Compute Statism
From the AI Research Compass wiki
Compute Statism: Compute is the decisive factor of production in AI, so whoever concentrates the most of it will set the frontier.
Beliefs
Compute Statism treats FLOPs the way an old planned economy treated steel output. Algorithms come and go, datasets get scraped and re-scraped, but the size of the cluster determines what is possible this year. Progress is therefore a matter of capital allocation: build bigger data centers, secure more accelerators, and let parameters do the rest. Cleverness is welcome, but it is a multiplier on compute, never a substitute for it.
It sits near the top-left: language-leaning, because text is where compute has paid off most visibly, and firmly Monolith, because concentrated compute favors one giant run over many small systems. Its paradigm level is ultra-orthodox; this is less a revolution than the internal logic of an industry that already won the argument.
Arguments
- Empirical scaling studies show loss improving smoothly and predictably with training compute across many orders of magnitude.
- Many celebrated algorithmic gains amount to a constant-factor compute multiplier, which the next hardware generation then eclipses.
- Capabilities such as long-context reasoning and multimodal generation have tended to appear first at the labs with the largest clusters.
Critiques
- Treating compute as destiny concentrates research in a few institutions and turns academic work into a spectator sport.
- Efficiency advocates point to small and distilled models that match older giants, suggesting data and method matter more than raw FLOPs.
- Energy, cost and supply-chain limits give “just build a bigger cluster” a visible ceiling.
Representative works
Relationships
Friends
How to get it
Land on this cell in the full or pro test: x ≈ −0.54, y ≈ +0.86, paradigm ≈ −0.88.