There is a style to how these LLMs write. I have seen enough now that I can tell almost immediately. The main tell of something LLM generated is how vacuous it tends to read.

For these kind of things, you need a letter/statement that matches the fact of your application.

Now if anyone was just banning phrases, that would be stupid. Part of the reason some terms show up is that some sub community use them a lot. For decades now, about 50% of the cold emails I would get from the Indian region was using "esteemed faculty"

I don't know why, but I fond MINI's vehicle pretty heavy for their size.
Haven't looked recently, but when I bought my 4 door, I think it was the heaviest of its category.

I don't know that SAT measures merit.

My son is running an SAT prep for high school students. And I can guarantee you that anyone who can afford the expensive tutors will get about 100 points on the SAT higher than those who can't, even if they study for the same amount of time.

So yeah, to get a high SAT score, a large component is whether you are an idiot or not, whether you study or not, and whether you pay attention or not.
But in lots of ways it also measures whether your parents could afford tutors, whether you were born in a native english speaking family, whether you are room at home to study or have access to a decent quiet space, whether you need to work two shifts on the week end to help parents pay rent.

I am glad they decided to use the SAT, at $LOCAL_UNIVERSITY use the SAT too. And in graduate admission, we use the GRE for foreign applicants. They really give you an idea of what is going on and let you discard scores that are just too low. But I wouldn't sort by SAT and take top-k; that's stupid! That just selects people in ideal conditions and people who managed to cheat.

But I don't think the use case here is to review with AI when there is an allegation of misconduct.
I think the purpose here is to automatically review the thousands of hours of footage that $LOCAL_PD produces so that they can audit their own officers and identify bad conduct that did not raise to a complain before the officer goes postal.

Ideally that would have been better.

But it is possible that it took them that amount of time to verify that it was feasible and made sense policy wise.

I work for $LOCALUNIVERSITY and sometimes decision can take time because at the scale of ~50,000 people even a simple decision can have deep consequences.

Here we are talking billions and nuclear power plant. I would not be surprised if it took them a year to do understand the problem, understand the impact, see a possible solution and do the study by running numbers and scenarios.

Going car free can work depending on where you live. I grew up in the western suburbs of Paris. I didn't learn to drive until I was 25. I was living with my parents about 10 km from the Paris city limit. Paris is a decently provisioned city in term of public transportation. Yet, most families had 1 or 2 cars.

While I agree that there is a need to reduce car dependencies, especially in the US (where I now live), it is unlikely we can get rid of cars for much more than half the adult population until we massively reorganize our global lives.

And the only way we can realistically get there is by densifying suburbs into small towns. At that scale, we won't be able to charge in a garage, so we will need "public" charging infrastructure which is what the article linked is essentially talking about.

Now, I agree we should include externalities in costs where it makes sense.

I imagine you refer to Steamdecks. Steamdeck are more expensive, heavier, bulkier, with less batteries.
Nintendo has great games and a fair number of licenses that I want to play and that won't release on other systems.

Not that I am saying that steamdeck users are idiot or anything. If you have a huge steam collection, that probably can make sense. But I'll probably be getting a Switch 2 on release!

At the scale they seem to need to do it, it is not clear that hardware/software gain efficiency are the whole solution. Don't get me wrong we should do it. But our current ML stacks are already using our hardware decently well. That's the point where I don't expect we can really gain more than a factor 100 anymore.
And based on the type of usage that they believe we'll get, we more likely need to gain factors in energy consumption in the order of a million.

And training is not that irrelevant. If you look at the cost of training the final model that gets deployed. Yes, that's pretty negligible. But to have an honest metric, to train one model, you typically tried thousands, maybe millions of variants on hyper parameters. And that definitely takes time and power.

Even though I agree that eventually, commercially, the cost of running inference will have to be much smaller than the (full) cost of training, simply to be commercially viable.

I wonder whether they are used for cheating at chess!

Your own link shows that 40% of electricity production is renewable and nuclear. So, maybe your point isn't as valid as you think.

Also, 40% seem to be natural gas, isn't burning natural gas much better than petroleum?

Also, aren't efficiencies in energy production higher for a plant than for small mobile engines? Also, isn't it easier to filter/recapture emissions at a plant than in each individual car?

I write academic papers all the time. I have nothing to hide. Please run plagiarism detection of all dissertation and published papers. I'm sure there are a few frauds. Let's out them! Academia would be better without them.

Now, there is plagiarism and plagiarism. I really don't give a shit about "oh, they forget the quote around the passage they lifted even though it is cited." or the "they defined their problem by copying the standard definition of the problem". It might be bad form, but no one deeply cares about that.

I do care about papers lifting results from other papers though. Or I do care about doctored results.

Plagiarism checkers are a bit better now though. The one my grad school uses knows to exclude the reference pages for instance.

Obviously you shouldn't.
But in a dissertation you expect models to be set, and related works to be discussed. My dissertation has definition of classical combinatorial problems for completeness. I wrote them myself, but frankly a copy pasting of wikipedia wouldn't be much different.

people have been trying to convince me that we should only focys on studying quantum computers casue they are the future.
That was 2005. Glad I took the stance "I'll care once I can ssh on a useful one"

AFAIU quantum machines don't do exponential calculation. They are operating along a different model than regular machines. And as such they can do things that the other model can't do. (And vice versa)

It's a bit like saying that you can't go very far on scissors, but you can on a wheel.