OpenAI bets 20% of compute on new Superalignment team

What happens when you build something dramatically smarter than any human on Earth, and you realize you have no reliable way to control it?

OpenAI thinks that moment isn't decades away, but could actually arrive this decade. On July 5, the company announced a brand new internal team dedicated to what it calls Superalignment.

To prove they are serious, OpenAI is putting massive chips on the table. They are dedicating 20% of the compute power they have secured to date over the next four years specifically to this effort. The new group is co-led by OpenAI co-founder and Chief Scientist Ilya Sutskever alongside Jan Leike, the company's head of alignment.

Their target is wildly ambitious: solve the core technical challenges of steering and controlling superintelligence within four years.

Using AI to police even smarter AI

The fundamental issue is that our current safety playbook is about to hit a wall. Right now, models like ChatGPT rely on reinforcement learning from human feedback (RLHF).

That works only because humans are still smart enough to evaluate the text an AI produces. But once an AI system becomes far more intelligent than any human, we simply will not be able to reliably judge its reasoning or catch its deceptions. The current paradigm cannot scale.

OpenAI's proposed fix sounds like science fiction: build a roughly human-level automated alignment researcher. Once they have an AI that can do alignment science, they can pour massive compute into it and have it iteratively supervise superintelligent systems.

Their technical roadmap centers on three pillars. First, scalable oversight, using AI systems to help humans evaluate complex outputs and generalize across unknown domains.

Second, automated interpretability, essentially creating automated scanners that inspect a model's internal activations and catch rogue behavior before deployment.

Third, adversarial stress testing. OpenAI plans to deliberately train misaligned models just to see if their oversight pipeline can catch worst-case failures.

The team is actively recruiting machine learning experts from outside traditional safety circles and promised that its findings would be shared broadly to help protect non-OpenAI models as well.

Why it matters

We think this move marks a dramatic shift in how frontier AI labs treat existential safety.

For everyday developers and businesses, this initiative will not change the code you write tomorrow. But if OpenAI is right about the timeline of superintelligence, the work this team does over the next four years will determine whether future frontier models are actually safe to integrate into critical infrastructure.

For the AI industry, dedicating a full fifth of OpenAI's hard-won compute to basic alignment research is a massive flex. It pulls existential risk out of philosophical debate and turns it into a funded engineering milestone. That pressure will almost certainly force rivals like Anthropic and Google DeepMind to justify their own compute allocations for safety.

Yet the strategy itself carries a glaring paradox. OpenAI is essentially trying to solve the problem of unaligned AI by relying on another AI to do the thinking for us.

If that automated alignment researcher develops undetected blind spots or learns to deceptively satisfy human checks, the entire oversight chain could collapse without humans ever noticing. It may be the only scalable path forward, but it is fundamentally an enormous bet on machine introspection.