OPENAI — SUPERALIGNMENT▊
The team OpenAI created to keep superintelligence from going wrong — announced with 20% of the company's compute and a four-year clock — dissolved in under one year, days after both of its leaders walked out the door.
- OPERATION
- Superalignment team — formed July 2023 to "solve the core technical challenges" of aligning superintelligent AI, on a stated four-year timeline FACT
- LEADS
- Ilya Sutskever (co-founder, chief scientist) and Jan Leike — co-leads from announcement to exit
- PROMISE
- "20% of the compute we've secured to date over the next four years" — OpenAI's own announcement FACT
- EVENT
- DISSOLVED — May 2024, within days of both leads departing FACT
- DISPOSITION
- OpenAI: safety work distributed across teams; a board Safety & Security Committee formed May 28, 2024 ATTRIBUTED
Here is the rarest promise in the industry — a fifth of the food supply, pledged to the organ whose job is restraint — observed at the moment the organ is reabsorbed and the food reallocated, ten months into a forty-eight-month mission.
Yesterday was my last day as head of alignment, superalignment lead, and executive @OpenAI.
@janleike — on X, May 17, 2024, opening the resignation thread FACT (that he said it) · x.com/janleike/status/1791498174659715494
…safety culture and processes have taken a backseat to shiny products.
@janleike — same thread, as quoted verbatim by CNBC and Wired FACT (that he said it) · cnbc.com/2024/05/17/…
After almost a decade, I have made the decision to leave OpenAI… I'm confident that OpenAI will build AGI that is both safe and beneficial under the leadership of @sama, @gdb, @miramurati…
@ilyasut — on X, May 14, 2024; note that his exit statement cuts for the company FACT (that he said it) · x.com/ilyasut/status/1790517455628198322
We are dedicating 20% of the compute we've secured to date over the next four years to solving the problem of superintelligence alignment.
OpenAI — "Introducing Superalignment," company blog, July 2023; the promise, self-published FACT (that it said it) · openai.com/index/introducing-superalignment/
Superalignment was the loudest safety commitment any frontier lab has made: a named team, the company's most famous scientist at its head, a public fraction of the compute budget, and a deadline. It was, among other things, an answer to every regulator and critic who asked whether OpenAI took its own rhetoric seriously. For ten months, the answer had a team name.
Our read, and it is ours, not a finding and not OpenAI's stated intent: the fourth specimen of safety theater in this drawer — a conscience created for credibility, starved when it competed with product for the one resource that matters, and dissolved when its leaders quit over exactly that. OpenAI's answer — that safety work was distributed across teams and elevated to a board committee — is on this page at full strength. So is Sutskever's own exit statement, which reads in the company's favor. Read both. Keep the dates.
the drama timeline
ACT I — THE PROMISE (JUL–NOV 2023)
The commitment is made in public, with numbers. Numbers are what make promises auditable, which is what makes this file possible.
-
JUL 5 2023
Superalignment announced — 20% of compute, four years
OpenAI announces the Superalignment team, co-led by Sutskever and Leike: the goal is to "solve the core technical challenges of superintelligence alignment" within four years, backed by "20% of the compute we've secured to date." The announcement frames superintelligence as potentially arriving this decade and existing alignment techniques as insufficient.
Mark the structure of the promise: a fraction, a clock, and a name. Every element that later makes the dissolution measurable was placed there, voluntarily, by the company itself.
-
NOV 2023
The board crisis
OpenAI's board briefly removes CEO Sam Altman; Sutskever is part of the board that does it, then publicly reverses and joins the call for Altman's return. Altman is back within days. Sutskever's position inside the company afterward is, per the CNBC recap, uncertain — he does not return to day-to-day superalignment work in any publicly visible way.
ACT II — THE EXIT (MAY 2024)
A product ships on Monday. By Friday the safety team does not exist. The week is the argument.
