Interesting Links for 14-04-2026

[andrewducker]

1. How much co2 does training AI models produce?

(tags:ai co2 )

2. Consuming artificial sweeteners may raise diabetes risk for your children and grandchildren by modifying gene expression (in mice)

(tags:diabetes epigenetics taste )

3. Iconic Edinburgh city centre toy shop "Wonderland Models" set to close after plans to turn shop into two restaurants approved

(tags:toys Edinburgh shopping )

4. Birds do not, in fact, hit wind turbines

(tags:windpower birds GoodNews )

5. 'Bloodborne' Video Game Getting R-Rated Animated Movie Adaptation

(tags:movies games animation )

Comments

[calimac]

Trump didn't say that birds hit wind turbines. He said they hit windmills.

No, seriously, that'd probably be his defense if he's called on it.

[coth]

Although very small percentages of very large numbers can still be taken out of context and expressed as large numbers.

[nancylebov]

1. There's little agitation to get rid of gasohol, even though that's pure waste in the US.

#1

[mellowtigger]

That's a surprisingly low amount of CO2 for training a system on the sum total of human knowledge. Implementing those models repeatedly at scale is, I'm sure, where the main energy/environmental cost accrues. We can hope that AI compute centers switch to renewables sooner rather than later.

Re: #1

[andrewducker]

Yeah, the sooner we move everything to renewables the better!

Re: #1

[zz]

yeah, training the model might be a big lump of resource use, but running the model for millions of users every day probably eclipses that.

[channelpenguin]

1. the water use for cooling for data centres eclipses all as an environmental cost.

[bens_dad]

As long as the water is returned to the river and is clean the only environmental cost is the heat, which is significant, but probably doesn't eclipses everything else.

If you used local solar and wind to power the data centre you have only moved the heat from the air and the ground to the water. It might even be possible to use a water-source heat pump to undo that - if you even need the pump - I think the heat would be going "down-hill".

This is not (or should not be) like growing crops which take the water out of the river and put it into the air or the product which is then shipped across the world.

[channelpenguin]

Warmer water can be a big environmental problem (algal blooms, negative effects on water life and diversity).

From what I know of plant biology (uni WAS 30 years ago, mind you) most of the water used goes right been into the air, as plants use it a lot for cooling as well as building tissue. You can see that with a small pot plant on a sunny day (as I'm just seeing right now with mine). Water in the air comes back down as rain, that's absolutely normal. I find it more than a bit odd to see that as "waste". The availability of water to houses and farms has noticeably reduced in the areas surrounding centres (at least in the US), plus the costs have risen. Sorry if I think we need to eat and drink much more urgently than talking crap with your fave LLM or generating AI slop videos. We absolutely should stop using drinking water for toilets though. (Or any water. There are excellent composting solution but that would take time and support services to get people used to)

[jducoeur]

I sort of feel like (1) could be summarized as, "Amortization is a thing -- learn it!"

That said, it also kinda feels like yesterday's news: precisely because of how amortization works, the current topic is less about the cost of training, and more about the cost of inference. The former is largely a one-and-done, which amortizes nicely; the latter scales as O(n) with the number of users.

(Even inference doesn't appear to be necessarily bad, when you compare it with the energy cost of people doing the same things themselves, but it's a more interesting question IMO.)

As for (2), I'm kind of amused by the implication of the headline, that modifying gene expression in mice can raise diabetes risk in the grandchildren I don't have...

[andrewducker]

We are giving your children diabetes in alternative universes.

And the thing about training is that it's ongoing. Nobody is stopping training, because they all want their next models to be better than everyone else's next models. OpenAI and Anthropic both plan to make losses for the next few years because all of their money is going into training.