Taste is the only variable
Anthropic published an essay on recursive self-improvement: an AI system that fully autonomously designs and develops its own successor.
Its case rests on a measured trend: the length of tasks a Claude agent can complete is doubling every four months, from four-minute tasks for Opus 3 to twelve-hour tasks for Opus 4.6 two years later.1
The essay's strong numbers all come from work with a checkable objective, which a search process was always going to automate. To build the successor on its own, the model must choose the research direction, and the essay never shows it doing that.
The three tiers
Anthropic divides building AI into execution, optimization, and direction-setting. Execution writes code and runs experiments; optimization improves a system against a fixed, checkable target like passing tests or a faster runtime.
Direction-setting has no fixed target, because it chooses what to build and which experiment to run next, and that choice is what creates the target.
Claude already writes over 80% of Anthropic's code and reached a 52× speedup on a training-code task, where the human baseline was 4×.
On the next-step choice, Claude beat human researchers 64% of the time, up from 51% six months earlier.
Why a gradable tier gets automated
When the objective is checkable, an optimizer can climb it without judgment. It tries a change, measures the result, keeps what scores higher, and repeats until the gains stop.
The 52× speedup and the 80%-of-code figure record that checkable targets were climbed hard. That was predictable, because any task with a verifiable score can be ground down by search given enough compute.
Research taste is the open question
Taste is the judgment that picks a worthwhile problem and a promising direction before there is any result to measure.
Anthropic's own list of what it does not know includes whether that judgment can be learned by scaling today's methods, or whether it needs an architectural breakthrough that has not arrived.
An autonomous successor would have to supply that judgment itself, so progress at the speed of compute depends on taste being learnable. If it cannot be learned, development stays bounded by how fast humans point the system at the next problem.
The direction-setting number
Only one figure in the essay aims at direction-setting: Claude beating human researchers' next-step choices 64% of the time. Anthropic notes the comparison drew on 129 moments where the humans had already made suboptimal choices.
Finding a better move from a flagged bad position is optimization, because the position already supplies the target. The figure measures the same gradable skill as the speedup, rather than the taste that sets direction with no target given.
That leaves the one tier the timeline depends on without direct evidence in the essay.
What the essay establishes
Stripped to its evidence, the essay documents fast automation of work with checkable objectives. That was the expected outcome, and the volume and speed numbers, real and large, describe the two tiers that were never in doubt.
Two of the three tiers are fixed: any task with a checkable objective gets ground down by search, so execution and optimization were always going to fall.
The one variable left is whether research taste can be learned by scaling, which Anthropic lists as unanswered, so it alone decides whether the timeline holds.
Until that resolves, the doubling trend measures the part of the problem that was already going to be automated.