What a weekend. Two new wars in Asia don’t qualify as top news.
My first reaction to Hegseth’s conflict with Anthropic was along the
lines of: I expected an attempt at quasi-nationalization of AI, but not
this soon. And I expected it to look like it was managed by national
security professionals. Hegseth doesn’t look like he’s trying to avoid
the role of cartoon villain.
On closer inspection, it doesn’t look very much like nationalization. A
significant part of what’s going on is bribery. OpenAI’s president
donated $25 million to a Trump PAC. Dario supported Harris in 2024, and
hasn’t shown signs of shifting his support. The speed with which the
Department of War started negotiating with OpenAI suggests that
rewarding OpenAI was one of their motivations. If Hegseth wanted to
avoid the appearance of corruption, he’d have waited a bit, and
pretended to shop around. But bribery seems to be currently legal, and
advertising the benefits is likely to be good for business.
On the other hand, his attempts to look like he’s punishing Anthropic
look sufficiently clumsy that I’m confused as to whether he wants them
to be effective. He advertised Anthropic as both having the best AI and
as having the most integrity. I’m pretty sure that’s good for
Anthropic’s business.
The breadth of Hegseth’s proposed supply chain risk order is well in
excess of what he can plausibly enforce. Polymarket predicts almost no
net harm
to Anthropic. I’m confused as to what Hegseth expects, and what will
happen when his expectations bump up against reality.
Is it plausible that a deal with OpenAI will serve purposes other than
discouraging domestic dissent? Sam Altman is presumably persuading
Hegseth that OpenAI will be loyal to Trump’s goals. Altman’s track
record suggests that Altman is dramatically less trustworthy than Dario.
It sure looks like Hegseth’s position is that the contract with OpenAI
would be more favorable to the military. Yet Altman is trying to give
different constituencies different impressions about what interpretation
of the contract he will follow. Why should we expect the resulting AI to
care about the safety of anyone other than Altman?
Does Hegseth believe that the Department of War can verify whether an
OpenAI (or Anthropic) AI meets the military safety standards? The
military will run tests on the AI. But it’s pretty hard to mislead an
AI today as to whether it’s being tested versus in a real war. It’s
likely to be harder next year. Can OpenAI or Anthropic train an AI to
act obedient during tests, yet behave more ethically or more loyal to
someone else during an actual war? It’s hard to say.
But not all of Hegseth’s rants are as stupid as critics say. I want to
focus on the alleged contradiction between wanting to use the Defense
Production Act and a supply chain risk order. Anthropic
writes:
They have threatened to remove us from their systems if we maintain
these safeguards; they have also threatened to designate us a “supply
chain risk”—a label reserved for US adversaries, never before
applied to an American company—and to invoke the Defense Production
Act to force the safeguards’ removal. These latter two threats are
inherently contradictory: one labels us a security risk; the other
labels Claude as essential to national security.
While implementing both threats simultaneously would presumably involve
sending contradictory orders, I see nothing contradictory about making
the two threats.
The scariest part of this situation is that there are multiple national
security risks from AI.
It’s very plausible that in the not too distant future, having the best
AI will be one of the most important factors in military power. This
almost justifies using the Defense Production Act, but there are
problems with verifying whether the AI that the military gets would work
the way they want.
There’s also a real risk that an AI company could use the AI it has
deployed in the military to stage a coup. Remember that Sam Altman has
shown more success at handling
coups
than has Trump. This risk might be mitigated by some very select uses of
the supply chain risk order (i.e. something close to the opposite of how
Hegseth is using it).
I see nothing that prevents these two risks from becoming important at
the same time.
The Trump administration doesn’t take AI seriously enough to help with
either of these risks.
The Department of War desperately needs full control over the
development of any AI used to control their weapons. Yet they haven’t
been able to hire the kind of employees who could keep up with frontier
companies. The recent fireworks will make such hiring harder. And the
closer they come to nationalizing OpenAI, the more likely it is that key
employees will leave.
The closest that I’ve found to a good answer is that the Department of
War should use multiple AIs, including at least one open weight AI, and
at least one AI developed within the military, with no single AI coming
close to controlling half of the forces.
P.S. – Trump has occasionally hired competent people. Read more about
this topic from one such person, Dean
Ball.
[Update: Zvi has a long post indicating that the bribe probably wasn’t very important, that Altman has done some good things recently, and that Musk deserves some blame. Zvi seems to have more evidence than I will be able to digest.]
Discuss Read More