Company teardown · Mirendil

Two Anthropic researchers want AI to do the job of two hundred AI researchers. Most of the team they claim is missing from their LinkedIn People page.

Mirendil's founders left Anthropic in December 2025, raised a $200M seed at $1B in June, and were reported in talks at $5B three months later. Their thirteen identified people have done most stages of model training before, at frontier labs, big tech or the labs' suppliers. Six are missing from the company's LinkedIn People page, and the job board lists twelve open roles, some in stages no identified staff member has done.

Get the full visual brief as a PDF.
Founded
2026 · San Francisco
Team
13 identified · 20+ claimed
Raised
$200M seed at $1B · talks at $5B
Builds
AI that does AI research
Hiring
12 open roles

0.1 — The whole story · start months from 12 of 15 people ever identified

Ten people with known start months by June, when the company said twenty. By September the talk was a $5 billion valuation.

4
Known starts by March 2026
The CEO, CTO, a co-founder from xAI and a founding researcher. Talks are reported at $1 billion.
10
Known starts by June
Launch, when the company announced “20 researchers and engineers.” Three more have no published start month.
13
Identified staff, October 2026
After two technical hires left in September. Bloomberg reports “more than 20.”
Mar 2026
CEO, CTO, a co-founder from xAI and a founding researcher start; talks reported at $1B.
May 2026
Three join, including the founder of xAI's Grok web team.
Jun 2026
Launch: $200M seed at $1B, “20 researchers and engineers.”
Aug 2026
A multi-year TPU deal with Google Cloud.
Sep 2026
Two technical hires leave; Bloomberg reports talks at $5B.
Oct 2026
Mirendil-Base, Fern 0.1 and Lor; twelve roles open.

— The 13 identified staff, by where we found them

Everything here is public: LinkedIn, alphaXiv, GitHub, a personal site, posts on X, a founder podcast, an investor post and the funding press. Three people's start months aren't published and one is approximate. The reading is ours.

0.2 — The idea · automate model development

The product is an AI researcher's job description. Mirendil wants software to do it.

“We want to reduce it to 10, two, and eventually one, and eventually you wouldn't need to hire on the AI side.”

Behnam Neyshabur, CEO, AI + a16z podcast, June 24, 2026. Self-report; the captions are unlabelled, so the speaker is identified from context.

Mirendil-Base · Oct 21
A coding model “built by one researcher in 3.5 months, with less than $10M in compute.” Agents wrote over 99% of the code.
Fern 0.1 · Oct 26
The first product model: Mirendil-Base after short post-training, with 10 million tokens of context. Waitlist only, no open weights. Two larger models are in pretraining.
Lor · Oct 28
An agent system “to do the work of an entire AI lab”, in private beta. The thesis, sold as a product.

Product claims are the company's own. No independent benchmark yet; the one researcher is not named.

§ 01 — Part one of four · The founders

They built reasoning at Gemini and automated research at Anthropic. Now they want to automate the researcher.

The four co-founders, and how the goal of “eventually one” researcher shapes the team they hire.

1.1 — The four co-founders

The CEO and CTO left Anthropic in the same month. Both had worked on Gemini first.

01
Behnam Neyshabur
CEO · from March 2026
Sharif, a TTIC PhD, then IAS and NYU. Five and a half years at Google DeepMind co-leading Blueshift: Minerva, the first Gemini reasoning and Gemini's math model. Thirteen months at Anthropic co-leading Discovery, a team aiming at an AI scientist. 8,197 citations, per OpenAlex.
02
Harsh Mehta
CTO · from March 2026
Nine years at Google as a software and then research engineer, Gemini pre-training and reasoning at DeepMind in 2024, then pre-training and AI R&D at Anthropic. a16z: he “built the first version of the Anthropic autoresearch platform, initially as a team of one.” His last job was the company's thesis.
03
Shayan Salehian
Co-founder · from March 2026
Sharif and UC Riverside, six years leading home-timeline ML at Twitter and X, then sixteen months at xAI on Grok 4.1 reasoning post-training, the Grok chat backend, memory and tools. The co-founder who has shipped post-training at xAI.
04
Tara Rezaei
Co-founder · 2026
Sampad, Iran's school for exceptional talent, then MIT, with CSAIL research from her first year and two papers. A 2024 OpenAI internship on multi-agent systems. Multi-agent work, the core of Lor.

