Nathan Lambert: AI progress rapid, but not superintelligence

The balance between research and engineering in AI has shifted before. Before deep learning took off, AI was a research endeavor, but today's best researchers are judged by their ability to implement and scale ideas in complex infrastructure. Lambert argues the field is entering a new transition where engineering bottlenecks ease, making good ideas more valuable than good execution. One piece of evidence he points to is the RL environment data market, where new companies have crossed $100M or $1B in revenue despite outputs that many researchers describe as low-quality. Meta's Muse agent is an early indicator of agent-focused experiences, and Lambert expects more Muse-like products for different audiences. Even if models become superhuman at crawling literature and making connections across sparse networks, Lambert suggests there is a fine line between heralding a new era of scientific discovery — such as cures for most cancers — and simply accelerating the arc science was already on.