This passage makes me want to punch a wall in frustration. *even if* generative AI proves useful for tracking emissions and optimizing the power grid, that’s like a 1-2% efficiency gain. If AI consumes 10x the power of existing systems, and yields a marginal efficiency gain, THAT’S BAD. - ThreadSky

davekarpf.bsky.social • 263 days ago

This passage makes me want to punch a wall in frustration.

*even if* generative AI proves useful for tracking emissions and optimizing the power grid, that’s like a 1-2% efficiency gain.

If AI consumes 10x the power of existing systems, and yields a marginal efficiency gain, THAT’S BAD.

Comments

coloneloolong.bsky.social•263 days ago

Make Jevons Paradox Look Good Challenge

nparikh.org•263 days ago

The other problem is that it’s not obvious generative AI is the most useful AI for this. There are many other tools that aren’t power hungry.

nparikh.org•262 days ago

A lot of AI doesn’t work via gigantic models, but rather tiny specialized models that can run even in embedded kind of settings.

gothlorien.bsky.social•263 days ago

These same jackals were making a nearly identical argument about bitcoin—that it would incentivize renewable energy production—for years.

davekarpf.bsky.social•263 days ago

Yuuuup.

apearlma.bsky.social•263 days ago

AI tracking emissions: "detect a lot of emissions happening at data centers used to calculate emissions"

verdverm.com•262 days ago

you cannot improve what you don't measure, but if you do, you can do things like this

https://research.google/pubs/autopilot-workload-autoscaling-at-google-scale/

sharphall.org•263 days ago

Similar vibes

davekarpf.bsky.social•263 days ago

It’s like when the crypto-boomers insisted that their energy-guzzling mining operations were AKSHUALLY good for renewables, because they were creating so much additional demand, spurring more development in the renewable sector.

SHUT UP YOU FUCKING HACKS HOW DOES ANYONE YAKE YOU SERIOUSLY?

dextrosity.bsky.social•262 days ago

You don’t need to head over to the physics department on campus to know that perpetual motion isn’t a thing.

pfessenbecker.bsky.social•258 days ago

I get annoyed too at the implication that we need AI to make intelligent electricity usage predictions, as if no electrical engineer has tried predicting demand before.

brian.gawalt.com•263 days ago

I say gently, 1%-2% is an understatement. Dynamic line rating is not a "gen AI"* problem, but it is an ML+CV problem, and the upside is 40% capacity boosts. https://www.volts.wtf/p/getting-more-out-of-the-grid-weve

* (do not, under any circumstances, have to hand it to them)

davekarpf.bsky.social•263 days ago

Okay those efficiency gains would be awesome. And/but, the delta between gen AI and other machine learning probably amounts to only a chunk of that, no? (I think that’s what you’re saying anyway. Multitasking this thought while running errands.)

r.v.cx•262 days ago

Every other “AI” technique I’m familiar with is much, much better at this than LLMs. It’s right up there with “the solution to the housing crisis is advances in 3d printing technology”.

brian.gawalt.com•262 days ago

I will cosign that the relevant problem right now is "we're not looking at the line conditions in real time, at all" and not "we should be using trillion-parameter vision models to monitor the lines." The returns will diminish, to also cosign Dave's hunch.

brian.gawalt.com•262 days ago

The best case I can make for Open AI to contributing: the current monitoring tools are under installed, and the volts interview quotes a cost of ~$250k a pop for the LIDARs. "We can get that down to $10K per line if we combine ten normal cameras plus a trillion-parameter image model" is plausible.

r.v.cx•262 days ago

I know very little about the problem domain, but I am *extremely* skeptical that that spending $100M training a general-purpose model is at all appropriate for this kind of image processing. Even granting that image processing is a bottleneck, which 🤷.

verdverm.com•262 days ago

It can easily be offset by things like this: https://research.google/blog/heuristics-on-the-high-seas-mathematical-optimization-for-cargo-ships/

What you are effectively saying is that we shouldn't use computers to help us make things better because they use more energy tomorrow than they do today

Comments

Posting Rules

Reply