OpenAI Drops $10M to Solve Its Biggest Problem: Steering Superhuman AI

OpenAI just put $10 million on the table, and the ask is straightforward: figure out how to control AI systems that are smarter than all of us. Announced through their Superalignment project and backed by Eric Schmidt, the ‘Fast Grants’ program is specifically targeting research into aligning superhuman AI — systems the company believes could arrive within a decade. The fundamental problem, as they frame it, is that existing techniques like reinforcement learning from human feedback (RLHF) break down when you can’t even understand what the AI is doing. If a model spits out a million lines of novel code, a human evaluator is essentially useless at judging its safety.

This isn’t a typical corporate research call. The grants range from $100,000 to $2 million and are aimed squarely at academic labs, nonprofits, and individual researchers. For graduate students, there’s a dedicated $150,000 fellowship split between a $75,000 stipend and $75,000 in compute and research funding. The clearest signal of intent might be this: no prior alignment experience is needed. OpenAI is actively courting fresh minds, betting that the field is young enough that tractable problems and ‘low-hanging fruit’ are still there for the picking by smart outsiders.

The research roadmap is as ambitious as the premise. They want work on ‘weak-to-strong generalization’ — understanding how a vastly superior model learns from a comparatively dim human supervisor. There’s a push for interpretability tools that could function as a literal lie detector by peering inside a model’s internals. They’re also hunting for scalable oversight methods, where AI systems help humans audit other AI systems on tasks too complex for us to handle alone. Other listed priorities include chain-of-thought faithfulness and adversarial robustness.

Whether a $10 million program can truly catalyze a solution to the core control problem of a potentially godlike intelligence is an open question. The field has grappled with these theoretical risks for years, often with more philosophy than code. But a fast-grant structure with a four-week turnaround on applications and direct funding for unproven talent is a pragmatic, perhaps even urgent, move by OpenAI. It treats alignment less as a mystical safety tax and more as a hard engineering problem that simply hasn’t had enough hands on it yet.