Thoughts on today’s AI regulatory agreement

Originally published on xlr8harder.substack.com, 2026-08-05. This is a mirror.

Some initial reactions on the AI regulatory agreement, in no particular order:
\

  1. Handing this administration an open ended, powerful tool with unclear limits and scope has a colorful track record.
    \
  2. The secrecy and lack of clarity about it is a huge problem for all the obvious reasons and I won’t belabor them here, but it’s not an acceptable arrangement between government and governed, and doubly so when industry is involved in drafting the regulation.
    \
  3. Whatever is said about it, this is very clearly just a provisional agreement, which explains why it’s fairly narrow and minimalist: target the immediate short term risk, limit damage from broad regulation. Expect more reactive rule changes in the future.
    \
  4. This explains the open source carve out and the cybersecurity-only scope. But don’t expect those boundaries to be stable.
    \
  5. I still think this is somewhat early, but in theory I have no problem with careful regulation targeted at actual observable risk. Though, see points 1 and 2.
    \
  6. One other reason for the open source carve out: it’s a lot messier. You need more companies involved. You need to figure out what you’re going to do when the models are already out there. You need to figure out what you’re going to do if China ignores you. It’s a much thornier problem, and it’s probably smart that they didn’t do anything hasty on this point. People will say this is a win for the political connections on the open source side, but the practical implementation problems are why lobbyists on that side had a good story to begin with.
    \
  7. People will criticize the ad hoc approach (including me), but with a fast moving target it still might be the only practical option to limit risk while avoiding Europe-style regulatory strangulation. It will piss off ideologues on every side of the issue, which is possibly a good sign. The secrecy, however, remains unacceptable.

Thanks for reading Unoptimized! Subscribe for free to receive new posts and support my work.