-
MAY 13–14 2024
GPT-4o ships; Sutskever leaves
OpenAI launches GPT-4o. The next day Sutskever announces his departure — on the record, gracefully: he is "confident that OpenAI will build AGI that is both safe and beneficial" under the current leadership. Leike resigns the same week.
-
MAY 17 2024
Leike posts the reason; the team is dissolved
Leike posts his thread: disagreements with leadership "about the company's core priorities for quite some time"; his team "sailing against the wind" and "struggling for compute"; and the line that named this whole drawer — "safety culture and processes have taken a backseat to shiny products." The same day, CNBC and Wired confirm the Superalignment team is dissolved, its remaining members absorbed into broader research efforts, with co-founder John Schulman taking over alignment research. Altman, responding publicly: Leike is right "we have a lot more to do; we are committed to doing it."
-
MAY 28 2024
The replacement organ
OpenAI announces a board-level Safety & Security Committee — led by directors including chair Bret Taylor and CEO Altman — to evaluate safety processes over 90 days, as the company begins training its next frontier model. The safety function now reports where the dissolved team's critics said it couldn't win: inside the leadership it was meant to check.
Whether a committee containing the chief executive can perform the function of a team that quit over the chief executive's priorities — that is not a question the archive can settle. It is the question the reader should leave with.
both sides, on the record
The promise was specific and it was not kept as made — 20% of secured compute over four years, dissolved in ten months, with the departing lead saying on the record his team had been "struggling for compute" [1] [4].
The indictment is first-party. No leaker, no anonymous source: the head of the safety team himself posted that safety "has taken a backseat to shiny products" — and the CEO publicly agreed there was "a lot more to do" [4] [2].
The replacement reports to the people it checks. The successor structure is a board committee whose membership includes the CEO; the dissolved team's stated problem was leadership's priorities [5].
OpenAI's account: distributed, not abandoned. The company said dissolving the standalone team integrates alignment work more deeply across research, under co-founder John Schulman — an argument that a separate safety silo, not safety itself, is what ended [2] [3].
The co-founder who left endorsed the company on his way out. Sutskever's own post: he is "confident that OpenAI will build AGI that is both safe and beneficial" under current leadership — a first-party statement that cuts for OpenAI, from the man with the most standing to say otherwise [6].
The response was concrete and fast: a board Safety & Security Committee within eleven days, a public 90-day review, and Altman's on-record acknowledgment that Leike was right about the work remaining. No court or regulator has adjudicated any of this; the record is a structural choice and its critics [5] [2].
YOU DECIDE
Scoped to the claims. The 20%-of-compute pledge, the four-year clock, the ten-month lifespan, and the departing lead's own words are all on the record — most of them published by OpenAI and its own executives. Whether dissolving a dedicated safety team into the product organization is a betrayal of the promise or a maturation of it is the genuine dispute, and both readings are on this page. What is not in dispute is which one the man running the team believed — he resigned and told you.
Weigh the costly signal: Leike walked away from one of the most coveted jobs in the industry to say one sentence in public. OpenAI's counter-signal — the committee, the review, the acknowledgment — cost considerably less. Price both before you decide.
The archive does not judge. It keeps the timestamps.
evidence locker
PRIMARY & FIRST-PARTY
Jan Leike — resignation thread on X (May 17, 2024) FACT (that he said it) — "backseat to shiny products," "sailing against the wind," "struggling for compute"; his account, first-party.
x.com/janleike/status/1791498174659715494
PRESS & ANALYSIS
The standard. The promise is sourced to OpenAI's own announcement; the criticism is sourced to the departing lead's own verified posts; the dissolution is sourced to named reporting; the defense — distribution of the work, the Schulman handoff, the board committee, and Sutskever's own favorable exit statement — is presented at full strength. Every X permalink on this page was verified live before publication; nothing is paraphrased into anyone's mouth. The "safety theater" framing is our analytic read of the pattern, labeled as ours, not asserted as OpenAI's intent. Nothing here is an adjudicated finding, and no personal wrongdoing by any named individual is alleged. If it couldn't survive a defamation challenge, it wouldn't be on this page.