Four scaling paradigms keep postponing the wall: params, CoT, recurrent depth, multi-agent
gordic_aleksa · x · 2026-09-25
Aleksa Gordić argues that every time scaling seems to hit a wall, a new axis opens up—four so far:
- Data & params (Kaplan scaling laws), the original pretraining engine;
- Chain-of-thought, the inference-time scaling exemplified by OpenAI's o-series;
- Recurrent depth—looping computation in latent space;
- Multi-agent scaling, exemplified by Claude.
His take: "there's not enough compute in the galaxy for what's coming"—a good time to be selling compute or converting compute into intelligence (or both). It's a bullish big-picture read on sustained compute demand.
More from AGI Musings
- Harvard Kennedy School Researcher Critiques Dario Amodei's AI Slowdown Essay — StephenLCasper · 2026-09-25
- Ezra Klein proposes banning AI from writing code, stunning the AI community — soleio · 2026-09-25
- ICLR debate: bull-spotting reviewers opting out skews the reviewer pool — sarahwiegreffe · 2026-09-25
- ICLR reviewer pool debate: when experts opt out, bad papers gain an edge — BlancheMinerva · 2026-09-25
- Hugging Face CEO: AI risk comes from secret frontier labs, open-source is the fix — ivan_bezdomny · 2026-09-25
- Anti-AI sentiment turns to street violence, scientist calls for honest public framing — josephdviviano · 2026-09-25