[gradus_ad]

These very valid points apply to all companies trying to make money off of proprietary models, which means margins are going to collapse in a vicious price war that will make Uber vs Lyft seem tame.

As margins collapse capex will collapse. Unfortunately valuations have become so tied to AI hype any reduction in capex will signal maybe the hype has gotten ahead of itself, meaning valuations have gotten ahead of themselves. So capex keeps escalating.

None of this takes into account the hoarding effects at play with regards to GPU acquisition. It's really a dangerous situation the industry is caught in.

[wombatpm]

Couple of observations:

Companies use to hoard talent. Now they are hoarding compute, RAM, and GPUs.

Deepseek showed that there are possibly less expensive ways to train, meaning the future eye watering expenses may not happen.

Bigger models may not scale. The future may be federations of smaller expert models. Chat GPTX doesn’t need to know everything about mental health, it just needs to recognize the the Sigmund von Shrink mental health model needs to answer some of my questions.

[johnsmith1840]

Echoing the other comment they showed another big thing which is that the output if an AI model is the AI model. If you mass prompt scrape their AI you can recreate it almost exactly.

Very dangerous if you think about it that the product itself is the raw building block for itself.

Openai spends 1B$ on their model, releases it and instantly it gets scrapped by a million bots to build some country or company their own model.

[chipgap98]

Deepseek showed that distillation is possible. Their results are possible without someone else doing the leading edge training