Lambert: expect 'lossy self-improvement,' not fast takeoff

Nathan Lambert has a name for the version of AI progress he thinks is actually coming, and it is not recursive self-improvement. Writing at Interconnects, Lambert calls it *lossy self-improvement* (LSI): "the models become core to the development loop but friction breaks down all the core assumptions of RSI."

The essay walks through three reasons the loop leaks. First, the research work that actually gets automated is narrow: things like "lowering the test loss of a model," the kind of target Karpathy's recently-launched autoresearch is built for. That is real, but as Lambert puts it, "the leap required to navigate many metrics at once is a very different skill set."

Second, parallelizing agents runs into Amdahl's law and, more mundanely, into humans. "How many people do you think could come up with 300-400 tasks for AI agents every day?" Lambert asks. Adding more agents sampling from similar solution distributions is not the same as adding more researchers.

Third, compute is political. "Billions of dollars of compute resources for research are unlikely to be totally isolated for end-to-end experimentation," Lambert writes. Allocation still routes through people.

He leans on Paul Allen's "complexity brake," the idea that "the more progress science makes towards understanding intelligence, the more difficult it becomes to make additional progress," and warns that "the bottom of every sigmoid feels like an exponential." He is not calling AGI off. He still expects drop-in remote worker capability within years. He is saying the shape of the curve is linear, not explosive: "We will be in this era of lossy self-improvement (LSI) for a few years, but it is not enough for a fast takeoff."

Two researchers we track shared the piece after it went up.