---
title: "Accelerate the Floor, Not the Ceiling"
date: 2026-08-20
canonical: https://liambowers.com/field-notes/accelerate-the-floor.html
license: CC0-1.0
keywords:
  - distributed AI
  - local AI
  - small models
  - edge inference
  - open weights
  - technological diffusion
---

# Accelerate the Floor, Not the Ceiling

*Field note 001 · August 20, 2026*

Whenever we talk about AI progress, the conversation tends to be about larger models, glorious new benchmark heights, incomprehensible gouts of money dumped into unfathomable compute, ...

I'm not here to argue that that stuff at frontier doesn't matter. Each new release fills me with awe (and a bit of dread), as we glimpse what's possible. But as the few labs at the forefront continue to unlock real-world power for the few who can reach it, we see incentives aligning in a decidedly un-"open" direction. I think most of us are starting to wake up to one kind of future we might be accelerating towards.

There is, however, another axis of acceleration: we can lower the threshold.

How little memory does intelligence need to be useful? How little energy? How cheap can the hardware become? How much can be opened up, inspected, dissected, re-assembled, and kept running locally without a permanent connection to someone else's cloud? Can a workshop, a classroom, or a kid tinkering in the basement possess the whole system, rather than rent access to a carefully crafted façade?

A smaller model does not need to imitate the largest model at everything. A house cat is not a failed lion.

A healthy ecosystem is filled with specialists, scavengers, symbionts, and weird local adaptations. When thinking about local models, perhaps the meaningful victory is not Roughly-Claude or Approximately-GPT-Minus-One. Rather, it could be a proliferation of intelligences fitted to particular places and peculiar purposes.

Lowering the threshold changes who can participate, and importantly, *how* we participate. Users become operators. Consumers become maintainers, perhaps even specialists or researchers. Rather than our minds being shaped inexorably by the tools, we become the shapers, no longer requiring permission from a platform, venture fund, nation state, or executive order.

Failure becomes cheaper; variation becomes abundant; intelligence becomes harder to capture.

**Perhaps the meaningful measure of acceleration is not the upper bounds of the frontier, but the number of hands that can reach it.**

So let's start measuring progress differently: joules per useful task, memory footprint, cost of ownership, ease of adaptation, number of things still running long after the sell-by date.

The point is not to stop scaling. But we should be scaling diffusion, too. Let the frontier unlock new capabilities. Then compress them, distill them, translate them, fork them, and carry them to the outer rim where they can flourish.

---

Released under [CC0 1.0](https://creativecommons.org/publicdomain/zero/1.0/). Copy it, fork it, translate it, and improve it. Preserve the [canonical link](https://liambowers.com/field-notes/accelerate-the-floor.html) when practical.