Ilya Sutskever leaves OpenAI after nearly a decade

On May 14, 2024, Ilya Sutskever left OpenAI after nearly a decade and Jakub Pachocki became chief scientist, according to the company’s official announcement. The departure mattered because of the function changing hands—the scientific direction of the company that had popularised ChatGPT—not merely because one of its founders was famous.

In a message posted on X, Sutskever described OpenAI’s trajectory as extraordinary and expressed confidence that the company will build safe and beneficial artificial general intelligence — AI capable of handling a wide range of intellectual tasks — under the leadership of Sam Altman and Greg Brockman. He also said he is working on a personal project that he will share more details about later.

OpenAI said that Jakub Pachocki, previously its head of research, will take over as chief scientist.

One of the architects of today’s OpenAI

Sutskever founded OpenAI in 2015 alongside Altman, Brockman, Elon Musk and other researchers and entrepreneurs. He had previously worked at Google, where he contributed to foundational research in deep learning, the technique that allows machines to find patterns in vast amounts of data.

His name is linked to some of the breakthroughs that laid the groundwork for today’s systems. He coauthored influential papers on neural networks for computer vision and on neural machine translation. At OpenAI, he led a substantial portion of the research that produced the GPT families, the models that power ChatGPT.

The importance of that track record goes beyond the titles. In just a few years, OpenAI has become a company with mass-market products and enormous influence over the commercial direction of generative AI. The departure of its top scientific leader shifts the balance between research, product development and safety inside a company facing growing competitive pressure.

The shadow of the November crisis

Sutskever’s departure comes six months after the most turbulent episode in OpenAI’s history. The board removed Sam Altman on November 17, 2023, a decision Sutskever took part in as a board member. The news triggered an internal revolt: the great majority of employees threatened to resign if Altman did not return.

Days later, Sutskever publicly said he regretted taking part in the decision to remove Altman and pledged to do everything he could to reunite the company. Altman returned as chief executive on November 22, with a revamped board that no longer included Sutskever.

That crisis exposed a tension that remains essential to understanding OpenAI. The organization was founded with a mission focused on developing advanced AI that would benefit all of humanity, but it now also competes in a market of enormous economic value against Google, Anthropic, Meta and other technology companies. The pace of product launches and the need to manage risks do not always move in lockstep.

Sutskever was particularly associated with the latter. In 2023, he co-led the Superalignment team, created to research how to supervise and control AI systems far more capable than today’s models. The goal was to solve, within four years, safety problems that could emerge with future models. His departure leaves open the question of how much weight that line of research will retain within OpenAI.

A scientific handover at a decisive moment

Jakub Pachocki takes over as chief scientist after leading OpenAI’s research. He has worked on the development of GPT-4 and other company systems, making the appointment a bet on technical continuity rather than an external hire.

Sutskever’s departure does not, by itself, signal a change in strategy or reveal the nature of his next project. But it does remove one of OpenAI’s founders — and a voice with unique authority in the debate over how to build increasingly powerful AI systems — from the company’s front line.

For OpenAI, the immediate challenge will be to show that it can maintain its research capabilities while commercializing its products and addressing the safety demands that come with its most advanced models. For the industry, Sutskever’s next initiative is worth watching: few people have had as much influence on the technology that has brought generative AI to the center of the economy and public debate.

How to read a departure without turning it into psychodrama

The first discipline is to separate facts from interpretations. The role a person leaves, the date, the announced successor and the public responsibilities of both are facts. Claims about an internal defeat, a commercial turn or the private reason for leaving are interpretations unless attributable documents or testimony support them. An organisation can lose a central figure without automatically abandoning its scientific programme; it can also preserve the organisation chart while stripping a function of resources.

The official announcement verifies the succession but has a limit: it is the company’s account and says nothing about the new chief scientist’s budget, authority or measurable objectives. To evaluate continuity, follow artefacts rather than adjectives. They include research papers, system cards, published evaluations, hiring and the identity of the person who approves deployment. Changes in those outputs are evidence of an operational shift; a different press-release tone alone is not.

A function is measured through its decisions

At a frontier lab, “chief scientist” may mean setting the agenda, coordinating teams, providing technical review or acting as a public representative. The title alone does not reveal which decisions the holder can stop. A useful record asks who controls compute and recruitment, who defines pre-release tests, who receives adverse findings and which body resolves a disagreement between safety and product.

Research and corporate governance also need to be kept apart. A brilliant researcher does not replace rules for handling conflicts of interest, documenting decisions or supervising executives. Likewise, a reconstituted board does not prove that a scientific method is sound. Combining the two levels produces forceful headlines and weak diagnoses.

The transferable skill is to audit a leadership change through a simple chain: function, authority, resources, succession and outputs. When a prominent person leaves any laboratory, preserve that record and update it with public evidence. It lets you detect genuine change without inventing motives or mistaking a leader’s biography for the institution’s operation.

What to watch after the announcement

An honest follow-up fixes the initial snapshot and does not fill gaps with hindsight. On the departure date, only what Sutskever and OpenAI had disclosed could be stated. Later projects, if announced, belong to later reporting; they cannot be presented as motives already known. This date discipline prevents an old article from seeming to predict the future.

Scientific output provides slower signals than an organisation chart. A published paper may have begun long before a departure and retain authors who have since changed employer; a new recruit may take time to produce results. Date every artefact and distinguish initiation, publication and institutional responsibility. Continuity is not settled by one signature but by observing a sequence.

The minimum follow-up record fits on one page: announcement date, role, known responsibilities, successor, decision bodies and expected artefacts. Add each piece of evidence without rewriting the earlier snapshot. That history preserves uncertainty and shows when a conclusion became reasonable.

A documented sequence is worth more than an instant explanation.