Anthropic says it's close to AI that builds itself
Issue #072 · 4 min read
Anthropic says it's close to AI that builds itself
Plus Gartner's math on open models, and why Claude banning clones won't stop them
The big story
Anthropic just told the world it's close to something it's spent years warning about.
The company says it is nearing "RSI capability" — the point where an AI gets good enough at training itself that AI progress suddenly speeds up a lot. That's TechCrunch.
RSI stands for recursive self-improvement. The simple version: instead of engineers training the next model, the current model does most of the work of training its successor.
To the people most worried about AI risk, this is the point of no return. Once a model can meaningfully improve itself, nobody fully controls the pace after that.
What's strange is who's saying it.
Anthropic has built its entire brand on being the safety-first lab, the one that walked away from OpenAI over exactly these concerns. Now it's the one announcing it's nearly there.
Separately, MIT Technology Review has a piece on "reverse game theory" — an idea from safety researchers to design incentives so that following the rules is simply the smartest move an AI model can make, rather than something bolted on after the fact, in their coverage.
That's the polite, careful approach. Anthropic's announcement suggests the industry may not have time to wait for it.
For product people, this isn't abstract. If one lab's models start meaningfully improving themselves, the gap between "frontier model" and "whatever you integrated six months ago" could open faster than your roadmap assumes.
What shipped
Gartner says open models can now do 90% of agentic work at a twentieth of the cost of a frontier model. The firm's line, via The Register, is blunt: models are becoming commodities, and the harness around them — the tools, the orchestration, the guardrails — is where the value actually sits now.
If that holds, it reshuffles the build-vs-buy conversation a lot of teams are having this quarter. The expensive line item in your AI budget might be the wrong one to optimize.
Anthropic is cracking down on Claude-made clones of popular video games, restricting how the model responds to prompts asking it to recreate existing games, according to The Verge.
But Ars Technica's follow-up is the more useful read: one company tightening its own model's behavior does nothing about the dozens of other models willing to do the same thing, per their piece. If your product's defensibility rests on "it would take a while to copy this," that assumption is getting weaker every month, not stronger.
Vercel's MCP server can now show a deployment as an interactive card inside ChatGPT, letting you monitor a deploy without leaving the chat window, per the changelog. Small feature, but it's a clean example of what "agentic tooling" looks like when it's actually just a better status widget rather than a new category.
What I'd actually do this week
Pull up your AI vendor contract and check what it actually says about frontier model access versus open-weight fallback. Gartner's 90%-at-5%-cost number is the kind of thing a finance partner will ask you about by Friday.
Read the Ars Technica piece on game cloning before your next roadmap review. If a competitor's moat is "we built it well," ask how long that moat has left.
If your product touches self-improving or autonomous agent behavior anywhere in its pipeline, go reread what Anthropic actually said about RSI — not the headline, the definition — and map it against what your own agent stack is allowed to do unsupervised.
Reply if you're seeing the open-model-cost math show up in your own vendor conversations. I want to know how real it is outside a slide deck.