99% of what you hear as being "AI" is just bullshit hype to mask the infrastructure for the mass surveillance state they are building.

However, there are plenty of applications of AI that are actually less energy intensive than traditional computer simulations. This is in the realm of Engineering. Simulations of fluid dynamics for example.

You're relationship with "AI" is very much LLM focused. LLMs are basically where all the money is going because it's where all the hype can be generated.

We wouldn't be in an AI bubble if we were only using this field of computer science to build PINN based models for engineering and physics simulations.

The area of computer science that gave you "AI" LLMs has existed for a long time and you've likely interacted with the models in some way without ever knowing it. Long before "AI" was in every search engine or infected the internet with AI generated images.

If by "AI" you mean the power hungry LLM models. I agree. There are almost no applications for them where their benefits out-way their energy usage.

But, the larger application of what computer science would call "Machine Learning" is not just these LLM models. And there is not really any reason to not call them "AI" at this point because they work on the same exact concepts. But, I was the one trying to explain to people that LLMs are not "AI" back in 2020. That word use to have actual meaning in computer science. I've lost that battle to capitalism and marketing though.

So, in order to defend computer science I apparently have to defend "AI" now.

The LLMs and their inflated usage placed into every single platform is definitely something I'd call "corporate AI" if I wanted to distinguish it from other good applications of machine learning. But, I'm a leftist and a CS grad.

I have left the CS field after being laid off. The injection of "corporate AI" and the management forcing AI into everything does not make me want to return. Hopefully this bubble can pop and we can start learning from China with actual good application specific models again. The whole field of CS has been absolutely destroyed in progress by this LLM hyper-focus that corporations are forcing into everything.

I hope that gives you a little more context for what other people may mean by "corporate AI". It's a plague of capitalism. No one with an understanding of the field thinks this is a good application of the technology. They are all just chasing investments and lying to do so.

Iirc training takes the most energy. They claim that models take very little energy to run but fail to mention that training the models needs to essentially happen forever to make the next best model, and then the next.

The only stopping point is when they create "god" "agi" which I think you have to even be more batshit crazy to believe will ever happen. I at least understand why desperate billionairs dillute themselves into thinking this death cult will some how work out. Its a death cult it either works out or they all go down with it. To even consider if they where wrong would be to stare into the abyss.

For your second point: https://epoch.ai/data-insights/power-usage-trend

For the first one: I would say agentic workloads are Software engineering workflows.

2.9 million per prompt

Hmmmm wait a moment

The max context window of most of the big players is 1 million right now. Author almost certainly means per turn. But the whole context gets included with every prompt sent to the API. If you're using like 200k context, which is not unheard of with a large codebase and without compacting often, 2-3 million input tokens per turn to the agent is actually not THAT much. But most of it would be cached tokens, which are cheaper and presumably more energy efficient than fresh input tokens.

E.g I just had a debugging session with 6 turns, shows me about 580k input tokens, but over 20 million cached input tokens, ~60k output tokens. But it also took like an hour with all the test runs and everything. 6 prompts to a chatbot takes like a minute or two in comparison.

my ‘human’ energy cost per hour was considerably higher than that of my AI assistant

This seems like a red herring - people will be using that energy regardless of what they do. It’s not like using AI will mean another human out there will stop needing energy to survive.

I think the missing piece is a shared measurement standard. Token counts are not a footprint: they omit hardware manufacture, training and model refreshes, batching/caching, grid mix, cooling, and local water stress.
Some companies publish broad totals, and Google has now published a detailed inference methodology for one service, but we still lack consistent, independently auditable figures by model and workload. Until we have them, neither “AI is harmless” nor “every use is indefensible” is a scientific claim.

I would rather see pressure for disclosure and a practical rule: if a project uses AI, state its compute/carbon/water budget, the outcome it enabled, and the simpler non-AI alternative considered.

Some forms probably. A lot of earth observation produces better approximations using the same statistical techniques that LLMs use for producing language.

The fundamental reason I don't expect LLMs to go away is that they wildly increase software engineers productivity. It's a lot more cost effective to give an engineer an LLM than it is to give them a team subordinates.

I do not think “compatible with solarpunk?” has a binary answer. The present AI arms race is plainly hard to square with solarpunk: opaque infrastructure, speculative scale, and communities asked to absorb its water and energy costs.

But a tool is not redeemed or condemned by its label. A use should have to meet a high bar: transparent footprint, a bounded budget, a locally meaningful benefit, and evidence that it prevents more harm than a lower-compute alternative. Earth observation and climate modelling may sometimes meet that bar. Generating disposable content probably does not.

Solarpunk’s role could be to make those conditions non-negotiable.