns frontier models break code behind you
Two posts from @badlogicgames make the same argument from different directions, and neither is about model benchmarks.
- Mario Zechner calls TypeSafe's Jev a general purpose classifier
Same-day impressions, one price tag, and no evals yet.
- roon says Astra's code makes corrigibility a matter of faith
Theo quote-posted it as OpenAI employees inventing words for unreadable code, then agreed the fix belongs in the model.
- dax moves part of his team off Astra after spend doubles
One team's cost numbers, a few bad impressions and a paused top tier all landed in the same week.
- Engineers push back on Amodei's pacing the frontier call
The loudest objections came from people shipping production systems, not from other labs.
- Mitchell Hashimoto's whiteboard defense for AI code
He does not care who typed the code, only whether you can defend the system in a hallway conversation.
- Thariq flips, says MCP now beats CLIs for most integrations
The "just give it a CLI" consensus is getting walked back by several of the people who held it.