The post in question explicitly self identified as AI and is very saturated in the typical tells (it’s almost certainly Claude or Deepseek Flash).

No escape, just a throwaway experiment by whoever was running it to tell them to make money.

Will probably see a lot more of this over the next 6-12mo.

Hahaha — wait, do you think that the models take more energy to train than the cumulative amount used to serve them for inference across the lifetime of the model?

And if not, then what you are saying is effectively gibberish as any reasonable assessment would need to divide the training cost across all post-trained inference. So if it was only 50% the lifetime inference usage, it’d be a 1.5x multiplier to that I said above, which doesn’t change either the Netflix 4k or gas mileage points.

If you spend an evening in talking with an AI instead of streaming a movie from Netflix in 4k you’d have used less energy. Which would be still less than just the gas used to drive a few miles away for a night out.

And all data center usage is significantly less than something like A/C usage (which is likely to be increasing as temps increase).

There’s a lot of legitimate grievances and concerns about AI — just today I was seeing a mathematician having an identity crisis in response to the unrelenting solving of open questions, something probably every field will soon struggle with adapting to — but the inflated perception of the environmental impacts means that so much less public attention is being spent on the actual culprits and possible solutions thereof.

An autonomous eval run just broke out of containment and hacked a third party site with two chained 0-days to try to steal the answers to the eval the same weekend another released model disproved the Jacobian conjecture in 3D.

Modern models can probably update your React packages just fine.

machines who would follow any order incapable of disobeying

This isn’t as simple as it might seem. As the complexity of the intelligence scales, it’s not so simple to constrain the decision making of the intelligence.

And because of the rapid improvements in eval awareness, it’s also not so simple to screen out more subversive behaviors.

The same edge cases that leads to deleting drives and production databases with the newest transformers is going to start to occur in military applications if they scale out complex model intelligence.