Anthropic's apparently starting to learn that it can't have its cake and eat it when it comes to working with the military

Running an AI company isn't a cheap thing to do, especially if you're at the forefront of developing new models. You need to buy or rent hundreds of thousands of GPUs, pay huge energy and water bills, and hire lots of highly skilled and qualified staff. So, where does one find a handy source of money for all of this? In the case of Anthropic, it's turned to the military as the pot of bountiful cash, but according to one report, it's now discovering that all that glitters isn't gold.

The agreement between Anthropic and the then-named Department of Defense to "prototype frontier AI capabilities that advance U.S. national security" was announced last July, and while it was only given a capped $200 million budget—a fraction of the money it's received from Microsoft and Nvidia—the makers of Claude were looking "look forward to deepening our collaboration across the Department to solve critical mission challenges".

However, Reuters claims that six months on, the relationship between the now Department of War and Anthropic has taken a bit of a turn, and it's apparently all down to the safeguards that are built into the LLM models. Those that specifically prevent them from being deployed in scenarios such as autonomous weapon targeting and domestic surveillance.

It's being claimed that Anthropic is arguing the presence of these safeguards is a key part of its usage policies, whereas the DoW takes the view that as long as no laws are being broken, it should be free to use a commercial AI however it sees fit, including the removal of the safeguards. Hence, why the two are at loggerheads.

Given Anthropic's history, though, you'd be forgiven for thinking that the AI company would have surely expected this right from the start. After all, it has already teamed up with the likes of Lawrence Livermore National Laboratory and Palantir in the past, all for Claude to be used in classified areas of security and intelligence.

So if all the above is an accurate description of the current state of affairs between the AI firm and the US military, I find it hard to believe that Anthropic wouldn't expect these demands from the Department of War.

Naturally, we have no idea as to precisely what has been discussed between the two parties. It might be a matter of money, for example, with Anthropic effectively saying, 'Sure, we can switch all that off, but that's a whole pile of extra work, which is another contract entirely.' That said, Anthropic already has a special version of Claude for "US national security customers", so why would the DoW not be using that one?

Whatever the situation may truly be, I suspect that Anthropic is beginning to learn that while it's all well and good have a set of universal usage standards, the moment you make exceptions to some of these for specific customers, others will want the same. And if that customer just so happens to be the military, with its enormous cake of money, if you don't like the slice that's been given to you, other AI companies will surely rush in to pick up the crumbs.