For two weeks everyone has been talking about recursive self-improvement — RSI, models that speed themselves up until nobody can keep up. Nathan Lambert, who writes Interconnects for 83,000 subscribers from inside the frontier-lab world, spent Friday explaining why he still doesn’t buy it.
His counter-model is lossy self-improvement. Three things hold it up: the automatable slice of research is too narrow to beat the exponential costs baked into scaling laws. More agents running in parallel hit diminishing returns. And what actually gates a strong model — data centres, power, politics — is barely touched by AI at all.
The line from the system card
Lambert quotes the system card for Claude Fable 5.1 and Mythos 5.1. In it, Anthropic says internal use of recent models has been a key factor in holding the current rate of progress — but that it sees no clear signs of dramatic acceleration beyond that rate.
That squares with what we wrote last week about the R&D automation index: plenty of automation, no explosion. Lambert’s read of the internal numbers from OpenAI and Anthropic is that the automation runs deepest in software engineering, log monitoring and working through planned experiments. The chores, in other words.
The numbers from the podcast
More useful than the thesis are the timelines Lambert pulls out of two Dwarkesh episodes. John Schulman, Beren Millidge and Charlie O’Neill were asked when AI clears specific bars:
- Drop-in remote worker for a full month of white-collar work: O’Neill says about a year with programmatic access to the tools, two years if it has to go through a browser. Schulman says a year for an okay version. Millidge says three years for full generality.
- 10x productivity for AI researchers: Schulman two years. O’Neill five to ten.
- Beating top human experts at everything done on a computer: Schulman three to four years, Millidge around five, O’Neill five to ten.
Richard Ngo, whom Lambert quotes approvingly, frames the underlying problem neatly: the safety community will end up directionally right and factually wrong. No superintelligence in the next eight years — but things moving fast enough that it will feel like the short-timeline camp won.
Where machines keep failing at post-training
Schulman explains on the podcast why post-training teams stay large: somebody has to decide, area by area, how the model should behave. Then the line anyone who has ever made a model worse will recognise — it is very easy to break post-training in a way no benchmark catches.
Two exponentials, only one of them bends
Lambert’s strongest point is a distinction the debate keeps blurring. RSI bites where the target is clear and measurable: serving models, efficiency, cost per token. Those curves go steep, and Jevons’ paradox turns that into a good business.
The other curve is peak intelligence, and it still demands exponential compute for linear gains. His conclusion: until better evidence shows up, lossy self-improvement stays his baseline, and all the extinction talk strikes him as misplaced.
You don’t have to agree. But of everything written against the panic these past two weeks, this is the piece that states most precisely what would prove it wrong.