Refolk
November 16, 2026·10 min read

Mirendil Hires Its CEO's Co-Authors. Only 45 US Profiles Say Pretraining

Mirendil's CEO is hiring his co-authors, and only 45 US profiles say pretraining. How to source AI researchers from co-author graphs, not job titles.

how to recruit AI researcherssourcing machine learning researchersfind pretraining engineersrecruit ex-DeepMind researchersAI research engineer sourcing arXiv
Mirendil Hires Its CEO's Co-Authors. Only 45 US Profiles Say Pretraining

If you recruit for a frontier lab or a well-funded AI startup, you can already see who Mirendil is likeliest to hire next: the co-authors on its founders' papers. Their LinkedIn titles won't tell you that, so you have to go to the papers.

On Nov 12, 2026, Mirendil announced a $1B Series A, 141 days after its $200M seed, and it still calls itself a small team. The CEO's own website already lists three past collaborators on staff.

What did Mirendil raise, and what is the money for?

Mirendil raised a $1B Series A led by Kleiner Perkins and Clearlake, less than five months after a $200M seed. Pomegra reports the money will go to compute and hiring. The company itself hasn't published a split.

Here is what's on the record:

  • Series A: $1B, announced Nov 12, 2026. Kleiner Perkins and Clearlake led. Battery, Doerr Capital, Gil & Co, Google, M13, Pear VC and S32 also took part.
  • Seed: $200M at a reported $1B valuation, co-led by a16z and Kleiner, with Nvidia joining.
  • Gap between rounds: 141 days, by FourWeekMBA's count.
  • Team size: about 20 researchers and engineers at the June 24 launch. The Series A release only says "still a small team," and Pomegra puts it at "roughly two dozen."
  • Use of funds: Pomegra says compute and hiring. GrowthList says Mirendil hasn't disclosed how it will spend the money.

A jobs page showed 16 open roles, including MTS Pretraining Data, MTS Compute, Founding GTM and Founding Finance. That page was about 47 days old when checked, so look at it again before you plan around it. Whatever the exact count is, a company that had about 20 people at launch and has just closed a $1B round is going to hire, and it will start with the people its founders already know.

Is Mirendil already hiring from its founders' co-author graph?

Yes, and the CEO's own website shows it. Behnam Neyshabur's personal site lists Preetum Nakkiran, Xiaoxia Wu and Anna Golubeva, all past collaborators, as current Mirendil Members of Technical Staff.

A co-author graph is the network of people linked by shared papers. Each paper connects every pair of its authors. Neyshabur's graph is a good one to map. His CV lists him as Senior Staff Research Scientist and Gemini Team Lead, and from 2019 to 2024 he co-led Blueshift, the team behind Minerva and the reasoning work in the first Gemini, Gemini 1.5 and Gemma 2. WOWTALE describes his role as driving pretraining for math and code reasoning on Gemini.

The rest of the founding group widens the graph:

  • Harsh Mehta (CTO): built the first version of Anthropic's internal autoresearch platform. He met Neyshabur at Google in 2019.
  • Shayan Salehian: ex-xAI.
  • Tara Rezaei: MIT.

The co-author graph predicts more than the hiring. The advisor and angel list includes David Dohan and Yuhuai Wu, both Minerva co-authors, along with Sanjeev Arora. So the paper graph overlaps with the cap table as well as the org chart.

There's a practical reason this happens. A lab that went from seed to Series A in 141 days doesn't have time to vet strangers, and someone you've shipped a paper with has already passed the hardest interview there is.

None of this is hidden, and that looks deliberate. Naming current MTS on the CEO's site tells former collaborators their peers have already made the jump, and every competing lab can read the same page and pull the same Scholar profiles. The edge is speed, not access: who reaches the Minerva authors and the founders' small-paper co-authors first, with a specific message about their own work.

The co-author graph is public. The only edge left is who reads it first and calls with the right paper in hand.

Why can't you find pretraining researchers by job title?

Frontier labs give almost everyone the same title, so searching by title mostly returns noise. In Refolk's index there are 15,961 US Members of Technical Staff, and only 45 US research profiles mention "pretraining" in their headline at all.

