I keep reading ‘AI isn’t going away’, but I don’t think the people saying it have thought through the economics.

An LLM is, roughly speaking, two parts. One defines the structure of the model: the kinds of layers, their arrangement, and the connections between them. The other is the weights. The first part is quite similar to any other software artefact. Once you have a copy of it, it keeps working. This is the cheap bit to build, but the tricky bit to design.

The weights are the result of training. You need to throw a lot of data and a lot of compute at a system to create the weights. Once you have done this, you can use them indefinitely. The problem is that the weights include all of the data that is embedded into a model.

If you train a model today to use for programming, it will embed almost nothing about C++26, for example. If you train it on news, it will not be able to answer any questions about things that happened after today. Weights from today quickly become outdated.

This is one of the big costs for LLM vendors. Just as a snapshot of Google or Bing’s index rapidly decays in value and needs constantly updating, so do LLM weights.

Training these things costs a lot of money (DeepSeek claims only a few tens of millions, but it’s not clear the extent to which that was an accounting trick: how many millions did they spend training models that didn’t work?). For ‘AI’ companies, this cost is a feature. It s a barrier to entry in the market. You need to have a load of data (almost all of which appears to have been used without consent) and a huge pile of very expensive GPUs to do the training. All of this is predicated on the idea that you can then sell access to the models and recoup the training costs (something that isn’t really working, because the inference costs are also high and no one is willing to pay even the break-even price for these things).

So if the companies building these things speculatively can’t make money, what happens? Eventually, they burn through the capital that they have available. Some weights are published (e.g. LLaMA), but those will become increasingly stale. Who do you expect to spend real money (and legal liability) training LLMs for no projected return?

If you’re a publicly traded company and are building anything around LLMs, you probably had a legal duty to disclose these risks to your shareholders.

#Breaking - Innenminister Karner will die Zahl der überwachten öffentlichen Orte in Österreich verfünffachen – per Erlass, mitten in der Sommerpause. Ohne parlamentarische Debatte, ohne wissenschaftliche Grundlage, ohne transparente Begründung.

Mehr dazu 👉 https://epicenter.works/content/verfuenffachung-der-videoueberwachung-in-oesterreich-geplant