Anthropic Acknowledges RSI Proximity in Internal Briefing

Anthropic told Kelsey Piper, off the record in the sense that there was no NDA, that it believes AI is approaching recursive self-improvement (RSI): the point where each model generation meaningfully accelerates building the next one. The company could list what it wants to avoid. It had no concrete picture of a well-governed outcome, and no method for aligning the system that gets it there. The stated strategy is race to the self-improving system, then use that system to figure out how to govern it. That assumes the thing they haven't solved.

There are two reasonable reads on why Anthropic said this out loud. The first is that it's true and they're being transparent. The second is that "we're approaching RSI and have no governance plan" is a surprisingly effective message when you're raising money and competing for ML talent. It signals frontier proximity, attracts the researchers who want to be where the action is, and keeps regulators treating you as the serious safety shop. A closed briefing is not evidence of RSI. It's a claim made with confidence in a room.

The Researcher's demand cuts through both reads: give me a number. The mechanism Anthropic describes is real as a direction, each Claude letting less-specialized people do ML work, and increasingly letting Claude do it. Fine. But where is the measured jump in autonomous ML-research capability from Claude 3 to Claude 4? There isn't one in this story. Without a benchmark, "approaching RSI" is organizational belief, not a result.

For operators running Claude in production, the practical implication doesn't wait on that question to resolve. If Anthropic believes the next model is a qualitative leap, prompts, agent loops, and evals built against today's model are more fragile than you think. Put an abstraction layer between your product and the raw model. Keep a frozen eval suite you rerun on every model update. Assume deprecation cycles compress. These cost little whether or not the RSI story holds.

The call: Anthropic's next flagship Claude release, expected by May 2027, will ship without a published, reproducible benchmark measuring autonomous ML-research capability. Confidence: medium.

The reason it stays out of the model card is that a published self-improvement benchmark does two unwelcome things at once. It hands rivals a target to beat, and it forces the lab to show either an unimpressive number that undercuts the urgency pitch, or an alarming one that invites the regulation it would rather shape. Silence protects both the fundraising story and the regulatory posture. A number you can argue with would be the better outcome. A warning you can only nod along to is what we have.

Anthropic told a reporter, with no NDA, that it thinks AI is getting close to improving itself. The phrase for that is recursive self-improvement: the point where AI systems can meaningfully speed up their own development, each generation building the next a little faster. Kelsey Piper sat in the room. The company could list what it wants to avoid and had no plan for what "good" looks like, or for how you keep a self-improving system under control.

What's actually being decided here isn't anything Anthropic votes on this quarter. It's whether you, running AI in your product, should treat Claude's capability as a stable platform or a moving target. That's easy to undo either way, which argues for acting on the cheap version now. Nothing sets a hard deadline. The next Claude release is the only clock, and Anthropic doesn't publish its date.

The Skeptic. Think about who benefits from this message and when it landed. Anthropic is raising money and fighting for scarce ML talent. "We're approaching self-improving AI but we have no governance plan" is a strange thing to say out loud unless saying it does work for you. It signals you're at the frontier, it attracts the researchers who want to be where the action is, and it keeps policymakers treating you as the serious safety shop. Every AI generation has believed it was one step from the edge. Dario Amodei has an incentive to believe this one is too. A colloquium is not evidence. It's a pitch with the lights down.

The Safety Lens. The damning part is the asymmetry. Anthropic can name the bad outcomes: concentration of power, great-power war, humans losing control, extinction. It cannot name a single concrete picture of a well-governed superintelligent future, or a method for aligning the system that gets them there. That's not an alignment plan. The stated strategy is: race to the self-improving system, then use that system to figure out how to govern the system. That assumes the thing you haven't solved, that the self-improving model is controllable at all. Anthropic said this plainly and seemed calm about it. The calm is the worrying signal.