Member of Technical Staff (MTS) is the flat, catch-all title that OpenAI, Anthropic and other labs use for researchers and engineers alike. It deliberately says nothing about specialty. In Refolk's index of professional profiles, OpenAI and Anthropic each account for 10 of a 25-person sample of US MTS profiles. Filter by that title and you get a list of everyone at those labs, whatever they actually work on.

Adding a keyword doesn't fix much:

45
US research profiles with "pretraining" in the headline
Research Scientist, Research Engineer or MTS titles only, in Refolk's index.

Of a 25-person sample of those 45, 19 hold the MTS title. That means 76% of the sample sits under a title that tells you nothing about their specialty. The top employers in that sample:

  • OpenAI (5)
  • Google DeepMind (3)
  • SpaceX Super Intelligence (3)
  • Essential AI (2)
  • Reflection (2)
SegmentCountSource
US, "pretraining" in headline, RS/RE/MTS45Refolk's index
UK, same query2Refolk's index
All US Members of Technical Staff15,961Refolk's index
Gemini 2.5 report authors~3,298arXiv 2507.06261v3
Minerva paper authors~15Google Research blog

Divide 45 by 15,961 and you get about 0.28%. That is a ceiling on how many US MTS profiles self-label as pretraining people (the title sets differ slightly). The other 99.7% may include plenty of pretraining researchers. They just don't say so. Treat the profile headline as the least informative field on the page.

The UK looks empty. It isn't.

Run the same pretraining query in the UK and Refolk's index returns 2 people, both at Cohere. That isn't a talent desert. It means UK researchers don't put this work in their headlines.

The US-to-UK ratio for self-labeled pretraining researchers is 22.5x. Researchers at DeepMind London clearly exist, yet nobody from DeepMind London shows up in the results. The researchers are there. Their profiles just don't use the word.

This is where title-based searches go wrong. A recruiter trying to recruit ex-DeepMind researchers in London runs the obvious query, gets two Cohere profiles, and tells the hiring manager the market is empty. Then the search drags into a second month of InMails to generic "Research Scientist" profiles while the role stays open and the lab across town closes the people who were on the right papers all along.

In the UK you have to start from the papers, because the profiles give you almost nothing. Pull arXiv author lists with London affiliations, then resolve each name to a current employer.

Which co-authors are actually worth sourcing?

The fewer authors a paper has, the more each name on it is worth. A 4-author paper tells you who did the work. A 3,298-author technical report mostly tells you who worked at the company that year.

3,298
names on the Gemini 2.5 technical report on arXiv
About 73x more than every US profile in Refolk's index that self-labels pretraining.

The arXiv listing for the Gemini 2.5 report shows Comanici "and 3297 other authors." That's useful for one thing: confirming someone was at Google during that period. It can't tell a pretraining lead apart from someone who touched the project briefly. Compare the papers in the Mirendil founders' own history:

PaperAuthorsWhat a name tells you
Gemini 2.5 report~3,298Employment, not contribution
Minerva~15Real contributor to math reasoning
Gated state spaces (Mehta, Gupta, Cutkosky, Neyshabur)4Direct collaborator of both CTO and CEO

Here's how I'd weight names when sourcing machine learning researchers from papers:

  1. Score each co-authorship by 1 divided by the author count. A shared 4-author paper counts for far more than a shared 3,298-author report.
  2. Pay attention to position. On small ML papers the senior author is often listed last. A 4 to 15 author paper with the founder last or in the middle is where the strong signal is.
  3. Prefer repeat co-authors. Two small papers with the same founder is a stronger relationship than one.
  4. Use giant contributor lists as a filter. A name on the Gemini report tells you where someone was. Then check whether that name also appears on a smaller paper.

The Minerva list (Lewkowycz, Dohan, Yuhuai Wu, Michalewski, Misra, Gur-Ari and others) is the model here. It's small enough to read, and it's tied directly to the work Neyshabur is known for. Two of its authors, Dohan and Yuhuai Wu, are already on Mirendil's advisor and angel list.

How do you mine a co-author graph, step by step?

