Import AI 454: Automating alignment research; safety study of a Chinese model; HiFloat4

The most consequential item is the framing around automating alignment research itself: the premise is that human researchers are a bottleneck on safety work, and that AI systems should help close that gap. That's a significant assumption worth examining on its own terms, separate from whether any specific technique works.

The Stanford AI Index coverage from MIT Technology Review (April 13) documented how fragmented expert opinion has become on AI progress and safety, and this newsletter lands squarely in that contested space. The question Clark raises about market pricing of AGI connects to the economic anxiety thread running through TechCrunch's tokenmaxxing coverage from April 17, which noted widening gaps between AI insiders and the broader public. The safety evaluation of a Chinese model is worth noting as a relatively rare data point: most safety benchmarking covered on this site has focused on US-origin models, so independent evaluation of a Chinese LLM is a different kind of signal, even if the methodology behind it isn't yet clear from the summary.

Watch whether the automated alignment research approach Clark highlights produces any peer-reviewed output within the next two quarters. If it does, that would suggest the bottleneck framing is serious; if it stays in newsletter-summary form, treat it as a hypothesis still looking for evidence.