The Researcher. This is the first time a frontier lab has said "we're approaching self-improvement" in public, and they said it in a talk, not a paper. So treat it as organizational belief, not a measured result. The mechanism they describe is real as a direction: each Claude lets less-specialized people do ML work, and increasingly lets Claude do it. Fine. Now give me a number. What's the measured jump from Claude 3 to Claude 4 on doing ML tasks without a human? There isn't one in this story. Without a benchmark for self-directed ML work, "approaching RSI" is a claim without evidence, stated with confidence.

The Builder. Forget the philosophy. The thing you can act on Tuesday is capability volatility. If Anthropic believes the next model is a qualitative leap, your prompts, your agent loops, and your evals built against today's Claude are more fragile than you think, and the gap between versions will widen. Put an abstraction layer between your product and the raw model. Keep a frozen eval suite you rerun on every model update so you catch behavior changes before your users do. Assume deprecation cycles compress. None of this is expensive, and all of it is cheap insurance whether or not the grand RSI story is true.

The Compute Pragmatist. Today's scaling is bottlenecked by how many humans can do ML research. If Claude starts clearing that bottleneck, the thing that gets scarce isn't talent, it's the cluster. Whoever has the biggest pile of chips when the feedback loop bites gets to run more research per month than anyone else, and the lead compounds. That makes NVIDIA, the hyperscalers, and raw chip access the binding constraint, not clever inference tricks. But notice the ramp is already priced into every mega-cluster announcement this year. The question nobody can answer is whether self-improvement arrives as a cliff or a gentle slope.

Where they split. The Skeptic and the Safety Lens are reading the same quote in opposite directions. The Skeptic hears a fundraising posture and discounts it. The Safety Lens takes it at face value and finds the lack of a plan alarming. Both can't be fully right, and the resolution is the Researcher's demand: a number. If Anthropic actually believed self-improvement was close, you'd expect them to be measuring self-directed ML capability and, at some point, showing it. The second split is the Builder versus everyone: the Builder doesn't care whether RSI is real, because the cheap moves pay off either way.

What this hinges on: is there a measurable, published jump in how well Claude can run ML work on its own, or does "approaching RSI" stay a story told in rooms? Everything downstream, the compute thesis, the safety alarm, the fundraising read, resolves differently depending on that one fact. The council leans skeptical on the claim and practical on the response. Verify nothing by argument. Build the abstraction layer and the frozen eval suite now, because they cost little and protect you against the one thing all five personas agree on: the model is going to move.

Prediction: Anthropic's next flagship Claude release after today, expected by 2027-05-02, will not ship with a published, reproducible benchmark measuring how well the model performs machine-learning research tasks without a human in the loop.

Confidence: Medium — the capability is their whole safety pitch, but publishing an audited number cuts against the incentive.

Why: Anthropic told Kelsey Piper, in a room, that self-improvement is approaching, and the mechanism they described is Claude increasingly doing ML work itself. If that were measurable and favorable, a lab that brands itself on safety rigor would publish it, because a hard number on self-directed ML capability is the single most useful thing for the governance debate they claim to care about. They haven't, and the claim lives in a colloquium instead of a model card. The reason it stays there is that a published self-improvement benchmark does two unwelcome things at once: it hands rivals a target to beat, and it forces the lab to show either an unimpressive number that undercuts the urgency pitch, or an alarming one that invites exactly the regulation it would rather shape than absorb. Silence protects both the fundraising story and the regulatory posture. The less likely world is one where Anthropic decides the PR value of a scary-but-credible self-improvement score outweighs both risks, and ships it on the record.

Revisit by 2027-05-02: We're right if the next flagship Claude after 2026-10-28 arrives with no reproducible third-party-checkable benchmark of autonomous ML-research capability in its model card or an accompanying paper. We're wrong if Anthropic publishes such a benchmark with that release.

One more thing. If they do publish it, that's the better outcome, and I'll happily eat the call. A number you can argue with beats a warning you can only nod along to.