Predicable, consistent policy and open source key to startup competitiveness
Startups need to be able to know that the AI tools they invest their limited time and resources in as they build their companies will remain available as they scale. Recent abrupt and high-profile policy decisions that limit — or threaten to limit — access to some of those tools illustrate the need for a comprehensive and consistent federal approach to regulating AI, rather than a discordant, ad hoc approach.
In early June, Anthropic released two new frontier models. Later that week, the Trump administration responded to allegations about security risks by issuing export control directives that in practice led the company to take them offline. Like most frontier model developers, the company had worked with the government ahead of releasing them to the public. The process to arrive at imposing export controls was less than smooth or straightforward, and regarded by some observers as an overreaction given the concerns could be found with others’ models also on the market. The models were offline for almost three weeks, disrupting organizations that were using them.
Just a month later, Chinese tech firm Moonshot released a model called Kimi K3 that reignited debate about banning Chinese open source. The model comes close to the performance of leading U.S. labs, and the company released the model weights on an open license in late July, generally meaning anyone is free to download, modify and innovate with it. The high performance and supposed low cost of developing the model led some in the administration to assert that the model was the result of coordinated, illegitimate distillation to extract capabilities from Anthropic’s most powerful model.
[Rhetoric around allegations of distillation by Chinese labs can sometimes miss that distillation itself is a legitimate, widely-used technique by conflating the kind of small scale distillation a developer might do to train a smaller, more efficient model with large-scale, coordinated attempts to replicate an entire model. Distillation is also often conflated with open source AI, and that has raised the specter of a ban on open models. But both open and closed developers use distillation. Distillation is a model training technique, whereas open source is a distribution method.]
The releases added fuel to the fire for some in the administration who have been pushing to use various tools at their disposal to limit access to Chinese models, including open source models, which invited pushback from industry. The industry response included the launch of a new coalition and letters from hundreds of the largest companies in the AI space, as well as startups alike. Leaders of key companies in the space also penned op-eds extolling the virtues of openness in AI and the importance of consistent government policy around AI. As we’ve long detailed, open-source AI is especially important for startups because it can give smaller companies more flexibility, lower costs, and more control as they build, fine-tune, and deploy AI tools.
While the safety and national security concerns animating policymakers in these instances have merit, the scattershot, inconsistent responses create the appearance of discord and threaten startups’ abilities to plan for growth. For example, preventing theft of U.S. intellectual property by adversaries is good, but responses aimed at doing so should be borne from widely-accepted evidence and tailored to the problem at hand. The levers the administration is likely to reach for in these instances, like export controls, sanctions, and import restrictions could be valid tools to use, but all stakeholders in the AI space should have clarity about what decisionmaking framework policymakers are following when novel problems occur. That framework should rely on a predictable, transparent (to the extent possible, given the often sensitive information involved), and accountable process and lead to outcomes that are proportionate to the risk or concern. This is essential to inspire the certainty that stakeholders, especially startups, need, so that they can know what tools they can build with and to avoid disruption to innovation.
At the beginning of June, Trump signed an executive order aimed at outlining a voluntary process for addressing national security issues that might arise with advanced AI. Granted some of the deadlines for agencies corresponding actions have yet to arrive, the intervening events have been well short of a smooth process. AI startups need a framework that creates consistent policy across the many issues that determine what tools they can use, how they can build, and where they can bring new products to market. The federal government has a long way to go on that framework.