You can use it, just don’t marry it

Yesterday's email drew a bunch of replies, all saying roughly the same thing: I've repeatedly mentioned how skeptical I was of generative AI, so why am I now saying it's fine for some things?

I actually never said otherwise. Narrow uses do exist. I'm not the AI police, and I'm not under the illusion that scolding people will stop them using the tools that are sitting right there on their desktop. If you're going to use it, I'd rather you went in with your eyes open – though, right now, that’s about as useful as arguing for restraint during a gold rush.

What I am saying is: don't build anything you care about on top of it. Not yet, maybe not ever, the way things are going.

Let’s start with the money. The unit economics are fiction. Every query you run is subsidised by someone else burning a pile of cash. And subsidies are, per definition, temporary.

Ask what happens when the price tag reflects the actual compute; not the VC-funded fantasy version. Enterprise buyers are already flinching as per-token prices creep toward reality. And those prices are still below cost for the labs selling the tokens. Ed Zitron has been tracking this collapse-in-slow-motion better than anyone, if you need the receipts.

Of course, there's a second layer of subsidy underneath the first (defence contracts, tech nationalism, etc) which might keep the lights on longer. But that’s not forever.

When I raise the financials, people point to “open-source” models as the answer. Two problems:

• First, they’re not open source, they’re open-weights; the source and training data are not public.
• Second, much of their openness exists as a counterweight to US frontier labs. And Chinese backers will only fund that for as long as it serves their interests. That’s already shifting: Alibaba now releases models via its API first, Kimi requires attribution, and even Meta’s latest “Muse Spark” model has abandoned open weights.

We risk ending up with frontier models locked behind a few monopolies at very high cost, while open-weights models stay frozen in time and gradually become less useful.

Then there’s everything else:

• Hallucination is baked into the technology. OpenAI itself says it’s mathematically inevitable, and most people don’t check outputs.
• It’s making people less capable, including students.
• The power, water, and land use of data centres is substantial. And big tech have discretely dropped environmental pledges from their sites while chasing market share.
• And that’s before we even get to the ethics: it’s all built on extraction and plagiarism, concentrating public knowledge into a few private hands.

I could go on. But here's the actual answer: want to poke at it, play with it, see what it's good for? Knock yourself silly, ethics notwithstanding.

Want to depend on it – build your workflow, your product, your job around it? You’re putting critical work in the hands of a service that can raise prices, degrade, or vanish between one quarter’s earnings call and the next.

Colin