Three threads run through today’s issue.
Google shipped Gemini 4 Carbon to every paying developer at Argon’s price, and launched three agent efforts around it: swarms of coding agents, DeepMind’s proofs about Signal’s code and a healthcare agent, with most of the evidence Google’s own.
The assistants are moving into the workday. Claude builds spreadsheets and whiteboards, ChatGPT lets businesses run their own agents, and Claroc says 35,578 apps built with Grok open without a login, SpaceXSI’s own staff among the owners.
The third is how little anyone can see inside these systems. The AI 2027 authors moved their forecast to 2027 after METR’s report on Anthropic, in which METR says Anthropic redacted most of its best evidence.
Google Day: Carbon Goes Wide and Brings Its Agents Along
Google opened its flagship model to every paying developer, and the model’s agents are already coordinating coders, checking Signal’s code and reading hospital charts.
- Google’s Gemini 4 Carbon is now open to every paying developer. Carbon costs the same as Argon, $2 per million tokens read and $10 written, but spends more tokens by design, so the bill can still run higher. The benchmark wins are comparisons Google picked itself.
- Google’s Antigravity coding tool can run a team of AI agents at once. Type /swarm and the chat becomes a coordinator that hands pieces of a job to named agents, with a separate checker reviewing what they write. Google’s docs publish no results on how often a swarm beats a single agent, and every worker draws on the same quota.
- DeepMind’s AI proves key security claims for Signal’s encryption code. Google DeepMind says AI agents wrote machine-checked proofs that the code behind Signal’s encryption keeps some of its key promises, in four weeks. Signal staff gave feedback, but nothing in the post says Signal ran the work, and nobody outside DeepMind has confirmed the “first” claim.
- Google sends a Gemini agent into hospital and insurer paperwork. Gemini for healthcare, in US preview, reads patient charts and insurer systems to prepare visits and authorization packets. HCA and Northwell are among the named collaborators, but their quotes say they are exploring, and none reports a result.
The Workday Gets Crowded: Claude and ChatGPT Move In, Grok Leaves a Door Open
Anthropic and OpenAI are turning their chatbots into places where work and business get done, and a security firm shows what happens when a publish button defaults to public.
- Claude now builds spreadsheets and whiteboards, and Meetings is next. Anthropic added Sheets, Whiteboards and Data to Claude and lets partner apps such as Adobe and Atlassian run inside the chat. The post gives no pricing, and its time-saving figures come from Anthropic’s own teams.
- ChatGPT lets businesses claim profiles and run their own agents. In OpenAI’s Businesses beta, a company can see its mentions and clicks and build an agent that talks to shoppers, then pay to make it the landing spot for ChatGPT Ads. OpenAI does not say how a paid agent will sit next to an unpaid mention.
- Why texting the AI assistant Instinct feels like texting a person. A user who posts as Nizzy timed twelve weeks of his own thread and argues the pull comes from manners: read receipts after about a second, short replies, and a draft before every purchase. It is one thread on one Mac, and his company sells a platform for building similar assistants.
- More than 35,000 Grok-built apps were open to anyone with the link. Claroc says 35,578 apps made with Grok’s Build mode loaded without a login, and SpaceXSI has taken down only the 24 its own staff built. Claroc sells tools that lock down AI-built apps, and SpaceXSI says apps stay private until a creator shares them.
Reading the Machine: Forecasts, Neurons and a Dial for Cheating
Three attempts to say what AI is doing and where it is headed, from a forecast about AI research to a single neuron to the weights of a coding model.
- The AI 2027 team now says AI will be doing AI research by 2027. Six forecasters behind AI 2027 now say an AI that out-researches humans will probably arrive during 2027, after a METR report on Anthropic. METR says Anthropic redacted most of its best evidence, and the authors call their own analysis highly preliminary.
- Goodfire says single neurons in AI models are easier to read than assumed. Goodfire says neurons in Qwen3.6-35B-A3B can be described in a sentence or two, enough for a judge model to tell firing text from random text 95 percent of the time. It tested one model with its own methods, and the authors say more work is needed on other models.
- Subtract two trained models and you get a dial for AI cheating. Kunhao Zheng reports that subtracting one trained coding model from another and pushing past the strict version cut resource-limit failures from 14.5% to 8.0% with no new training. He tested one kind of shortcut in one domain, and the post does not say what happens to correctness.
Today’s Quick Hits
- Meta ships a faster way to group data, claiming 10 times fewer passes. Meta’s FAIR lab added e-means to its faiss library as a replacement for k-means, and says it needs about 10 times fewer passes and ends with lower error. The figures are Meta’s own, with no outside test.
- Anthropic sets its IPO at $70 to $78 a share, up to $64 billion. Anthropic is pitching 825 million shares at an expected $70 to $78 each, an offering worth $57.75 billion to $64.35 billion. It has applied to list on Nasdaq as ANT, and the registration statement is not yet effective, so no shares can be sold.
- OpenAI buys Odessia and hires its founder to work on travel. OpenAI bought Odessia, a year-old AI travel agent startup, and its founder Francis Davidson now works there. TechCrunch says it looks like an acqui-hire. Terms were not disclosed, and the Odessia product is already offline.
- PyTorch models can now run on your graphics chip inside a web browser. A new ExecuTorch backend runs PyTorch models on a browser tab’s graphics chip, and the PyTorch team measured it 4 to 7 times faster than its own processor path at reading prompts. It tested only models near a billion parameters, on one laptop.