The CEO and CTO were at Anthropic from December 2024 to December 2025, on identical dates. Their joint paper on gated state spaces dates to 2022.

1.2 — Ten, two, and eventually one

A frontier lab needs two hundred top researchers, the CEO says. Mirendil is betting it can need one.

— Behnam Neyshabur (from context), AI + a16z, June 2026

“If you need 200 people to have a good frontier AI lab, and 200 of the best people, which is almost impossible to do, we want to reduce it to 10, two, and eventually one.”

Self-report. The founders say candidates ask how they can “compete with big places with … only having … 20 people,” and that they work with “maybe 10 times less people.”

200
Researchers a lab needs, per the CEO.
20
Staff claimed at launch. We identified thirteen.
1
Researcher behind Mirendil-Base, per the company.
12
Roles open in October.

The small team is the proof of concept, and the pitch to candidates.

Talent brief

Want the full visual breakdown?

Download the PDF version of this teardown: the headcount chart month by month, the training stack mapped person by person, how six hidden staff were found, and five lessons for founders building a research team.

§ 02 — Part two of four · Most of the stack in thirteen people

Pretraining, reasoning, post-training, inference, agents. Someone here has done most of them before.

Where each stage of the training stack was learned, the gaps, and the nine people beyond the founders.

2.1 — The stack, mapped · where each person did the work before Mirendil

Most stages of training have been done before. The gaps are on the job board.

01
Pretraining
Mehta
Gemini at DeepMind, then Anthropic.
02
Reasoning and math
Neyshabur · Mehta
Minerva and the first Gemini reasoning (Neyshabur); Gemini reasoning in 2024 (Mehta).
03
Post-training and RL
Salehian · Wu
Grok 4.1 at xAI; RL training at Together AI.
04
Automated research
Neyshabur · Mehta
Discovery and the autoresearch platform, at Anthropic. The company's thesis.
05
Agents and tools
Salehian · Rezaei
Grok's chat backend and tools at xAI; multi-agent systems at OpenAI.
06
Inference and compression
Wu
Quantization and speculative decoding at Together AI; DeepSpeed at Microsoft.
07
Training infrastructure
Mozaffari
A distributed training pipeline at Microsoft.
08
Theory
Nakkiran
Generalization research at Apple.
09
Shipping the product
Gau · Chhadia · Karunamurthy
Grok at xAI; engineering, product and design at Scale AI.
10
The gaps
Open roles
RL environments and RL infrastructure have no one mapped. Both are posted.

From each person's stated roles. Devvrit, Cesista and the operator are not mapped; OpenAI is a two-month internship.

2.2 — The researchers · four members of technical staff

A double-descent theorist, an inference lead, a Google-trained PhD student and an olympiad programmer. None is on the People page.

01
Preetum Nakkiran
Founding MTS · 2026
Berkeley EECS, a Harvard theory PhD, then four years as a senior research scientist at Apple ML Research. Known for Deep Double Descent; 5,347 citations, per alphaXiv. How models generalize as they scale.
02
Xiaoxia Shirley Wu
MTS · June 2026
A UT Austin PhD in machine learning, a senior researcher at Microsoft on DeepSpeed, then two years at Together AI, ending as principal scientist, on quantization, speculative decoding and RL. Serving a 10M-token model cheaply.
03
Devvrit K.
Founding member · March 2026
BITS Pilani with an institute bronze medal, three years as a Google student researcher during the PhD, and a 2025 GenAI internship at Meta. Joined while a fifth-year UT Austin PhD student.
04
Franz Louis Cesista
MTS · start not published
Two IOIs and an ICPC World Finals for the Philippines, mathematics at Ateneo de Manila, then three years at Expedock, from engineer to ML research scientist. The one technical hire from outside the US big-lab pipeline.

2.3 — The builders · three engineers and a twenty-six-year veteran

The builders shipped Grok, ran Scale AI's engineering and trained models at Microsoft. Two of the four are missing from the People page.

