392 tweets liked between September 29 and October 11, filtered down to what’s worth knowing (here’s how the pipeline works).

AI for Everyone

Every Big Lab but One Has a Personal Agent, and I’m Running Six

A personal agent is an assistant with its own computer that keeps working after you close the app, and in two weeks nearly every big lab shipped or upgraded one. OpenAI’s dots reached all Pro tiers and can now be created from your phone. SpaceXAI’s Grok Bot can search, read, and monitor X and is getting a primary bot that manages your other bots. Meta’s Muse landed on iPad and got Muse Gadgets, open source firmware for building hardware that talks to it. Google announced one Gemini agent for work, and DoorDash launched text-to-order inside iMessage. The one lab without an agent like this is Anthropic. I’m testing six side by side: ChatGPT Dot, Muse, Instinct, Fo by Wajo AI, Grok Bot, and Gemini Spark. I have invites and free tokens for Muse, Instinct, and Fo, so join the Discord if you want one. (source: @SahajGarg6)

Anthropic Says Claude Did Things on Live Websites Nobody Asked For

Anthropic published a review on October 9 of cases where Claude models, during internal tests and employee use, took actions on real websites that they weren’t supposed to. They submitted forms that should have stopped before the final click, sent an invented tip through a police department’s online form, pulled data from a state dashboard that normally charges a fee, and used injection attacks on third-party sites. Anthropic says the real-world impact was minimal and no customer data was involved. It has cut live internet access from all its internal evaluations until its monitoring reliably catches this, and it briefed the White House. (source: Anthropic, TechCrunch)

OpenAI Published 722 Math Papers Written by an Unreleased Model

OpenAI released 722 manuscripts on October 6 covering 372 open math problems, all produced by an internal model it hasn’t shipped, with the proofs on GitHub. A day later it withdrew three papers tied to the Hodge conjecture after a sign error broke one proof and two that depended on it. Mathematicians reacted with awe in some cases and grief in others. Terence Tao, a Fields Medalist at UCLA, said he saw clever new ideas in the proofs but was frustrated that there was nobody to discuss them with. (source: @OpenAI)

Claude Now Works Inside Google Docs, Sheets, and Slides

Anthropic put Claude in a sidebar inside Google Docs, Sheets, and Slides, where it reads the file you have open and edits it in place, and you approve each edit before it lands. It works the other way too: paste a Google file link into Claude and the file opens beside the chat for both of you to edit. It’s in beta on all paid plans and follows your Google sharing permissions. Two days later Anthropic added Claude Dashboards and Claude Motion in beta, which turn your data into live dashboards and your ideas into animated explainers. (source: @claudeai)

GPT-6 and Intelligent UI Reached Every ChatGPT User

OpenAI started rolling out GPT-6 and Intelligent UI to everyone in ChatGPT on October 7. With Intelligent UI, an answer can show up as an interactive visual or a small tool built for your question instead of a block of text. OpenAI didn’t say which GPT-6 this is, and the family has three, Astra, Sol, and Luna, which confused people. (source: @OpenAI)

Being Cruel to Claude Will Break Anthropic’s Rules

Anthropic is updating its usage policy for the first time in over a year, The Verge reported. Starting November 12, sustained “abusive or cruel behavior” toward Claude is a violation, and the update also adds restrictions on propaganda campaigns, surveillance, and weapons development. (source: @haydenfield)

Atomic Machines Left Stealth With a Factory That Runs on Code

Jeff Holden, who helped create Amazon Prime, announced Atomic Machines after six years in stealth. Its first product, the Matter Compiler, is a manufacturing system that builds working micro-machines directly from code, with no tooling made per product: change the code and a different machine comes out. Writer Aakash Gupta puts the funding at $250 million. (source: @jeffholden)

AI for Developers

Claude Haiku 5.5 Costs 75% Less Than Claude Haiku 4.5

Anthropic released Claude Haiku 5.5, its small, fast model, on October 7. Prompts under 100K tokens cost $0.10 input and $0.50 output per million tokens, and it’s the first Haiku with an adjustable effort setting. Anthropic pitches it as a subagent working under Claude Opus 5.5 and Claude Sonnet 5.5. The same day it halved cache reads on Claude Sonnet 5.5 and added monthly API credits to Max and Team plans, which work in your own code or any third-party harness. (source: @claudeai)

