Skip to content
Patchnotes

Anthropic's RSI warning and Google's AI-only Search

Issue 14 min read

Anthropic's IPO filing raises recursive self-improvement, Google's school settings bundle Search with AI, and DeepSeek proposes agents that manage other agents.

In this issue

Anthropic says self-improving AI may be closer than expected

Anthropic’s IPO filing says future Claude models may soon reach the point where AI greatly speeds up AI research. People link that milestone to recursive self-improvement (RSI).

Claude already helps Anthropic run experiments, generate training data, write code and evaluate new models. That part was known. The new part is that Anthropic admits it has missed big speed-ups in its own development before and only noticed them after they happened.

No current Claude model has crossed that threshold. Anthropic does not claim AI can build its successors without humans. The line sits in a risk disclosure, the section where companies list what could go wrong. It is not a product announcement.

Sundar Pichai talked about RSI a few days earlier. Now Anthropic has put the same idea in front of investors. Two rival labs raising it in the same week is notable. One of them saying it might not see it coming is the bigger point.

Google will turn off Search for schools that reject AI

Google’s upcoming K-12 settings give schools more control over AI, with a catch. Search, Gemini and other AI tools will be off by default until administrators turn them on. There is no option to turn on the old Google Search without AI.

Google says Search and AI are now too tied together to separate. A school that wants no AI loses Google Search too.

The timing is awkward:

  • New York City and Los Angeles have already restricted student AI use.
  • Two US senators have pressed Google over its classroom push.
  • Former FTC chair Lina Khan accused Google of using its search dominance to pressure schools into accepting AI.

Google says this gives schools more choice, and on paper it does. Administrators can turn on Search without the separate Gemini app. They cannot get Search without AI. Schools that already run on Chromebooks, Docs and Classroom have little room to leave Google.

DeepSeek wants AI agents to manage themselves

DeepSeek published a technical write-up on Super-agent, an architecture for running many AI tasks without a person watching every session.

The starting problem: DeepSeek estimates people can track only 3 to 5 agent sessions at once. Past that, the human is the bottleneck, not the model.

The design looks like a small company:

  • A Router hands out work to Workers that run in parallel.
  • Agents monitor GitHub discussions, investigate bugs and open pull requests.
  • A Watchdog checks whether any agent is stuck or drifting off task.

In simulated tests, the 95th-percentile task time dropped from 29.5 to 15.7 minutes, nearly half. It used many more tokens, and the tests did not measure whether the code was good.

Getting one agent to do useful work keeps getting easier. Managing ten is a different job. DeepSeek is betting the next jump comes from agents coordinating with each other, so people review results instead of supervising every step.

The DeepSeek Harness prototype is open source. Super-agent itself is still a proposal, not a product.

Rebuilt with Opus 5.5

People are using Claude Opus 5.5 to rebuild large, well-known software from scratch and then open-sourcing it. Apps that took big teams years to build are being copied in days. Examples:

  • agent-desktop. Like agent-browser, but for your whole computer. Gives any coding agent computer use, and several agents can work in parallel, each with its own cursor.
  • Reverse Engineer Anything. Open-source toolkit that gives your AI agent reverse-engineering abilities through MCP and a CLI.
  • terminal-mac. A full macOS that runs inside your terminal. Computer use built in through MCP, a CLI and a skill. Free and open source.
  • Kino. Edit video with your agents.
  • Agent37. A low-cost sandbox for agents.
  • Archon. A video foundation model that turns raw footage into a finished edit in minutes, zero-shot.
  • Instinct. A personal assistant that learns what you’re working on and what matters to you.
  • Bastion. Closes your router’s backdoors, shows where your traffic goes, speeds up your Wi-Fi, and lets AI agents watch your network.
  • Paper. One design space, built on web standards, that connects your teams, agents, code and data. There is a video guide too.
  • Super-Agent. DeepSeek’s new open-source AI coding agent prototype.
  • Capy. Run a fleet of cloud coding agents inside the apps your team already uses.
  • Grok Bot. AI teammates you can hand real work to.
  • AlignCanvas. Turn images into 3D motion.
  • ChatGPT Voice. New Finances integration for talking to AI about your money.

Worth reading