An OpenAI exec, two weeks out of Trump's AI team, wants Washington to make Kimi K3 harder to use

Two weeks into his new job at OpenAI, and two weeks out of the Trump administration's AI policy shop, Dean Ball posted a long thread on X in early August calling China's release of a free, top-tier language model "reckless" and asking the Trump administration to create "large amounts of regulatory risk" around Americans using it. The post came days after Moonshot AI, a Beijing-based lab, open-sourced Kimi K3 and posted the trained weights on GitHub for anyone to download, a contrast to OpenAI's and Anthropic's closed products. The release is unusually complete: weights, a tech report, and a working download. US labs have to match all three without a price to charge against.

Kimi K3 reportedly trades benchmark blows with leading closed-weight systems at a sliver of the cost. Trade press has put it at 2.8 trillion parameters, and the published weights let any company or developer run and fine-tune the model without paying Moonshot. The week of the release, the tech-heavy Nasdaq sold off and the S&P 500 slipped on the read that a cheap, open alternative is good enough to dent demand for paid inference. Closed-weight lab executives have been publicly warning that the cost of training and serving frontier models is climbing, and Kimi K3 lands on top of that squeeze. The math behind Ball's post is straightforward: when a lab cannot win on price and the cost curve is bending the wrong way, the next move is to ask the government to add friction the market will not.

The Futurism piece that first surfaced the post called the conflict "glaring." Ball framed the ask plainly: "open-weight models are inherently decelerationist," a label for a worldview that argues slowing AI's spread is the responsible path. He then predicted the Trump administration would "create large amounts of regulatory risk around the use of open-weight Chinese models," and acknowledged the goal was to "spread FUD" (fear, uncertainty, and doubt) that deters American hyperscalers, the largest US cloud and AI buyers, from buying the Chinese alternative. The strategy is plain: change the cost equation through Washington, not through product.

Ball's incentives and his prior seat are the same desk. He is asking a regulator he just left to harden the ground against a rival that is beating his new employer on price. If the Trump administration takes the ask, the friction lands on US cloud and AI buyers, not on Moonshot. That posture is the part a wire story will skip.

The dual-use concern is real, and the steelman is this: a frontier model whose weights anyone can download is a national-security object the moment a foreign adversary decides it is, and a domestic closed-weight lab is not paid to internalize that externality. Capable open-weight models can be fine-tuned for harmful use, and once the weights are out, no recall mechanism exists. The policy question is whether the US government should respond to the spread of the weights, and at what point the response stops being a market-shaping tool. The ask, not the concern, is the story, and the ask lands from a seat the asker just left.

Pushback came fast. Alexander Green, founder of the AI company Littlebird, replied that Ball's line of thinking was "really crazy stuff to believe." Commenters on the thread called Ball "a little rattled." Ball did not walk the post back.

The next test is whether the Trump administration writes Ball's framing into AI procurement policy, and whether any hyperscaler publicly picks Kimi K3 or its derivatives for a production workload.