Hacker News new | ask | show | jobs
by cmiles74 2003 days ago
Sorry, I'm having trouble following this comment. In a public institution (as well as at Google) research papers are subject to peer review before publication. Now Google is adding _another review process_ after peer review, I don't believe this is something that would happen at a public university.

Do you have an example of a public university censoring research papers that passed peer review because that university believed the paper cast "an innacurate negative light"... on what, exactly?
1 comments
free_rms 2003 days ago
The tone of parent comment was "Google should tolerate research that's critical of Google", which I'm sympathetic to.

My rejoinder was that if it's inaccurately critical of Google, like hyping carbon impact without mentioning a decades-long carbon mitigation program, I get a lot more sympathetic to Google's position. Why should they pay someone to spread falsehoods about them?

I don't really care about the process minutiae.

link
joshuamorton 2003 days ago
Does it make sense to mention Google's carbon mitigation program in a paper not specifically about Google?

"Large language models can have an outsized carbon emissions impact, but not Google's, those are carbon neutral".

link
free_rms 2003 days ago
They'd probably want to finess it a little bit more than that, but yeah?

If you're going to bring carbon emissions into it, mentioning current mitigations as well seems relevant unless you're writing an op/ed piece.

link
joshuamorton 2003 days ago
> mentioning current mitigations

The "mitigation" in question is buying carbon offsets (I mean there are improvements in DC efficiency also, but those only do so much, and language models ballooning 100x isn't going to be fixed with 10 or 50% efficiency improvements). For the moment "carbon neutrality" is only achieved through the purchase of energy offsets.

That doesn't mitigate. It offsets. Don't get me wrong, still better than nothing, but its not a mitigation.

link
free_rms 2003 days ago
There's also

* this generation of language models leaning into transfer learning reducing the total number of training runs for different applications

* TPUs being more power efficient than GPUs (the numbers they used in the paper were based on GPUs)

* other energy-centric stuff that's not just offsets, efficiency like you mention in addition to sourcing from renewable

link