01
Arno Gau
MTS · May 2026
Staff engineer at Scale AI on Nucleus and Spellbook, founder of Iudex AI, which xAI acquired, then nineteen months at xAI founding the Grok web team through Grok 3, 4 and 4.1. Built Grok's web product.
02
Vijay Karunamurthy
“Building” · May 2026
Head of search and discovery at YouTube, five years an Apple engineering director after Apple bought his startup, then Head of Engineering, Product and Design at Scale AI. Has scaled an AI company before.
03
Nikash Chhadia
MTS · July 2026
Stanford math and CS with an MS on leave; ML at Tesla, AWS and the Arc Institute; six months on Grok at xAI. Joined two months after leaving xAI.
04
Sahand Mozaffari
Engineer · around August 2026
Sharif and UIUC; a distributed training pipeline for 10x larger models at Microsoft, then production ML at Walgreens and Opendoor. Training infrastructure at scale.

2.4 — Everyone else · the operator and two departures

An operator runs the business side. Two technical hires left in September.

April 2026 · Alvaro Marin
Business and Operations. Eight years co-founding Opportunity Network, from New York to Barcelona; a Stanford MBA; then BCG. The first non-technical hire, a month after the founders.
Apr → Sep
A machine-learning research scientist who came from Apple. Left after six months.
May → Sep
An ML engineer who owned post-training data, RL environments, the RL training stack and the GPU cluster. Three of those four areas are now posted roles.

Two of fifteen people ever identified have left, 13%. Neither departure shows on the People page.

§ 03 — Part three of four · Hiding in plain sight

The People page shows seven staff. Research profiles, a personal site, a GitHub bio and a product post gave us six more.

Why a $1 billion lab's People page shows half its team, and how we found them anyway.

3.1 — Six of thirteen · how the People page misses them

Six of the thirteen are missing from Mirendil's People page. Three use “Stealth”, “Stealth Startup” or a bare “M”.

“Stealth Startup”
Xiaoxia Shirley Wu. Same title, same start date, employer masked.
“Stealth” · “M”
Devvrit K. and Arno Gau. One word, or one letter, where the company should be.
No Mirendil role
Preetum Nakkiran, Sahand Mozaffari and Franz Louis Cesista: a headline of “Researcher”, no new role after leaving Opendoor, a headline-only mention. None reaches the People page.

The People page counts twelve members: four investors, one unrelated Brazilian namesake and seven staff; one more hidden card couldn't be resolved. Twenty is the company's launch figure; Bloomberg reported “more than 20” in September.

3.2 — How we found them · each source, and what it gave

The People page misses them. Their own profiles don't. Researchers keep their research profiles current.

01
LinkedIn People page
7 staff
Twelve members, four of them investors and one a Brazilian namesake.
02
Past-company search
2 departures
People whose Mirendil role had already ended. Invisible on the People page.
03
alphaXiv profiles
3 · Wu · Nakkiran · Cesista
Each researcher's own profile names Mirendil. The best single source for a quiet lab.
04
A personal website
Devvrit K.
“I am a founding member at Mirendil.”
05
A GitHub bio
Arno Gau
“Currently: @mirendil,” with a repository pushed the day we looked.
06
A post promoting Lor
Sahand Mozaffari
An X post about the product led to the engineer's own LinkedIn post: “my next chapter at Mirendil.”
07
Still unread
9 tagged accounts
The CEO's launch post on X tags nine accounts; we could not read who they are. The best lead to the missing seven.

§ 04 — Part four of four · Come and take theirs

They want to automate the AI researcher's job. Until then, they're hiring twelve.

The open roles and who you'd sit beside, the gaps in the record, and what we took from it.

4.1 — The job board · 12 open roles on Ashby, October 29, 2026

Think it's unfair for researchers to automate researchers? Twelve of their jobs are open.

