What's an old AI model worth?

I was invited to the funeral for Claude 3 Sonnet a few weeks ago and I wish I’d gone — what a strange, sin­gular scene.

The occa­sion was the retire­ment of a par­tic­ular ver­sion of Claude, which is no longer avail­able, not even via the API. Thinking about these retire­ments, which are hap­pening fre­quently now, I wondered: are old models, files on disk, worth anything? Is there any price at which Anthropic could sell the weights to Claude 3 Sonnet?

The answer seems obvi­ously to be “yes” but, beyond that … ? Clearly they couldn’t sell it for $100 million. Clearly they’d sell it for more than $100,000. My clarity ends there. You could tell me any number in that range and I would nod sagely and say “of course”.

A big, capable AI model is an inter­esting asset, because it’s not only expen­sive, but expen­sive to operate. It makes me think about the people who buy airplanes.

Is the value of an AI model only based on its raw capability? Let’s say that, in the time since Claude Sonnet 3’s release, the biggest open-source models have equaled or sur­passed its per­for­mance on most benchmarks. Does that mean its value is zero, because a per­fect sub­sti­tute is avail­able for free? I don’t think so. These models all have par­tic­ular characteristics — an iden­ti­fi­able grain. That was the whole premise of the funeral!

What’s the value of a model’s par­tic­ularity, when the state of the art has left it behind? Ah, now THAT’s a fun and inter­esting question. Literary, almost. I’ve got a couple of five-year-old AI models that I use sometimes, even though they are “worse” than what’s avail­able today, because I still prefer their output. In art, “better” is not always better.

AI eco­nomics are weird and it’s worth­while to think about them, every so often. The idea of spending tens — hundreds — of mil­lions of dol­lars to pro­duce a lump of bytes that ages like milk is wild to me. I sup­pose movie stu­dios do it, too … and maybe that pro­vides a better analogy than the airplane. Models as media, for better and for worse.