Tim Williams

The team out at GTC this week - go see James speak later!

Great discussion on spec driven agentic coding at the london agentic meetup. Thanks for organising @Shashikant86 !

Getting a lot of ollama posts now, dressed up as posts claiming to offer free Claude code. Seems like an awkward growth hack that ends only in disappointment

@tomas_hk Surprising graphs: how gepa-maxed the existing models already seem to be - was this under some constraint to gepa? And the huge gains possible to gpt-oss! Congrats on the release!

Today we are releasing a new Prompt Optimization algorithm with significant accuracy and efficiency gains, generally available now through our API.

What’s your agentic Dunbar number?

@ChadMoran @bcherny Likely a Ghostty thing: ghostty.org/docs/install/r…

@bcherny Will do tomorrow.

Claude Code 2.1.0 is officially out! claude update to get it We shipped: - Shift+enter for newlines, w/ zero setup - Add hooks directly to agents & skills frontmatter - Skills: forked context, hot reload, custom agent support, invoke with / - Agents no longer stop when you deny a tool use - Configure the model to respond in your language (eg. Japanese, Spanish) - Wildcard support for tool permissions: eg. Bash(*-h*) - /teleport your session to claude.ai/code - Overall: 1096 commits github.com/anthropics/cla… If you haven't tried Claude Code yet: code.claude.com/docs/en/setup Lmk what you think!

@mmoderwell creativity is a big factor, humans are at least currently SOTA in that regard

Compute is not the bottleneck for scientific breakthroughs. Human creativity is.

As an english cricket follower and developer, I think the term vibe-coding needs updating to Baz-coding. Great in familiar, or simple unfamiliar conditions but absolutely useless in complicated unfamiliar conditions

It bothers me that we call LLM function callers 'Agents'. The term 'Actor' seems so much closer semantically. A system prompt is the outline given as we begin improvisation. If we don't like how the scene plays out, we yell 'cut' and start again from the top with notes.

@brankopetric00 Give @useblacksmith a go - turbo boosts for your actions.

Developers complained: staging deployments take 35 minutes. Process: 1. Push code 2. Jenkins builds Docker image (12 min) 3. Push to ECR (4 min) 4. Update ECS task definition (2 min) 5. Deploy to staging (12 min) 6. Run smoke tests (5 min) Optimized: 1. Implemented layer caching in Docker: 12 min -> 3 min 2. Parallel ECR pushes: 4 min -> 3 min 3. Blue-green ECS deployment: 12 min -> 4 min 4. Parallel smoke tests: 5 min -> 2 min New total: 12 minutes Impact: - 15 deploys/day × 33 minutes saved = 8.25 hours/day saved - Team ships 2x more frequently - Feedback loop faster Time to implement: 3 days Slow CI/CD kills velocity. Every minute you shave off multiplies across the team.

I’m suddenly getting a ton of nerdy posts from European weather watchers seemingly desperate for cold conditions. I’m all for a nerdy post but I don’t understand why they want it to get so chilly

Wrapping python functions as tools can make the discovery phase of functionality quicker but robs the llm of its greatest ability to deal with and circumvent errors.

@ashpreetbedi Why not just to markdown file? Start simple

Whats the opinion on Agents generating user memories? Not a fan of the lag, but async/background can improve that. Curious how others feel about this, would you ship this in prod? (of course with a diff db not sqlite 😂)

@dani_avila7 Like the update phase in mem0 arxiv.org/pdf/2504.19413

Getting tired of Claude Code automatically running /compact and stopping my work every time the context gets full So I'm working on a background hook that automatically saves key decisions, code patterns, and session context when hitting 80% usage. Then it compacts everything seamlessly without interrupting my work. Here's the flow I'm building. Publishing it on aitmpl.com in a few days for anyone who wants it.

Scaffolding supports a building only until it can stand on its own. Agents and tools play the same role for LLMs. Until the bitter lesson renders them unnecessary, they remain a useful abstraction.

GPT5 in cursor. Eventually good, but it really takes the scenic route, stopping to read everything along the way

We’ve been listed on The Future 50 ! 🚀 Nominated by investors and selected by @thegeneralistco, we’re recognized as one of the world’s most exciting startups. Learn more about us here: orbitalmaterials.com/products

Us: ugh 4o is too sycophantic OpenAI rushing to update Kanban board, slipping on spilled LaCroix: ‘Absolutely! You’re right! We’ll fix that right away’

being cheap, I love this method. It has saved me many tokens over the years.

@Stammy I've been kicking this around, automatically collecting a context of interactions locally through your browser, to allow local personalisation, and eventually provide context to agents to work intelligently on my behalf.

The magic of browsing the web isn't quite gone, but it's waiting to be reinvented. There's a lot of detail I'm leaving out of this thread, read the full article here: paulstamatiou.com/browse-no-more

📝 just published a new article I love and use AI answer engines (Perplexity, ChatGPT, Grok, etc) all day. There's a problem though. I'm no longer browsing the web; I'm consuming AI answers instead. I'm missing out on what made the web magical to begin with.