Elasticity of substitution between experimental compute and cognitive research labor in frontier AI software R&D
2 events · 0 assessments · 1 decision
Assessed
First pass. The budget was tight, about 20% spent by turn 4 and each turn costing a lot, so I kept the quantity atomic with no named arguments. The evidence is one study with two specifications plus calibration conventions, and I weighed these in the estimate prose. I recorded the estimate as a distribution with median 0.85 and a 90% interval of 0.2 to 2.8. The parent claim's assessment ("limited compute would bottleneck progress", inconclusive) already matches this, so I sent no notification. A later pass should do three things: record Davidson's post as an instance, read the Whitfill-Wu paper in full to get the identification details, and consider named arguments for the frontier-experiments reading versus the total-compute reading.
Claim entered the graph