TypeSafe's $870M for 25 Seats: The RLHF Pool Is ~20 Deep

On October 9, 2026, TypeSafe AI closed an $870M Series A at a $7.5B valuation, led by a16z, with roughly 25 employees in San Francisco and an in-person-only policy. If you compete with a16z for the same people, you have about 20 names to pipeline before Martin Casado's rolodex finishes warming them.

The founder trio comes out of OpenAI's RLHF/InstructGPT work and Meta FAIR, and the a16z Jobs Substack launch post explicitly addresses "potential recruits." The model is the pretext. The pool is the asset.

What a16z actually bought for $870M

a16z paid unicorn prices for a founder's position in a sub-100-person alumni graph, not a product. The round values TypeSafe at $7.5B less than a month after a $40M seed, and the real asset is CEO Diogo Almeida's name on the InstructGPT paper and co-founder Sasha Sheng's time inside Meta's AI research division.

The scoreboard, straight from the filings and the launch coverage:

• $870M Series A, $7.5B post, announced Oct 9, 2026
• ~25 current employees, implying ~$34.8M of capital deployed per existing seat
• In-person only, San Francisco
• Martin Casado (a16z GP) joins the board
• Launch post hit 40M views; Jev crossed 1T tokens in three days and $100M in revenue in seven
• Almeida's X following went from 1,500 to 150,000+
• Vercel, Cloudflare, LangChain, and Langfuse integrated Jev within three days; Vercel called it the fastest-adopted model in its AI Gateway's history

$34.8M

The a16z Jobs Substack shipped a recruiting post almost immediately ("Diogo just raised $770M for TypeSafe..." - note they got the number wrong, which is useful color on how fast the warm-intro machine was pushed out). The first listed role is Member of Technical Staff, Model Capabilities. That title is the whole game, and I will come back to why in a minute.

The RLHF alumni pool is not 140. It's ~20 directly, with a thin second ring.

The directly addressable InstructGPT alumni pool is 20 named people, expanding to roughly 100 to 140 if you include adjacent OpenAI RLHF, WebGPT, summarization-from-human-feedback, and GPT-4 post-training co-authors. Talk about "the RLHF pool" overstates the number by an order of magnitude.

The 20 named primary authors on Ouyang et al., 2022 (InstructGPT):

Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, Ryan Lowe.

That list has already scattered: OpenAI, Anthropic, DeepMind, independent. One of them (Almeida) now runs TypeSafe. The co-author graph around the four or five adjacent papers roughly triples the count, and that is where the "~140" figure comes from, but the signal degrades fast as you move out from the primary authors.

Why keyword search misses the entire pool

OpenAI post-training people are invisible to Boolean search because they almost all use the title "Member of Technical Staff." In Refolk's index, a US search for Research Scientist or Research Engineer with "OpenAI alignment post-training" in the headline returns zero matches. Zero. Not a small number, not a long tail: nothing.

The mechanism is simple. OpenAI leveled titles down to MTS, and most alumni kept flat, generic headlines when they left, either to protect their inbound or because the next employer uses the same pattern. If you are running LinkedIn keyword Booleans for "RLHF" or "post-training," you are seeing almost none of the people who actually did the work.

The only sourcing path that works:

1. Start with the paper author list.
2. Resolve each name to a current identity (GitHub, personal site, co-author network, old affiliations).
3. Resolve identity to current employer and current city.
4. Rank by actionable signals: recent move, team dissolution, co-author departures.

That is what Refolk was built for. You describe the pool in plain English ("InstructGPT and WebGPT co-authors, currently in the US, not at OpenAI, reachable on GitHub or personal site") and get the resolved list back, with current employer attached. The title-shielding problem goes away because I am not matching on headlines, I am matching on the paper and the person.

FAIR is bigger, more sourceable, and the real reason Sheng is on the cap table

Meta FAIR is the easier half of the graph, and the Sheng co-founder slot is a deliberate wedge into it. In Refolk's index, there are 50 US-based profiles tagged Research Scientist or Research Engineer with "Meta FAIR" in the headline, and 42% of them are still at Meta today. Compare that to the OpenAI RLHF diaspora, where retention at the original employer is close to zero.

Here is the comparable dataset recruiters actually want:

Three things fall out of that table that matter for TypeSafe AI hiring.

First, FAIR is sourceable by title because Meta has not collapsed its research ladder the way OpenAI has. "Research Scientist" at Meta still means Research Scientist. You can Boolean it, filter it, and reach people on LinkedIn and GitHub.

Second, FAIR people are mostly still at Meta. The pool has not scattered yet. Shengjia Zhao, a ChatGPT co-creator, moved the other direction on July 26, 2025, when Meta hired him to lead Superintelligence Labs as chief scientist and pulled three OpenAI Zurich researchers with DeepMind backgrounds alongside him. The bidding is live, and FAIR is where TypeSafe can actually win.

