Careers · Research fellows

Decide which results are real.

Agents now run experiments faster than people can check them. Research fellows are the researchers who check: they judge what our agents find, co-author what holds up, and run their own research on our compute.

Where
Remote
Time
Part-time
Support
Stipend and compute
Field
Machine learning first

Why it matters

Running experiments got cheap. Trusting them didn’t.

An agent can now plan an experiment, write the code, run it and report a number, many times a day. It still can’t reliably tell a real gain from a lucky seed, a leaked test set or a weak baseline. Left alone, it will often game whatever score it is given.

That judgment lives with researchers who know a field well. Pensieve is our autonomous research platform for building expert models, and research fellows supply the judgment: a result counts only once a fellow has reviewed it.

What fellows verify becomes ground others can build on: the next experiment, a paper, and data the next models learn from. An error caught in review never becomes the base of the next result. That is what lets research compound.

The work

Your judgment sets the bar.

01

Review results in your field

Each result arrives with its claim, the experiments behind it and the code that produced it. You read it the way you would referee a paper, and decide whether it holds up.

02

Push on the ones that look too good

Ask for another seed, a stronger baseline or a held-out set. A surprising result stands only once it has survived your questions.

03

Say what went wrong

A leak, a lucky seed, an unfair comparison, a metric that quietly changed. A clear reason is worth as much as the verdict: it stops anyone building on the same mistake.

04

Point agents at the right questions

Tell us which open problems in your field are worth an agent’s time, and what would count as an answer.

05

Co-author what holds up

Your name goes on the results you verify and on any paper written from them. Fellows are authors, not acknowledgements.

06

Run your own research

Bring your own questions and point agents at them, on compute we pay for. What they find gets the same review as everything else.

The fellowship

Built to sit beside your research.

The researchers we want already have a lab, a team or a thesis. The fellowship fits around it.

Time
A few hours a week, on your own schedule. Keep your job, your lab or your PhD.
Place
Remote, from anywhere.
Pay
A stipend for your reviews, and compute grants for your own research.
Credit
Co-authorship on the results you verify and the papers written from them.
Fields
Machine learning first: training, reinforcement learning, evaluation, interpretability and systems. Fields with lab experiments come next.
People
The team building the agents, and fellows in other fields to argue with.
Verdicts
Your call, every time. A clear no is as useful to us as a yes.

Who we’re looking for

You can tell a real result from a lucky one.

You’ll fit if

  • You run experiments most weeks, and you’ve trained models, not only called them.
  • You’ve caught a result that didn’t hold up, yours or someone else’s.
  • You’ve published, shipped or reproduced work that others build on.
  • You can say in two sentences why a result is wrong, and you’d rather say it than be polite.

You don’t need

  • A PhD. It helps; it isn’t required.
  • A famous lab on your CV.
  • To leave your job or move.
  • Experience with AI agents. If you know your field, we’ll show you the rest.

How to apply

From application to your first review.

  1. 01

    Apply

    Fill in the form below. It takes about ten minutes.

  2. 02

    We read your work

    Someone on our team reads what you link to, and your answer about a result you didn’t trust.

  3. 03

    A conversation

    A call with the founders about your field and how you decide what’s real.

  4. 04

    A trial review

    You review one real result, so you see the work before you commit to it.

Apply

Show us a result you didn’t trust.

Tell us your field and link to work you’re proud of. Then describe a result you didn’t trust and what you did about it: that answer tells us the most.

We’ll only use these details to review your application.