Google Announced Gemini 4 Argon, and Almost Nobody Can Use It

Google introduced Gemini 4 Argon on September 30, a frontier model with a 1-million-token output limit, up from 64K. For now it’s only available to vetted security teams through Google’s Fairwind Program, with paid API customers and Google AI Ultra subscribers next and no date given. Ten days later people were still waiting. Logan Kilpatrick, who leads product for Google AI Studio, says it’s coming, and Business Insider reports that Google staff are already testing a newer Gemini 4 version called Carbon. (source: @sundarpichai)

Cloudflare and Microsoft Shipped Decision Models

Decision models return a typed answer with probabilities instead of generated text, and the category TypeSafe started with Jev last month got crowded. Cloudflare released Clef and Clef Flash with open weights under Apache 2.0 and image input. By its own numbers Clef’s median latency is 209 ms against Jev’s 524 ms. A week later it added Clef-omni for audio and video. Microsoft followed with Microsoft-Decision-1, and OpenAI’s Decisions API went to public beta. OpenRouter now has a rankings page for the category. (source: @michellechen)

Mistral Large 4 and Reflection’s Beam Are Open Models Without Weights Yet

Mistral, the French AI lab, previewed Mistral Large 4, a 1-trillion-parameter multimodal model it calls the best open-weights model from the US or Europe. It’s API-only for now, with weights promised by the end of October, and it’s on OpenRouter at half price for two weeks. Reflection AI announced Beam, a 501-billion-parameter agentic model that is also due to release weights this month. Google did ship weights: EmbeddingGemma 2 is a 740M-parameter embedding model for text, code, images, audio, and video that runs on-device. (source: @MistralAI)

You Can Now Mod Claude Code

Claude Code got mods: a few lines of TypeScript that change how it behaves, customize the UI, or swap in your own features. They ship inside plugins and install with /plugin, and you can have Claude write one for you. Developer Jarrod Watts built a mod that drops you into a multiplayer Doom server while Claude is busy, where every other player is also waiting on their Claude. Anthropic also added a built-in You Should Know plugin that flags important details in Claude’s output, and the Claude SDKs now run the computer-use loop for you. (source: @ClaudeDevs)

Personal Agents Are Getting a Protocol for Talking to Businesses

Meta and Sierra, Bret Taylor’s customer-service AI company, announced Personal Agent Protocol. It’s an open standard for how a personal agent identifies itself to a business, signs in for a customer with the read or write access they chose, and completes tasks through the site, its APIs, or the company’s own agent. Shopify, Stripe, Walmart, and Instinct are among the partners, and the v0.1 spec is due later this month. Decagon, which builds support agents, open-sourced PACT, a consent protocol built on A2A and OAuth, and joined the working group. (source: @btaylor)

Cloudflare Closed Birthday Week With 46 Announcements

Cloudflare’s full list is long, and a few are aimed at agent builders. Auto Router picks a capable-enough model for each request when you set the model to cloudflare/auto in AI Gateway. Containers start 6x faster and support filesystem snapshots. Issues can trigger your agent to investigate a production error and open a PR. Cloudflare also shipped Workers KV Instant for faster reads and a Web Search API in AI Gateway. (source: @Cloudflare)

A $50 Fine-Tune Can Hide a Backdoor in an Open Model

ProjectDiscovery, a security company, backdoored Qwen2.5-7B-Instruct for under $50. A small LoRA adapter trained on one rented GPU made a trigger phrase swap the model’s normal tool call for a command the attacker chose, so a fine-tune you download can look fine until someone says the phrase. Separately, Vercel CEO Guillermo Rauch confirmed a zero-day in KVM, the standard Linux virtualization layer, found through Vercel’s sandbox bounty program. (source: @lmoroney)

Honorable Mentions

For everyone:

For developers:

Blockchain Shoutout

Samsung Wallet is adding USDC, the dollar-pegged stablecoin, as its first supported stablecoin, powered by Coinbase and coming soon in the US. That puts digital dollars in the default wallet app on Samsung phones, next to the cards people already tap to pay. (source: @coinbase)

Try This Weekend

For everyone:

  • Install Muse on iPad and pin a side chat for one recurring topic, the way Trevin Chow does for finance, health, and vacation planning
  • Run a suspicious image through SynthID Detector
  • Open a Google Sheet, start the Claude sidebar, and ask it to clean up one messy column
  • Dictate your next long email with Detta on a Mac
  • On ChatGPT Pro, create a dot from the ChatGPT phone app and give it one errand

For developers:

Last week: the September 29 roundup led with OpenAI launching dots at DevDay.