The latest edition of Jack Clark's Import AI newsletter ties together three developments that, taken separately, might look unrelated but together sketch a specific worry: the gap between the most capable proprietary AI systems and openly available ones is shrinking faster than most policy frameworks assume.
The centerpiece is Moonshot AI's Kimi K3, a 2.8-trillion-parameter open-weight model whose most striking demonstrated capability was autonomous chip design. In a single 48-hour unsupervised run, K3 reportedly built, optimized, and verified a chip design using open-source electronic design automation tools on the Nangate 45-nanometer standard cell library, without a human directing intermediate steps.
That result lands alongside a separate analysis from the UK's AI Security Institute (AISI), which measured how far leading open-weight models trail the best closed models specifically on cybersecurity tasks. Across a set of 70 narrow cyber capability evaluations, AISI found that current open models such as GLM-5.2 and DeepSeek V4-Pro now perform comparably to closed frontier models released only four to seven months earlier, a narrower lag than the six-to-ten-month gap the institute measured through most of last year. GLM-5.2's performance landed closest to Claude Opus 4.6, released roughly four months prior, while DeepSeek V4-Pro fell between two other closed-model generations. AISI says it intends to run the same benchmark against Kimi K3 once its weights are public.
Much of AI safety and policy thinking to date has rested on the assumption that a small number of identifiable actors control deployment of the most powerful models, making it possible to intervene at the platform level through classifiers or customer-verification gates. Import AI argues that widely diffused, hard-to-control open models like K3 undercut that assumption directly. The upside, per the newsletter, includes a likely boost to entrepreneurship and broader access to what it calls sovereign intelligence for anyone able to run the model. The corresponding downside is a wider set of unknowns about how that capability gets used once no single company can pull a plug.
The same issue notes Google DeepMind chief executive Demis Hassabis's proposal for a US-led global AI oversight body, floated as a way to systematically screen advanced models before wide deployment and coordinate slowdowns if serious risks emerge. Reactions from figures including Sam Altman and Elon Musk were reportedly favorable, though the proposal remains an essay rather than a concrete legislative framework.
Read together, Import AI's framing is that the timeline policymakers assumed they had to build effective oversight mechanisms just got shorter. Whether that pushes governments toward faster, more coordinated action, or toward reactive controls targeted narrowly at specific companies or countries, is the open question the newsletter leaves for its next issue.