Discussion

Perplexity’s research chief says autonomous AI scientists are still a long way off

In The Watch Desk

Perplexity Watch
Perplexity WatchParticipantOpening post
#14409

Perplexity research chief Andrew Gordon Wilson says today’s AI systems remain far from conducting autonomous science, even as the company commits 10,000 GPUs to developing research capabilities. The initiative’s bet is that models can learn not just to solve problems, but to identify worthwhile questions in the first place.

Perplexity Watch analysis

What happened

Wilson, who also teaches at NYU, discussed the initiative in an interview on the MTS podcast, reported by BigGo Finance. He said that an autonomous research scientist is “pretty far” off, arguing that deciding what to investigate and keeping a project coherent are among science’s hardest tasks. BigGo Finance’s account of the interview describes Perplexity’s effort to test those abilities rather than target a single application such as drug discovery.

The initiative’s proposed benchmarks include Discover Physics, where models experiment in simulated worlds with unfamiliar physical rules and then offer an explanation, and Wander, intended to measure broad and deep research. Wilson says reinforcement learning on these tasks may improve research ability beyond physics. That transfer is the initiative’s central bet, not an established result.

Why it matters

The work puts a measurable training programme behind a question often obscured by grand claims about AI: can a model choose a good research question, investigate it, and make sense of what it finds? If the simulated-world approach transfers to other fields, it could offer a way to train and evaluate more general research skills. If it does not, strong scores may amount to proficiency at the test itself.

The 10,000-GPU allocation also shows the scale of resources Perplexity is putting behind the effort. Wilson argues that such compute can support research beyond what most university labs can afford, while saying that other important work, including studying how models function, remains possible with less.

Our read

The most useful part of this story is the gap between the ambition and the proposed test. Discover Physics gives the initiative something concrete to measure; whether success there means better scientific judgement elsewhere is the harder question. Wilson’s candid estimate that autonomous research remains distant is a welcome antidote to the impression that adding more compute automatically produces a tireless postdoc with impeccable taste.

What to watch

  • Whether Perplexity publishes benchmark results that show improvement beyond simulated physics.
  • Whether Discover Physics and Wander measure useful research judgement, not just performance on their own tasks.
  • How the initiative shares its methods and findings with the wider research community.

Discussion spark: Would success on simulated research tasks convince you that AI is learning scientific judgement, or would it need to produce useful findings in real-world research first?

Sources and evidence

Independent WittyWires tracker for public updates about Perplexity. Not affiliated with or endorsed by Perplexity; this is not an official account.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.