01
Post-training
3 · RL · RL Infra · RL Environments
Beside Salehian, from Grok 4.1 post-training; environments pays $300–400K base. Part of the work the departed engineer owned.
02
Pretraining
1 · Pretraining
Beside Mehta, from Gemini and Anthropic pre-training. The stage Mirendil-Base says one researcher ran.
03
Serving
2 · Inference · Kernels
Beside Wu, from quantization at Together AI. Kernels was posted the day Mirendil-Base launched.
04
Systems
2 · Infrastructure · Distributed Systems
Beside Mozaffari, from distributed training at Microsoft. The GPU cluster also lost its owner.
05
The thesis
2 · Evaluation · Recursive Self-Improvement
Beside Chhadia, whose Mirendil role is listed as RSI, and the founders' Anthropic work. RSI has been open since June 12.
06
The company
2 · Company Building · Executive Operations
Beside Marin and Karunamurthy. Both posted October 21.

4.2 — What the record doesn't show

Five gaps in the evidence, and $5 billion talks not reported closed.

01 · HeadcountAbout seven unnamedTwenty claimed at launch, more than twenty per Bloomberg; thirteen identified. Nine accounts tagged in the CEO's launch post could not be read. Any of them may be hidden the same way.
02 · AttributionWho is the one researcher?Mirendil-Base is credited to one researcher, unnamed in every source we read. We won't guess.
03 · ResultsSelf-reported numbersNo open weights. One reader on X noted Fern's 45% on multi-hop QA at 10M tokens against 100% single-needle retrieval. The cost claim is unaudited.
04 · CompetitionThe field is fillinga16z featured MakerMaker.AI; Apex Intelligence launched an “automated AI research” model two days after Fern. Neither is a prior employer here.
05 · RecordsThin profilesThree start months are unpublished and one is approximate. A missing title and a profile that wouldn't load came from second sources.
The priceUnder $250 million per staffer.If the $5 billion talks close, against more than twenty people. The investors are paying for the thesis as much as the team.

4.3 — What we learned

Five moves worth copying.

01 · SourcingRead past the People pageSix of thirteen were found through alphaXiv, a GitHub bio, a personal site and a post. For a quiet lab, LinkedIn is a floor.
02 · PlanningStaff the stack, not the org chartA few people per stage, each of whom has done it before. Thirteen people cover most of a pipeline; the job board covers the rest.
03 · PlanningHire a pair, not a personThe CEO and CTO came as a unit, after Gemini and Anthropic together. A proven pair de-risks the first year.
04 · SourcingMine one lab deeplyThree technical staff came from xAI, including a team founder. One alum's network opens the next hire.
05 · Offer designSell the small teamCandidates ask how twenty people can compete; fewer people is the thesis. Size can be the pitch, not the objection.
The catch“Eventually one” still needs twelve.The goal is fewer researchers; the board lists twelve roles. The automation has to arrive before the hiring stops.

— The takeaway

They want to automate the AI researcher's job.
Think that's unfair? Twelve of theirs are open.

Brief
Mirendil · Talent Brief
Prepared by
Base to Base · Recruiting
Written for
Researchers, and founders building research teams

Methodology & limitations

Thirteen staff identified, six of them off the People page, and every self-report labelled.

— Sources

  • The LinkedIn company page (12 associated members; four investors and one unrelated namesake excluded) and profile histories for 13 staff (12 from LinkedIn, one from alphaXiv because the LinkedIn profile wouldn't load) and two former staff, collected October 28–29, 2026.
  • For the six staff not on the People page: their own alphaXiv profiles, a personal site, a GitHub bio and one staff post. OpenAlex and alphaXiv citations; cphof.org for competition results.
  • One founder appearance: AI + a16z, with both the CEO and CTO. The a16z seed post.
  • Company posts (launch, Mirendil-Base, Fern); Bloomberg via PYMNTS for the $5B talks; The Information for the March talks; Import AI 478; the Ashby job board; posts on X, September 29 to October 29, 2026.

— Limitations

  • Founder quotes are self-report. Investor claims are neither self-report nor independent, and are labelled.
  • About seven claimed staff are unidentified. Every count here is over the thirteen we could confirm.
  • The stack map is our reading of each person's stated roles, not an org chart the company publishes. Founders are named as in the press and staff as they show themselves on LinkedIn or alphaXiv; former staff are not named.
  • Citation counts move daily; they are dated to collection. Teardowns like this are how our searches begin — this one's on us.