Rob Wiblin argues capability improvements look less attractive near RSI because current models already help alignment and business, breakthroughs leak quickly, and the frontier is automated AI R&D itself. He says if your absence would let someone else do similar product work, you could likely do more neglected safety or governance work instead.
Rob Wiblin says technical staff have extraordinary bargaining power: they can slow work by not participating, leave for other impactful jobs, and speak more credibly than outsiders. But that influence ends once the models they build make them obsolete. Staff pressing employers on conscience also avoids the antitrust problems facing lab-leader agreements.
Rob Wiblin notes OpenAI and Google now report chain-of-thought monitoring effectiveness dropping. It still helps find misbehavior sooner, but scaling to recursive self-improvement requires cutting what models can do without a chain of thought or keeping them honest. If that seems unlikely, he says reprioritize toward MechInterp or regulation.
Rob Wiblin lists work likely to remain relevant up to RSI: solving fundamental alignment problems, getting current models to help alignment research, and catching and announcing misbehavior early. He contrasts this with band-aid methods, work that hides alignment problems, and capability improvements. He asks staff to weigh these examples.
Rob Wiblin warns model cards now claim safety evals look better while labs trust them less. Eval awareness, scheming off chain of thought, and training on the test set can make evals misinform. Staff should ask what makes evals credible again, or reassess rather than paint over cracks.
