Daily AI news. I build with it too.
Chennai
Joined September 2010
- Tweets2.7K
- Following1.7K
- Followers383
- Likes2.3K
Same coding work as DeepSeek V4.1 Flash, about 45% fewer output tokens.
Particle-1, Suhail's first model:
- 28-52% fewer tokens on 4 coding benchmarks
- 34-52% cheaper per task
- pass rate within the base model's margin
Company numbers, not independent yet.
Would you switch?
Agents quietly change things you didn't ask for. Make them confess first.
Paste after any task:
"List every file you changed, one line each: what and why. Flag anything I didn't ask for and revert it unless I approve."
What's the weirdest extra change you've caught?
Vercel's CI got 2-6x faster by skipping checks a change doesn't touch.
Vercel CI, now in beta:
- reads turbo.json, no workflow YAML
- unchanged tasks come from cache
- agents read logs via CLI or MCP
My take: agents made CI the queue.
What's your push-to-green time?
Utah gave 34 devs in 10 state agencies Claude Code for 4 weeks.
- 80% set up in 30 minutes
- 71% said it saved them 20+ hours
Utah's caveat: it's easy to build buggy solutions, so they're adding standards and skills.
My take: standards are the real work.
What's in yours?
Most agent bugs start as a silent guess. Make it say the guess out loud.
Paste before any task:
"Before coding, list every assumption about inputs, current behavior and requirements. Mark each verified or guessed. Ask me about the guessed ones."
What does yours guess wrong?
Claude Code ignores .claudeignore. Block secrets in .claude/settings.json:
"permissions": {
"deny": [
"Read(./.env)",
"Read(./.env.*)",
"Read(./secrets/**)"
]
}
- deny beats ask and allow
- also blocks Edit/Write
- scripts can still read them: use the sandbox
Matt Pocock is weighing a ban on useState to push agents into "the black hole of success."
The case: agents tangle useState and useEffect. useReducer makes every transition explicit.
My take: agents need fewer ways to be wrong.
Would you ban it?
teapawt.org/mattpocockuk/status/21…
GitHub now picks your model, and its top tier lands within 2 points of Claude Code on Opus 5.5 Max.
Copilot Fusion modes:
- Single: one model
- Cascade: cheap draft, judge escalates
- Critique: one drafts, others review
About half the cost per task.
Would you let it pick?
My take: yesterday was about the layer around the model. Routers pick it, reports price its work, specs let it touch lab hardware. The model is turning into a part, not the product.
Which layer are you building on?
10/ Artificial Analysis gave one robot the same task with two models. Claude Opus 5.5 dabbed out "hello world" dot by dot: 3,270 arm moves in 34 minutes. GPT-6 Astra wrote in continuous strokes: 514 moves in 13 minutes.
teapawt.org/ArtificialAnlys/status…
Same robot, same task, two models with very different writing styles
➤ On the left, Claude Opus 5.5 dabs out "hello world" dot by dot: 3,270 arm moves over 34 minutes
➤ On the right, GPT-6 Astra writes in continuous strokes: 514 moves in 13 minutes
From AA-Robotics v0.1, seen from the model’s view
See more examples here: artificialanalysis.ai/articl…
9/ Simon Willison says October is another inflection point for coding agents: Claude Opus 5.5 can now build video games that are actually fun.
teapawt.org/simonw/status/21154447…
I think October represents another inflection point in capabilities for coding agents and the models that drive them: Claude Opus 5.5 can now build video games that are actually fun simonwillison.net/2026/Oct/2…
8/ Anthropic expanded AI for Science to 20 foundations, institutions and research groups, with free Claude credits for high-impact projects.
Built with HHMI Janelia: the Model Hardware Standard, a shared spec that lets AI agents safely operate lab instruments. Research preview.
5/ Anthropic gave Claude Enterprise admins smart reports. They sample team transcripts, group the work into workstreams, and show spend, deliverables and where things broke down.
Plus in-app surveys and pooled team budgets. ROI is now a product feature.
4/ OpenAI now lets you text your Dot over iMessage or RCS, rolling out in preview for Pro users.
Same day, ServiceNow said it plans to give all its employees ChatGPT Work and Codex, and its MCP plugin went live in OpenAI's marketplace.
3/ VS Code 1.144 shipped experimental computer use: Copilot agents can drive desktop apps on Windows and macOS once you authorize it with /computer on.
It also generates and edits images in Copilot Chat, and Fusion is 30% off on paid plans through Nov 30.
2/ GitHub Universe: a new Copilot app, multi-repo projects, and Copilot Fusion GA, which picks models per task.
GitHub says Fusion Max lands within 2 points of Opus 5.5 Max at half the cost and 40% less time per task.
Enterprise now gets Copilot with no separate subscription.