Lilian Weng Exits Thinking Machines for OpenAI's Self-Improvement Team

Lilian Weng’s whiplash-inducing career move this week perfectly encapsulates the paradoxical reality of the 2026 AI industry: the only thing more exhausting than building a frontier AI startup is trying to stay away from the AGI race.

On Monday, Weng announced she was stepping down as co-founder of Thinking Machines Lab (affectionately known as "Thinky") due to severe health issues and burnout. By Wednesday, OpenAI confirmed she was coming back to her old stomping grounds. Her new role? Leading a top-level team dedicated to one of the most high-stakes, sci-fi-sounding initiatives in the field: recursive self-improvement.

The rapid transition from a burned-out startup founder to the leader of OpenAI's self-improving AI division highlights a growing trend in the machine learning ecosystem. Top-tier researchers are realizing that the operational friction of building a startup from scratch often distracts from the actual science—and the gravity well of established giants like OpenAI is stronger than ever.

The Breaking Point at "Thinky"

Thinking Machines Lab was supposed to be the agile, open-weight counter-punch to the increasingly closed-off corporate AI labs. Founded in 2025 by former OpenAI CTO Mira Murati, the startup managed to attract a staggering roster of talent, including Weng, who had previously served as OpenAI's VP of AI Safety Research.

But the reality of the startup grind quickly took its toll. In an internal Slack message that Weng later shared on X (formerly Twitter), she was remarkably candid about the physical and mental cost of the frontier AI race.

"I don’t feel I’m able to continue at the pace a startup requires," Weng wrote.
"After thinking about it for several months, I ultimately have to admit that the amount of consistent stress and workload have pushed me beyond what my health can sustain physically."

Weng noted that she had been ill more frequently in the past seven months than at any other point in her life. The breaking point seemingly coincided with Thinking Machines' sprint to release Inkling, the company's first open-weight model. Weng admitted to feeling immense "guilt" when trying to take sick leave or rest while her team was working around the clock to ship the model.

Thinking Machines was built on a thesis that the future of AI shouldn't be entirely dictated by closed-source, API-gated monoliths. Inkling was designed as an open-weight alternative that developers could run and fine-tune locally. But the open-weight philosophy comes with a unique set of burdens. When a company releases an open-weight model, there is no "undo" button. The safety evaluations, red-teaming, and alignment checks must be exhaustive before the weights are published, because once they are on the internet, they are there forever. This reality likely contributed to the brutal sprint culture that Weng described.

Murati responded publicly with grace, writing on X: "We’ll miss you, it’s been wonderful building Thinky together. I’m glad that you’re putting your health first. Thank you for everything."

The Irony of a "Predictable" Role

What makes Weng's departure fascinating is her immediate pivot back to OpenAI. In her departure note, she mentioned that she still loved the field and reading papers about frontier AI progress, but realized that a "scoped role in a more predictable place, and not as a cofounder" would better suit her current physical needs.

To the outside observer, joining OpenAI to build self-improving superintelligence doesn't exactly sound like a low-stress, predictable gig. However, it speaks volumes about the difference between research stress and founder stress.

At a startup, a co-founder is responsible for everything: securing compute clusters, managing investor expectations, hiring, payroll, and the existential dread of running out of runway. At OpenAI, Weng can plug into an established, practically infinite compute infrastructure and focus entirely on the science. She doesn't have to worry about whether the GPUs will arrive on time; she just has to figure out how to make the models smarter.

Before her time at Thinking Machines, Weng built a formidable reputation across the tech industry. She joined Dropbox in 2014, moved to the fintech giant Affirm in 2016, and eventually found her calling at OpenAI in 2018. Over nearly seven years, she worked on robotics, founded the Applied AI Research team, and eventually took the reins of the company's Safety Systems. Her work was foundational to the pretraining data curation, safety evaluations, and eventual deployment of GPT-4—arguably the model that kicked off the current generative AI arms race.

The Holy Grail: Recursive Self-Improvement

Weng's new mandate at OpenAI is arguably the most important technical challenge of the late 2020s. She will lead a team supporting cross-research work on recursive self-improvement (RSI).

For the uninitiated, RSI is the theoretical process by which an AI system becomes capable of iterating on its own architecture, training data, and code to create a smarter version of itself. That smarter version then repeats the process, leading to an intelligence explosion.

Why is Weng the right person for this?

Safety Pedigree: During her first stint at OpenAI, Weng was instrumental in defining what model safety actually looks like in practice.
The Alignment Bottleneck: Recursive self-improvement is inherently dangerous. If a model's alignment degrades even slightly during a self-directed update, subsequent generations could rapidly diverge from human values. Weng's deep background in AI safety makes her uniquely qualified to build the guardrails for models that write their own updates.

OpenAI's decision to dedicate a top-level team to this specific vector indicates that the company believes we are on the cusp of models that can genuinely contribute to AI research. We are moving past AI as a coding assistant and into the era of AI as an autonomous AI researcher.

The Brain Drain at Thinking Machines

Weng's exit also shines a harsh spotlight on the retention crisis at Thinking Machines Lab. When the company launched, it boasted a formidable founding team of six industry heavyweights. Today, nearly a third of the original founding members have departed.

Lilian Weng: Returned to OpenAI.
Barret Zoph: Left for OpenAI (and subsequently left OpenAI again in June 2026).
Luke Metz: Left for OpenAI.
Andrew Tulloch: Left for Meta.

Currently, only CEO Mira Murati and Chief Scientist John Schulman remain from the original founding six.

This brain drain underscores a brutal reality for AI startups in 2026. Even with massive funding rounds and visionary leadership, competing with the "Big Three" (OpenAI, Anthropic, and Google) is an uphill battle. The sheer scale of resources required to train frontier models creates a grueling environment for engineers and researchers. When the going gets tough, the allure of returning to a mega-lab with unlimited compute and a massive support staff becomes irresistible.

What This Means for the Industry

Lilian Weng's return to OpenAI is more than just a high-profile game of musical chairs. It is a signal that the AI industry is entering a new phase of consolidation, driven not just by capital, but by human endurance.

The Limits of the Open-Weight Sprint: The pressure to release open-weight models is immense. The safety evaluations and red-teaming must be flawless before the weights are dropped, creating a pressure-cooker environment that is burning out top talent.
Safety is Now a Feature of RSI: By putting a safety veteran in charge of recursive self-improvement, OpenAI is signaling to regulators and the public that they are taking the intelligence explosion seriously. It’s a strategic move that blends aggressive capability scaling with defensive alignment research.
The Founder Penalty: We will likely see more "boomerang" employees in the coming months. Researchers who left big labs in 2024 and 2025 to start their own companies are realizing that being a CEO or CTO is a fundamentally different job than being a scientist.

As Weng noted in her farewell message to Thinky: "The future worth building is human." Ironically, she will now be spending her days building the very systems that might one day automate the researchers themselves. But for now, the AI revolution still relies on human minds—and those minds have limits.