Raw capture — raw/dean-ball-kimi-open-weight-thread.md
url: https://xcancel.com/deanwball/status/2078133895766114412 title: "Dean Ball on Kimi, Open Weights, and the Deceleration Thesis" author: Dean W. Ball (@deanwball) date_fetched: 2026-07-21 date_published: 2026-07-17
Dean W. Ball (@deanwball), Jul 17, 2026, 3:05 PM UTC
Ball describes Kimi as a strong model on par with top Q1 2026 models for agentic coding, though "very token hungry" and likely not cheap to run.
He expresses surprise that China permits open-sourcing of such capable models, attributing this roughly 75% to "strategic blindness/lack of AGI-pilledness" and what he calls the CCP being "very Yann Lecun-y" on AI, and 25% to export controls limiting Chinese inference compute plus an export-driven strategy.
He argues open-weight models are "inherently decelerationist" and questions why self-described accelerationists embrace them, suggesting the appeal is that they're "effectively ungovernable."
Ball claims an open-weight-dominant world leads to "full AI communism" — AI as state-provided "digital public infrastructure" — which he calls dystopian. He notes he's "never met an open-weight models advocate who doesn't ultimately concede this is where things end."
He predicts the Trump Administration will generate regulatory FUD around Chinese open-weight models via agency soft law rather than banning open source outright.
Finally, he says such models make the world "a bit more dangerous, but not so much more that you'll really notice" — until they're capable enough that you will.
@jsmorph (Jul 17): Says the "head of strategic futures @OpenAI seems a little rattled." Ball denies being rattled, says his analysis is "pretty level-headed" and he's "just pointing out obvious problems with where this goes."
@natolambert (Nathan Lambert, Jul 17): Says he doesn't "fully understand point 4 nor have I heard anyone think this is likely." Ball clarifies: if open-weight dominates, either a business or government must subsidize development costs, citing Yann LeCun's public argument that governments will all host their own models as "DPI" — digital public infrastructure.
@neil_chilson (Jul 17): Disputes that open-weight models are inherently decelerationist, arguing "if one is measuring the speed of real-world deployment and problem solving, open-weight models could still be quite accelerationist." Ball concedes short-term acceleration but argues it diminishes investment incentive for frontier advancement.
@KrishanuAR (Krish Ray, Jul 17): Argues open-weight gets capable AI into more hands, accelerating scientific progress, while closed systems create perverse incentives — citing Anthropic barring biology researchers from "Fable" as a decelerationist example. Ball draws a distinction: open-weight is roughly as diffusion-accelerationist as closed (maybe slightly more), but considerably development-decelerationist due to raising hyperscale cost of capital.
@xlr8harder (Jul 18): Argues open weights have accelerated progress, enabled startups, and kept pressure on labs to innovate — "Liberty helps accelerate." Ball retorts that "open-weight models obviously deter ai capex man. this is really basic economics."
@TimSweeneyEpic (Tim Sweeney, Jul 18): Compares Ball to "an executive at a taco company" speculating geopolitically about "advances in tacos." Ball responds simply "ok."