Introduction
If the alignment problem (Post 16) is the diagnosis, "superalignment" was the most ambitious treatment ever attempted: OpenAI's 2023 project to solve the control of Artificial Superintelligence (ASI), backed by a striking commitment of 20% of the company's compute. Two years later the team is gone, its leaders have departed, and the effort has scattered across the labs in different forms. This post reports what superalignment was, why it ended, what the labs actually do about ASI safety today, and the gap between the research and the race.
Quick Takeaways
- OpenAI announced its Superalignment team in July 2023, led by Ilya Sutskever and Jan Leike, with 20% of compute and a four-year goal: solve alignment for superintelligence.
- The team dissolved in May 2024 after both leaders departed; OpenAI said the work would continue in other teams.
- Today's safety work is spread across labs: Anthropic's alignment science programme, DeepMind's safety and responsibility teams, Meta's research — each with different emphases.
- The core technical bets — scalable oversight, interpretability, evaluations — are real research areas with real progress and real limits.
- The structural fact: safety research runs inside the same companies racing to build the systems it is meant to constrain.
What Superalignment Was
The July 2023 announcement was unusually candid. OpenAI wrote that superintelligence could arrive within the decade, that "we need to solve the alignment problem" for systems smarter than humans, and that current techniques — like learning from human feedback — would not scale, because humans cannot reliably evaluate work they cannot do. The proposed answer was to automate alignment research itself: build AI systems that help align their successors, bootstrapping from today's techniques to ones that work on superintelligence. Hence the name, the four-year timeline, and the 20% compute pledge — a scale of commitment the field had never seen.
The intellectual core was "scalable oversight": how do you supervise a system that outperforms you? Proposed approaches included having weaker models supervise stronger ones, using AI debate (systems arguing against each other with a human judge), and interpretability research aimed at reading a model's internal goals rather than trusting its outputs. None of these were claimed to be sufficient. The bet was that four years of focused effort would find out.
Why It Ended
In May 2024, within days of each other, both leaders left: Ilya Sutskever announced his departure on May 14, Jan Leike on May 15, with Leike writing publicly that he had been "disagreeing with OpenAI leadership about the company's core priorities for quite some time." OpenAI dissolved the team structure and said superalignment work would be absorbed elsewhere. Sutskever went on to found Safe Superintelligence (SSI) — the subject of Post 7 — explicitly organised around building superintelligence safely rather than racing first and aligning later.
What to make of this depends on who you ask, and this site reports rather than adjudicates. The sympathetic reading: the research directions survived, dispersed into other teams, and the departure drama was about governance, not science. The sceptical reading: the most resourced alignment effort in history lasted less than a year before the racing imperative swallowed it, which tells you how safety fares inside a race dynamic. Both readings agree on the facts above; they disagree on what the facts mean.
What the Labs Do Now
Anthropic runs the most academically visible alignment programme: interpretability research (attempting to reverse-engineer what models internally represent), work on scalable oversight, and published research on model behaviours like sycophancy and goal-persistence. Anthropic's public framing — CEO Dario Amodei's — treats alignment as the central technical challenge, and the company publishes more of its safety research than its competitors.
Google DeepMind maintains large safety, responsibility and alignment teams, publishes on evaluations and dangerous-capability testing, and has institutional history here: DeepMind published some of the earliest technical AI safety agendas in the mid-2010s. Its approach skews toward evaluations and governance frameworks — testing what systems can do and building processes around the results.
Meta's Superintelligence Labs (see Post 5) is the newest entrant, formed around advanced capability research; its safety posture is less publicly documented than Anthropic's or DeepMind's, which itself is a data point readers can weigh.
OpenAI post-superalignment continues safety research across restructured teams, publishes on topics including alignment faking and reward hacking, and — per the voluntary White House accord of September 2026 (Post 15) — has committed alongside five other labs to internal oversight teams, outside evaluators and board committees. Whether those commitments have teeth is, as TechCrunch's commentators noted, an open question: the accord is voluntary.
The Honest State of the Research
Three technical bets carry most of the current effort, and each deserves an honest status:
Scalable oversight — real progress (AI-assisted evaluation is genuinely better than it was), real limit (Post 16's verifier problem is unsolved; a supervisor that can't detect deception provides confidence without safety).
Interpretability — real progress (researchers can now identify some internal features of large models), real limit (understanding fragments of a model's computation is not the same as verifying its goals, and the techniques lag far behind capability growth).
Evaluations — real progress (dangerous-capability testing is now standard practice at frontier labs), real limit (evaluations measure what we think to test for; a system that behaves differently under evaluation than in deployment — "alignment faking," which labs themselves have now documented in experiments — defeats the test).
The pattern: every approach is genuinely advancing and genuinely insufficient, and capability is advancing faster than any of them. That is not an accusation; it is the field's own assessment, including from the researchers doing the work.
The Bottom Line
Superalignment was the field's biggest bet on solving ASI control — and it lasted less than a year as a team. What remains is serious, under-resourced-relative-to-capabilities safety research spread across competing labs, making real progress on problems that are genuinely hard. The structural tension has not changed since 2023: the same organisations racing to build superintelligence are the ones tasked with constraining it, on timelines set by the race. Watch what the labs publish, watch what they fund, and watch whether the September 2026 accord's oversight commitments acquire enforcement. The research is real. Whether it is enough, on time, is the open question this whole series orbits.




0 Response to "Superalignment: What Labs Actually Do About It Today"
Post a Comment