Miju Labs

← All buyers · Thinking Machines Lab

J

John Schulman

Co-founder & Chief Scientist, Thinking Machines Lab · Thinking Machines Lab

high confidenceTier A · Talk to firstSecond-wave labstext · multimodal

The inventor of PPO and one of the architects of RLHF and ChatGPT, now the last founding scientist running research at Thinking Machines Lab after a year of co-founder departures. He does not buy datasets, but he sets the lab's post-training direction — and in September 2026 he publicly framed the path to self-improving models as human feedback that lets models absorb researcher taste, which is the most quotable endorsement of taste data by anyone at this level.

Background

Caltech physics (2010), then a PhD at UC Berkeley with Pieter Abbeel, where he developed TRPO and then PPO — still the workhorse algorithm of RLHF. Co-founded OpenAI in December 2015, led the reinforcement learning team and was described as the architect of ChatGPT, staying nine years. He left for Anthropic in August 2024 saying he wanted to deepen his focus on alignment and do more hands-on technical work, then left again in February 2025 to become chief scientist at Mira Murati's Thinking Machines Lab.

Thinking Machines has since lost several co-founders — Zoph, Metz, Schoenholz, Tulloch, Chaplot, and Lilian Weng in July 2026 — leaving Schulman as the senior research anchor alongside CEO Murati and CTO Soumith Chintala.

What they run now

Research direction at Thinking Machines: the Tinker fine-tuning platform and Inkling, the lab's natively multimodal open-weight model (text, image, audio, video) released July 2026 and post-trained across math, agentic code, audio, image, chat and safety with more than 30 million RL rollouts. The lab's public thesis is human-in-the-loop customisation rather than one-size-fits-all models. Around him the lab has built the buying machinery: a dedicated Data Operations team (posted August 2026) that owns data projects from sourcing and vendor selection through quality review and delivery and explicitly builds partnerships with data vendors including international ones, plus open roles in post-training data, post-training evals for personalised preferences, and vision and audio expertise. Scale AI and Handshake AI are named as external evaluators in the Inkling safety post.

What they buy

Not personally. He sets the research agenda that generates demand; the lab's Data Operations team does sourcing and vendor selection, and Scale AI and Handshake AI already have a foot in the door as evaluators. Inkling's multimodal post-training (image, audio, video as well as text) implies demand for expert preference data across modalities, and the open 'post-training evals for personalised preferences' role points at subjective-quality evaluation specifically.

Career

  1. 2025 – presentCo-founder & Chief Scientist, Thinking Machines Lab
  2. 2024 – 2025Researcher (alignment), Anthropic — Joined Aug 2024, departed Feb 2025.
  3. 2015 – 2024Co-founder; led reinforcement learning, OpenAI — PPO, RLHF, ChatGPT.
  4. —UC Berkeley, Ph.D., EECS (advisor Pieter Abbeel)
  5. 2010California Institute of Technology, B.S., Physics

On the record

Taste as the bottleneck · 2026-09

Frames future self-improving models as trained with human feedback that absorbs researcher taste, plus research practice environments. source

Judgement, not capability · 2026-09

Argues choosing which experiment to run — the outer loop — stays human; judgement reappears as the bottleneck. source

Why post-training teams stay large · 2026-09

Says post-training teams remain large because someone must decide model behaviour across every domain. source

Interviews & talks

Why it matters here

Not a target, but the best quotable authority in Julian's whole pitch. 'Human feedback to absorb researcher taste' from the inventor of PPO is the line that makes taste data sound like frontier-lab infrastructure rather than a niche service, and it is fresh (September 2026). The real Thinking Machines entry point is the Data Operations team, whose posted remit is literally finding the best vendors and experts.

How to reach

Do not contact him. Speculation: use him as narrative, and approach Data Operations directly through the job-posting channel — that team is being built to meet vendors, is explicitly open to international suppliers, and has a lab-wide mandate. If Julian writes anything public about taste data, quoting Schulman's September 2026 framing and then showing how a designer panel operationalises it is the strongest available framing.

What we could not establish
  • Whether the Dwarkesh debate quotes are verbatim — I read them via a secondary write-up, not the primary recording.
  • Who leads Thinking Machines' Data Operations team; no name is public.
  • Whether Inkling's image and video post-training used external preference data or only internal and bootstrapped sources.

Why they are on the map

RLHF/PPO pioneer; remaining co-founder running research incl. post-training; authored 'LoRA Without Regret' (Sep 2025) source — evidence dated 2026-07

Decides whether the problem is funded.

How Thinking Machines Lab buys data

Dedicated 'Data Operations' team (job posted 25 Aug 2026): owns data projects 'from sourcing and vendor selection through quality review and delivery', finds 'the best vendors, experts, and data sources', builds partnerships with data vendors incl. international ones. 'Product Manager - Post Training' (Sep 2026) says Inkling was post-trained across math, agentic code, audio, image, chat and safety with >30M RL rollouts; also open: 'Research, Post-Training Evals' (evals for 'personalized preferences' with 'the data flywheel'), 'Research, Vision Expertise', 'Research, Audio Expertise', 'Research, Post-Training Data'. Inkling safety post (31 Jul 2026) names Scale AI and Handshake AI as external evaluators. Early Inkling post-training data was bootstrapped from Kimi K2.5 outputs (TechCrunch, Jul 2026).

Flat research org under CEO Murati, CTO Soumith Chintala (since Jan 2026) and Chief Scientist John Schulman; 'multimodal-first', one team across modalities. Data Operations sits in Product Management and serves researchers as an internal vendor-management function; post-training has its own PM. Heavy founder churn: Zoph, Metz, Schoenholz (OpenAI, Jan 2026), Tulloch, Joshua Gross (Meta), Chaplot (xAI, Mar 2026), Lilian Weng (left Jul 2026).

Known vendors: scale-ai, handshake-ai, FAR.AI (safety eval), Apollo Research (safety eval)

Sources

  1. Wikipedia — John Schulman (career, PPO, OpenAI/Anthropic/TML)
  2. StartupHub — Schulman on recursive self-improvement, Dwarkesh debate (2026-09-12)
  3. Thinking Machines — Data Operations job (2026-08-25): vendor selection and expert sourcing
  4. Thinking Machines — A safe path to open weights (Scale AI and Handshake AI as external evaluators)