← Back to feed

Lilian Weng is leaving Thinking Machines Lab and returning to OpenAI, moving from co-founder responsibility into a narrower research role after months of health strain and startup workload pressure.

The people move matters for two reasons. It puts one of OpenAI’s former Safety Systems leaders near recursive self-improvement work. It also leaves Thinking Machines with only two of its original six co-founders.

The OpenAI Role

OpenAI says Weng will lead a top-level group focused on accelerating internal research into recursive self-improvement. That phrase is narrow enough to be technical and broad enough to carry real organisational risk.

Recursive self-improvement work can mean several concrete things:

• Models improving training, evaluation, and inference systems
• Agents writing better code for model development workflows
• Automated research loops that generate, test, and refine hypotheses
• Internal systems that speed up safety, monitoring, and preparedness research

This is not a consumer product role. It sits close to the machinery that determines how quickly a lab can improve its own models and tooling. OpenAI has already been publishing around agentic scientific computing, harness design, and frontier efficiency. Weng’s return suggests that work is being organised as a first-class research track.

Why Weng Matters

Weng worked at OpenAI from 2018 to 2024, spanning GPT-4 development and leadership of Safety Systems. That combination is rare: deep institutional knowledge of how OpenAI builds frontier models, plus direct experience with the internal controls around deployment risk.

That background is especially relevant to recursive self-improvement. The technical upside is obvious: a lab that can automate more of its own research loop may compound faster. The risk is equally obvious: automated improvement loops can also hide failures, overfit internal benchmarks, accelerate unsafe capabilities, or push evaluation systems beyond human legibility.

Putting a safety leader near that work is not a guarantee of restraint. It is a signal that OpenAI understands the governance problem is inside the research workflow, not only at the product-launch gate.

Thinking Machines Takes Another Hit

Thinking Machines Lab launched with Mira Murati and a founding group heavily drawn from OpenAI. Weng’s departure leaves only two of the original six co-founders at the company.

The timing is awkward. Thinking Machines has been trying to prove that it can build a durable frontier lab outside the dominant OpenAI, Anthropic, Google, and xAI axis. Its first major release, Inkling, established credibility as a serious open-weights system. But model credibility and organisational stability are different things.

Frontier AI startups are now competing on two scarce resources at once: compute and people who can actually run model development at the frontier. Losing co-founders to Meta, OpenAI, or other labs is not just a headcount issue. It removes decision-makers who carry implicit knowledge about training recipes, evaluation standards, product taste, safety judgement, and research culture.

The Talent Market Is Becoming a Control Surface

The Weng move is part of a larger pattern: the most important AI hiring stories are no longer generic compensation fights. They are transfers of capability around specific research bottlenecks.

OpenAI gets a researcher who knows its internal systems and can operate near self-improvement work. Weng gets a scoped role with fewer co-founder obligations. Thinking Machines loses another senior founder while trying to turn Inkling from a strong first model into a durable platform.

The technical question is what Weng’s group actually ships. The governance question is whether recursive self-improvement can be made boring enough to operate inside a company whose incentives push relentlessly toward faster model progress.