I expect rapid progress but not towards general superintelligence

AI could make model development much cheaper and faster without resolving the harder question of general superintelligence. Nathan Lambert's essay grounds that forecast in engineering tasks with clear feedback: training speed and the computing cost of producing an answer can be measured, compared and optimized.

Lambert expects agents to automate more of that optimization, making experiments easier to run and shifting the research bottleneck toward good ideas. This is a forecast about improving the machinery of AI research. Its logic allows substantial economic and engineering progress even if those gains do not produce broadly superhuman scientific judgment. Argument and examples.