In the latest edition of Import AI, Jack Clark surveys the field and finds no single dominant story. Instead, he points to three areas where progress is real but unfinished: oversight, protein-folding scaling, and the economics of existential risk. These are not directly related, but together they sketch a landscape where AI's technical capabilities are outpacing our ability to govern and price their consequences.
On oversight, Clark notes that evaluating and controlling powerful models is proving difficult in practice. Even as systems become more capable, the tools to understand and constrain them lag behind, echoing a recurring theme in AI safety research. Separately, he highlights new work on scaling laws for protein-folding models, suggesting that these models may improve predictably with more compute and data, just as language models have done.
Finally, Clark discusses attempts to assign a monetary value to the risk of AI-driven extinction. This is a contentious exercise, but it forces a conversation about how much society should invest in safety measures. The three topics don't converge into a single thesis, but together they reflect a field that is simultaneously advancing technically and wrestling with hard governance questions.