I have been somehow repeatedly accidentally triggering the summarize this tab feature. It's driving me slightly insane.

You could try talking to them? Why you are presenting the options as
1. Yell at them about AI
2. Let them use good AI
Have conversation with them about harms of dangers of AI. Treat them with respect. They can and will access AI from evil or careless companies even if good alternatives exist.

People who know their results are much more likely to get published above some threshold have a motive for performing manipulations to their studies to achieve it. Doesn't always lead to outright data fraud, but there are many dishonest ways tune your questions and analyses to achieve a threshold.

Wrong to expect a normal distribution, but the extent to which > +/-2 is acting as a publication threshold does strike me as concerning. The main problem isn't what gets filtered out by publication bias but rather motivating effect awareness of this threshold creates among scientists.

My mom was excitedly telling me about 3I/Atlas a few weeks ago but unfortunately she had encountered the news in form of a bot-slop youtube video someone had posted to facebook. A whole diseased media ecosystem is eager to take advantage of people's natural interest in these events.

Really frustrating the people like that are part of many people's experience of learning about the genuinely exciting events of humanity first few observations of interstellar objects. Loeb is unfortunately just one part of the whole ecosystem of people exploiting this news to grab attention.

Awesome!

By presenting only the median energy use per prompt and not revealing mean energy use per prompt (or any other statistics about the distribution) they can hide where most of energy is being used. It doesn't matter is 90% of prompts are relatively cheap if the other 10% are using most of energy.

And yes, the most important bit to understand about this picture is that companies building these models have no intention of slowing the growth in their energy demand.

The acccelerators (TPUs/GPUs) AI data center use aren't designed for traditional compute workloads so I don't think they are often used for other stuff. The models are becoming more efficient, but this paper uses metrics that intentionally distort how energy intensive things currently are.

Or maybe they can already see that their models' aggregate energy use is being driven by prompts that use vastly more energy than median prompt. They use median as their aggregate stat don't even give the mean which suggests an intentional omission to me given how easy it would be compute.

That was my plan originally on seeing these comments, but eventually I decided a respectful attempt to disagree was probably wasted. You say people who share my perspective have been “sold a lobotomy.” Doesn’t sound like you’d have much curiosity about why I think you’re wrong.

Wasn’t offering one. Please continue enjoying your smug.

Yeah this is hot garbage take. Unfollowing.

I see a hummingbird in the first two seconds.

Would nice to have a government that is better at delivering, but that doesn’t mean the abundance agenda is a winning political message in a time when authoritarian overreach is the bigger threat than government incompetence.