niplav @niplav@schelling.pt

@chjara

@pixelherodev@fosstodon.org @drewdevault@fosstodon.org @dym
Makes sense. (I disagree, but I understand the sentiment)

@pixelherodev@fosstodon.org @drewdevault@fosstodon.org @dym Fair enough. Direct carbon capture is ~10x more expensive (as the MIT article claims, ~$1200/ton). This would indeed make it more difficult to pay the externality.

@pixelherodev@fosstodon.org @drewdevault@fosstodon.org @dym

So they might be able to pay off the training externalities in ~2 years, more if they also pay off the cost of running the model.

@pixelherodev@fosstodon.org @drewdevault@fosstodon.org @dym If you spend 1% of that on carbon offsetting you can purchase carbon offsets for $24k, which amounts to ~250 tons of carbon (at ~$100 for capturing a ton of carbon[1], using things like preventing rainforest deforestation, and this report[2] says it's more like $10). This page[3] estimates that training chatGPT produced ~550 tons of carbon.

[1]: https://climate.mit.edu/ask-mit/what-most-cost-effective-way-buy-carbon-offsets
[2]: https://founderspledge.com/research/fp-climate-change, p. 83
[3]: https://towardsdatascience.com/the-carbon-footprint-of-chatgpt-66932314627d

@pixelherodev@fosstodon.org @drewdevault@fosstodon.org @dym

I agree that instituting a carbon tax that captures the true externalities cost of the training & running would probably increase the cost to users, though I'd venture a guess that this would not dissuade OpenAI from training & running the system at the given demand.

Let's see. 100mio. customers. Let's say 1% of those pays for plus. $20/month for a year is $2.4 mio./year.

@drewdevault@fosstodon.org @pixelherodev@fosstodon.org @dym
That is a claim one could make (to that strength) of many new technologies (e.g.: microwaves, bicycles, internet wikis).

Also, I think the users, if asked, would claim they are now doing better.

@pixelherodev@fosstodon.org @drewdevault@fosstodon.org @dym
Hm. Would you claim that the people using GPT models right now *don't* need them?