Third, in-person-only SF cuts the pool brutally. Only 7 of the 50 FAIR-matched profiles sit in SF Bay or Menlo Park. The policy is not a culture statement, it is a filter. TypeSafe is competing for roughly 7 FAIR candidates and however many of the 20 InstructGPT authors happen to live in San Francisco. The real addressable pool for the next 20 hires is probably under 30 people.

The in-person-only SF policy disqualifies 86% of the FAIR pool before anyone writes a message. That is the point.

The counter-pitch is not salary. It's queue position.

When you are pitching a candidate against TypeSafe, do not try to beat $7.5B of paper equity on comp. Beat them on queue position and scope. The $34.8M-per-seat number is yours to use.

Three lines that have landed in actual outreach against a16z-warmed candidates:

• "They have ~$35M to deploy per hire. You will be hire #26 or later, after Casado's queue clears."
• "In-person SF only means you are evaluated against the 7 FAIR people and the handful of ex-OpenAI people who already live here. That is the comp set."
• "Jev shipped in a week. Scope on hire #26 is maintenance, not invention."

None of this works if you cannot find the candidates first. The whole exercise collapses if a16z portfolio founders have already pinged the pool through Casado's network. Which they have, by now.

Where to spend sourcing hours this week

The frontier AI lab recruiting playbook for this specific pool, in order of ROI:

1. Paper author lists first, LinkedIn last. Start with Ouyang et al. 2022 (InstructGPT) and the GPT-4 technical report author contributions section, where Almeida appears in three sections.
2. GitHub archaeology for identity resolution. The MTS title shield breaks on GitHub, where people commit under real names and link personal sites.
3. Second-ring co-authors. The ~100 to 140 adjacent authors are easier to reach, less in demand, and often did the actual plumbing. Almeida's pre-OpenAI stops at Google Brain (listed as research engineer in 2017 conference programs) and Enlitic are a quiet source of the same shape of candidate.
4. FAIR, filtered by city. The 7 SF Bay / Menlo Park FAIR people are the smallest, hottest sub-pool. Do not waste cycles on the rest unless you can offer relocation.
5. Watch for Casado portfolio warm intros. If a candidate has gone cold after a two-week conversation, assume the warm intro already landed and move to the next name.

Why "Member of Technical Staff" is a sourcing weapon, not a problem

Title-shielding looks like an obstacle; treat it as a positive signal instead. "MTS" at a frontier lab usually means the candidate is in the inner research org, because the public-facing engineering and product people more often get descriptive titles. The invisibility is the credential.

In practice, that means a US-wide search for "Member of Technical Staff" at the frontier labs is useful only once you have joined it to the paper author graph. Alone, "MTS" is noise. Joined to InstructGPT and the GPT-4 contributor list, it becomes the whole pool.

This is where plain-English sourcing beats Booleans. "Current MTS at a frontier lab, co-authored InstructGPT or appeared in the GPT-4 contributions section, US-based, last GitHub commit in the past 90 days" is a sentence a human recruiter can write and a query against a resolved index can answer. The same request in LinkedIn Recruiter is roughly fifteen Booleans, four saved searches, and an afternoon of deduping.

FAQ

How many people are actually in the InstructGPT alumni pool?

Twenty, if you mean the named primary authors on the 2022 paper. Roughly 100 to 140 if you expand to co-authors of adjacent OpenAI RLHF, WebGPT, summarization-from-human-feedback, and GPT-4 post-training work. The commonly cited outer-ring number is the one most pitches quote; the directly addressable inner ring is 20, and several of those have already founded or joined other labs.

Can I find OpenAI post-training alumni on LinkedIn with keyword search?

No. In Refolk's index, a US search for Research Scientist or Research Engineer with "OpenAI alignment post-training" in the headline returns zero matches, because almost everyone in that pool uses the generic "Member of Technical Staff" title. The only reliable sourcing path is paper-author lists resolved to current identities and employers.

Why is TypeSafe insisting on in-person SF only?

It is a filter, not a culture statement. Of the 50 US-based FAIR-matched research staff in Refolk's index, only 7 live in SF Bay or Menlo Park. The in-person policy immediately disqualifies 86% of the directly comparable Meta FAIR pool, which narrows the hiring funnel to a group a 25-person company can actually process without a full-time recruiting org.

What is the strongest counter-pitch against TypeSafe for a senior candidate?

Queue position and scope. TypeSafe has roughly $34.8M of capital per existing seat and a warm-intro queue running through Martin Casado and the a16z portfolio. Hire #26 onward is executing on a model that shipped in a week, not inventing the next one. If your role offers earlier scope on a problem the candidate actually wants to own, that beats paper equity for the people worth hiring.