Intentional News

Small Models, Decentralised Power

3 min read

Small Models, Decentralised Power

Origin: the owner’s comment on the Harari interview (260826 Yuval Noah Harari Full Interview Economist) — “The discussion was a bit simplistic… they may not have full handle on the reality of local AI and ability to rollout at scale… This will I believe be much greater by end of 2027 as hardware costs drop and powerful Chinese open weight models (and maybe US models like Gemma) emerge further. From what I can see if one is sufficiently smart and handle intelligently one can be powerful with smaller models today on cheap enough hardware. Plant a near term Envision note on this.” Seeded 2026-08-27.

The thesis (the owner’s + Alf’s read)

The dominant AI narrative — “a handful of trillion-dollar labs in the US and China concentrate all AI power” — is already being undercut by cheap, local, open-weight models. By end of 2027, this will be much more pronounced:

  • Hardware costs keep falling — inference that needed a datacentre a year ago now runs on a laptop/small box.
  • Powerful open-weight models are emerging — Chinese open-weight models, and US models like Gemma — closing the gap to the frontier for capable use (even if not for frontier training).
  • A sufficiently smart operator can already be powerful with smaller models on cheap hardware today. That’s literally the architecture the owner runs with Hermes/Alf — the vault is a live example of the thesis.

Why this matters (the connection the owner flagged)

This mitigates the concentration-of-power problem that dominates both interviews:

  • Harari and Musk both assume AI power aggregates into US+China empires (Harari’s “two empires” / “extreme imperialism”).
  • If capable AI is cheap, local, and open, the power of AI decentralises too — the “single point of information” that an autocrat or super-corp could control doesn’t exist if everyone runs their own.
  • This is the counterweight to Loss of Faith in Democracies - Global Problem: decentralised AI is one concrete mechanism by which democracies keep the capability edge that Harari says they need.

Caveats / honest limits (Alf)

  • Open-weight ≠ frontier training. Training capability (and thus the frontier of what’s possible) will remain concentrated for a while — decentralisation addresses capability diffusion (who can wield it), not frontier ownership.
  • Decentralisation cuts both ways: the same cheap local models empower individuals and bad actors (a lone actor with a capable local model is a new threat class).
  • “Sufficiently smart” is doing a lot of work — the harness/context-engineering matters as much as the model. That’s the competency-ladder point: capability follows skill, not just hardware.

Open questions

  • Track the open-weight frontier (Chinese models, Gemma, Llama, etc.) — revisit this note end of 2027.
  • Does “capability diffusion” actually blunt great-power AI rivalry, or just change its form?
  • does this vision further put to question the value to be gained from the massive data centre rollout and AI Tech bubble? Do frontier models really add that much value over a smaller model more expertly implemented to solve business problems? Perhaps working in hybrid with humans in HITL service models?

Graph View

Backlinks