Lossy self-improvement, not RSI: Anthropic sees no dramatic acceleration

Lambert is writing against a backdrop he describes as a cultural shift inside the frontier labs. He notes that a year ago, the general staff at OpenAI and Anthropic were already anxious about AI risks and the pace of progress, and that this anxiety grew as agents found stronger product-market fit at the start of 2026. Combined with the frenetic, competitive culture in the San Francisco AI scene, that precondition can amplify any concern — and Lambert says while this raises general awareness of AI (because fear sells), exaggerating risk timelines or severity has negative second-order effects. He points to the loud AI safety debates of 2023 and 2024, and says the primary risks forecast then did not arrive on those timelines.

He quotes Richard Ngo's summary that much of the safety community now expects an intelligence explosion within a few years, and that while this may be directionally right, it is factually wrong — no superintelligence within the next 8 years, but progress fast enough to feel like the short-timeline camp was right. Lambert's alternative, lossy self-improvement, rests on three claims: automatable research is too narrow to drive massive net acceleration given scaling laws' exponential costs; returns from adding more parallel agents diminish; and resource bottlenecks and politics remain major factors in building strong LLMs. He also draws on podcasts with Noam Brown and with John Schulman, Beren Millidge and Charlie O'Neill, summarizing their timeline guesses — for example, around one to three years for a drop-in remote worker, about two years (Schulman) or five to ten (O'Neill) for a 10× productivity uplift for AI researchers, and three to ten years for AI surpassing top human experts across computer-based work.

What all this hinges on, in his framing, is that the labs keep running on clear, measurable problems — software engineering, log monitoring, planned experiments — where agent swarms genuinely help, and where cheaper inference and better efficiency follow. Harder to automate are things like post-training recipes: Schulman's point that someone has to decide how the model should behave in each area, and that it is easy to screw up post-training in ways that do not show up on benchmarks. Lambert holds high uncertainty about whether unpublished breakthroughs could change his picture, and says the hardest exponential to move is peak intelligence. Until more evidence appears, he treats it as the less likely scenario that RSI delivers a near-term intelligence explosion rather than a broad, economically transformative diffusion of cheaper capability.