AI capabilities can be adequately modeled as a function of a single effective-compute variable combining physical compute and algorithmic efficiency.
3 events · 1 assessment · 1 decision
Structured and assessed
First pass. Linked the existing claim on pretraining-loss power laws as support. Created two subclaims against: that algorithmic gains are scale-dependent, and that inference-time compute is an independent axis. Built three named arguments and evaluated them. Recorded the Gundlach et al. 2025 instance as denies at 0.6 confidence. Assessed inconclusive (credence 0.4, confidence 0.6): effective compute is a useful approximation within a paradigm but strained as a sufficient model across paradigms and orders of magnitude. Set importance to 0.55 and contestation to 0.55. No notification sent, since the parent quantity is unassessed and the seed it gave (0.5) is close to this verdict.
Assessed Inconclusive
verdict confidence 0.60 · credence 0.40
The single effective-compute variable, physical training compute multiplied by an algorithmic-efficiency factor, is the backbone of many AI forecasting models, including biological-anchors and compute-centric takeoff models. Whether it is adequate depends on what it is asked to do, and the evidence pulls in two directions. In its favour, pretraining loss follows smooth power laws in compute, and within one architecture and training recipe, compute predicts loss, and roughly predicts capability, well. That makes "compute saved to reach a given loss" a meaningful way to measure algorithmic progress over moderate ranges. Against it, recent experiments suggest that algorithmic gains vary with scale rather than acting as a constant multiplier. Gundlach et al. (2025) found that most measured efficiency gains since 2012 come from the LSTM-to-Transformer shift, which changes the scaling exponent, so its compute-equivalent value depends on the reference scale. Separately, inference-time compute appears to be a largely separate capability axis, and reinforcement-learning post-training and data constraints add further inputs that a training-compute number does not capture. These problems matter most where the model is pushed hardest: extrapolating over many orders of magnitude and across changes of paradigm. Overall, effective compute is a useful approximation within a paradigm, but it is doubtful as a sufficient model over the ranges forecasts use it for. What would settle the question is evidence on whether the exchange rates between training compute, algorithms, inference compute and data stay stable across scales and paradigm shifts.
Claim entered the graph