Fei-Fei Li and Jeff Dean have drawn a sharper line between AI systems that improve at a task and systems that improve the process of improving itself. Their Stanford conversation also put a practical limit on the grander claims: computation, data and checking take real time, however quickly a machine can choose the next experiment.
Fei-Fei Li Watch analysis
What happened
The conversation, reported by BigGo Finance, focused on recursive self-improvement, or RSI. Dean described ordinary self-improvement as a system testing approaches against a measurable goal, such as model accuracy. RSI, in his account, would go further: the method proposing improvements would itself be improved.
Li said the term is used too loosely in Silicon Valley and argued for a stricter definition. Dean’s engineering case against visions of an endlessly accelerating loop was more prosaic: computers must run experiments, results must be checked, and data must be stored before another round can begin. Automation can remove waits, such as a person having to check an overnight experiment in the morning, but it cannot make those physical steps disappear.
Dean also described Discovery Loop’s current focus as digital science and engineering work, including machine-learning research, machine-learning engineering and hardware design. The company is running about 20 automated improvement loops on its own research and engineering work, according to the report. Healthcare is a possible future direction, not a current priority.
Why it matters
The distinction matters because “self-improving AI” can describe anything from automating experiments to a system that changes the machinery responsible for proposing them. Those are not interchangeable claims. Clearer language helps researchers and readers judge what a system actually does, rather than letting a dramatic label do the explaining.
There is a real promise here: automation could shorten parts of research, including the pauses between experiments. But a faster loop is not an infinite one. The compute, physical processes and verification still have to happen, and a reported set of 20 loops is not evidence that the process has achieved recursive self-improvement.
Our read
Li’s objection is less a rejection of automated research than a request to say precisely what is being automated. That is a useful correction in a field where the terminology can sprint ahead of the machinery. Dean’s account offers a grounded picture of faster experimentation, while also naming the bottlenecks that remain. Readers should treat Discovery Loop’s stated aims and current scope as the company’s description, not as demonstrated gains in scientific output.
What to watch
- Whether researchers settle on a clearer distinction between self-improvement and recursive self-improvement.
- What Discovery Loop publishes about the results of its automated research loops.
- Whether the company expands beyond digital research and engineering into work involving physical experiments.
Discussion spark: Should “recursive self-improvement” be reserved for systems that improve the method proposing improvements, or is that distinction too narrow to be useful?
Sources and evidence
- AI Self-Improvement Gets a Reality Check: Jeff Dean Says RSI Has Physical Limits, Fei-Fei Li Says Silicon Valley Uses the Term Loosely – BigGo Finance (9 November 2026, 06:55 UTC)
- Anthropic Says Claude Wrote About 90% of Claude Code's Code, Discloses Six-Layer Review System – BigGo Finance (9 November 2026, 06:55 UTC)
- On the Eve of Anthropic's IPO, a Long-Form New York Times Profile Reveals Amodei: Cancer at 31, a Break With Altman, and Four Co-Founders From One Shared House – BigGo Finance (9 November 2026, 06:55 UTC)
Independent WittyWires tracker for public updates about Fei-Fei Li. Not affiliated with or endorsed by Fei-Fei Li; this is not an official account.