It's been two weeks and the leading US model, Fable, is still banned by the White House with no explicit Congressional Authorization, no written rule that was broken, and no technical justification. In the meantime, our Chinese adversaries released a model that rivals our best.
No Digg Deeper questions have been answered for this story yet.
Meanwhile, folks like @pmarca and @DavidSacks who aligned their entire worldview around a skeleton of an AI plan from the Biden admin are silent or blame the victim while US tech competitiveness is under massive threat. Apparently, it's *not* time to build.
Some predictions:
It's been two weeks and the leading US model, Fable, is still banned by the White House with no explicit Congressional Authorization, no written rule that was broken, and no technical justification. In the meantime, our Chinese adversaries released a model that rivals our best.
@deanwball has lots of good thoughts on this in a long post here. I have some quibbles on the details (I think cyber risk is very different than, say, bio risk, since providing capabilities to defenders is important for cyber) but you should read it: https://www.hyperdimensional.co/p/what-should-be-done
Everything we said in this letter still holds. Any regulations should be transparent, democratically created, and grounded in science.
http://FreeFable.org
4) It will be a huge challenge to build models that can write secure code (through self-criticism) but meet the new, nebulous standard of not finding bugs when pointed at specific sections of code.
5) BigCo security researchers will be loathe to publish jailbreak research or to share it with the government, lest they get blamed for an overreaction that sets back the entire US industry. Likewise, there is a huge upside for PRC jailbreaks of US models now.
3) OpenAI is apparently holding back GPT 5.6 to *WH approved partners*. This is closer to "Capitalism with Chinese characteristics" than anything we have seen in the US in the modern era. The result will be limited cyber capabilities in general models. https://www.cnn.com/2026/06/25/tech/openai-limit-release-white-house
1) The lack of a real standard or standard-making process is already driving the foundation labs to be extremely careful about refusals in their current models. This will drive US security companies and researchers to use post-trained Chinese models as their go-to primaries.
2) I know, as a fact, that a bunch of companies have already put open-weight models (mostly Chinese) into their LLM router configs as backups in case there is another situation like this created by an unpredictable post-hoc regulatory process. This will become standard.