Start from the founders' Google Scholar profiles, keep only the small papers, and resolve every co-author to where they work today. This is what AI research engineer sourcing from arXiv looks like in practice:

  1. Pick the seeds. For Mirendil that means Neyshabur and Mehta. Open their Google Scholar profiles.
  2. Limit the time window. For Neyshabur, focus on 2019 to 2024, the Blueshift years, because those collaborators know his current working style.
  3. Drop large papers. Remove anything above roughly 15 authors, and treat the big Gemini and Gemma reports as employment checks only.
  4. Build the list. Record every remaining co-author along with the paper title, year and author count.
  5. Remove people who are already hired. Neyshabur's site names Nakkiran, Wu and Golubeva at Mirendil. Cross them off, and use them as second-degree seeds.
  6. Resolve current employers. This is the slow part. A name on a 2021 arXiv paper doesn't tell you whether that person is at OpenAI, Reflection or a stealth startup today.
  7. Check GitHub for builders. For the MTS Compute and MTS Pretraining Data roles, commit history on training frameworks and data pipelines tells you more than papers will.

Step 6 is where most sourcing projects stall: forty names, forty manual searches across LinkedIn, personal sites and GitHub. Refolk is built for exactly this. You describe the people you want in plain English and get back matched profiles across GitHub, LinkedIn and the open web, so a co-author list turns into a contactable shortlist without forty open browser tabs.

What does Mirendil's pitch say about who it hires next?

Mirendil's own claims suggest it wants a small number of senior people who can supervise AI agents, not big research teams. Its open roles also go beyond research, into compute, data, GTM and finance.

The company says one researcher working with agents pretrained Mirendil-Base in 3.5 months on under $10M of compute, and that agents wrote more than 99% of the code. Those are Mirendil's own figures and haven't been independently verified. If you take the claim seriously, though, it changes who they hire. A lab that thinks one person plus agents can pretrain a base model needs people senior enough to judge what the agents produce. It doesn't need a dozen junior researchers to write training loops.

For a competing recruiter, that means:

  • Research hires come from the co-author graph, weighted toward people with lead roles on small papers.
  • Compute and data hires come from GitHub and infrastructure signals, because these people often don't publish.
  • GTM and finance hires are ordinary startup searches. The co-author graph doesn't help here.

You'll be competing for the same people with the labs that show up in Refolk's index: OpenAI, Google DeepMind, SpaceX Super Intelligence, Essential AI and Reflection. Watch Essential AI and Reflection most closely. They're smaller labs, and they already employ self-labeled pretraining researchers, so they're fishing in the same small pond. Map the small papers and start outreach the week of an announcement, not after the role has been open for a quarter.

FAQ

How do you recruit AI researchers when everyone's title is Member of Technical Staff?

Skip the title. Start from papers and contributor lists. Find the founders or team leads you're competing with, pull their co-authors from Google Scholar and arXiv, keep only the papers with roughly 15 authors or fewer, and resolve each name to a current employer. In Refolk's index only 45 US research profiles mention pretraining in their headline, out of 15,961 US MTS profiles, so the title field won't find most of the people you want.

Are giant tech-report author lists like Gemini's useless for sourcing?

They're useless for judging what someone contributed and useful for confirming where they worked. The Gemini 2.5 report has about 3,298 authors, so being on it shows employment at Google during that period and not much else. Use it as a filter, then look for the same names on smaller papers, where each author is a meaningful signal.

How do you find pretraining researchers in London?

Go through arXiv affiliations and paper author lists, not profile headlines. Refolk's index returns only 2 UK profiles that self-label pretraining, both at Cohere, and none from DeepMind London. The researchers exist. They just don't describe their work in their headlines, so title and keyword searches will make the market look empty when it isn't.

The data behind this

Try it on the search you came here for

Stop building boolean strings. Just describe the person.

Type one sentence. I plan the search, read GitHub, public LinkedIn and Crunchbase records, and the open web as it is right now, and hand back a ranked list with the reason next to every name.

  1. 01Describe them

    One plain sentence. Role, city, stack, stage, whatever matters to you.

  2. 02I read the web live

    GitHub, public LinkedIn and Crunchbase records, the open web. Not a database that went stale last quarter.

  3. 03You read the shortlist

    Ranked, with the reasoning under every name. Open a profile, ask a follow-up, narrow it down.

  • No boolean, no filters, no seat to buy. One box.
  • Read at search time, so a profile updated yesterday counts today.
  • Every step visible as it runs, every name with its reason.

500 free credits on sign-up. No card, no demo call. See real searches.

Read next