Yes! The way those physics models are created is so cool. The article somewhat explains it, but it's mostly a fluff-piece for things unrelated to genAI. More in-depth:
The physically accurate simulation is great but slow. So we can create a neural network (there's a huge variety in shapes), and give it an example of physics, and tell it to make a guess as to what it'll look like in, say, 1ms. We make it improve at this billions of times, and eventually it becomes "good enough" in most cases. By doing those 1ms steps in a loop, we get a full simulation. Because we chose the shape of it, we can pick a shape that's quite fast to compute, and now we have a less-accurate but faster simulation.
The really cool thing is that sometimes, these models are better than the more expensive physics simulation, probably because real physics is logical and logical things are easier to learn.
We've done things like this for ages. One way we can improve them is by giving them multiple time steps. Unfortunately they kinda suck at seeing connections over time, so this is expensive. Luckily, transformers were invented! This is a neural network shape that is really good at seeing connections over one dimension, like time, while still being pretty cheap and really easy to do run in parallel (which is how you can go fast nowadays).
With a bunch of extra wiring, transformers also become GPT, i.e. text-based AIs. That's why they suddenly got way better; they went from being able to see connections with words maybe 3-4 steps back, to recently a literal million. This is basically the only relationship with "AI" this has.
I think these are probably the main bits they wanted to remove (both Article I section 9):
The Privilege of the Writ of Habeas Corpus shall not be suspended, unless when in Cases of Rebellion or Invasion the public Safety may require it.
No Bill of Attainder or ex post facto Law shall be passed.
No Money shall be drawn from the Treasury, but in Consequence of Appropriations made by Law; and a regular Statement and Account of the Receipts and Expenditures of all public Money shall be published from time to time.
Pure speculation to be clear, but they removed only the end of section 8, and probably removed section 10 because it would have to be numbered differently. Habeas Corpus has been causing them trouble, they've been stealing money pretty blatantly already, and bill of attainder/ex post facto laws seem convenient to them.
Having just read the problem, I'm curious how o3 solved it (and the human too tbh). My experience with LLMs says they'd be absolute complete crap at this, it's a very hard and open-ended problem. Intuitively I'd say it would just end up doing random changes tryjng to improve its score.
I think I could write the "trivial" solution but anything beyond seems... difficult. Congrats to the winner!
They can, but it's not trivial. The challenge uses a bunch of modern browser features that these scrapers don't use, regarding metadata and compression and a few other things. Things that are annoying to implement and not worth the effort. Check the recent discussion on lobste.rs if you're interested in the exact details.
It's (genuinely) interesting that big-time AI enthusiasts always assume that if you dislike AI, it's because you don't know how to use it (and not because they have different priorities or have conceptual issues with it or such). It was the same thing with crypto and NFTs.
I know a good amount of very skilled developers who tried AI several times in different contexts, and always found the tradeoffs much too bad to be worth it (in terms of accuracy, precision, cost, speed, whatever), and yet they still get "oh you just haven't tried my favorite AI tool".
Fair enough! The disadvantage is that, as opposed to Dropbox and similar, I have go into a file at the root of the synced folder, rather than keeping that config near to where itcs relevant.
I tried with both, but I didn't figure out how if such an option exists. I did manage to do the opposite (keeping files uploaded but not having them locally), both with and without VFS (with VFS it's in a context menu in nautilus, without it's in the desktop app).
I have. It hasn't worked very well for me, the docs weren't great (though I'm looking at them now and they do seem better?) and it broke in strange ways.
Wait, so OpenAI's whole kerfuffle here is based on nothing directly stated (e.g. in the paper like I thought), and worse, almost certainly completely unfounded?
Wow just when I thought they couldn't get more ridiculous...
Yes! The way those physics models are created is so cool. The article somewhat explains it, but it's mostly a fluff-piece for things unrelated to genAI. More in-depth:
The physically accurate simulation is great but slow. So we can create a neural network (there's a huge variety in shapes), and give it an example of physics, and tell it to make a guess as to what it'll look like in, say, 1ms. We make it improve at this billions of times, and eventually it becomes "good enough" in most cases. By doing those 1ms steps in a loop, we get a full simulation. Because we chose the shape of it, we can pick a shape that's quite fast to compute, and now we have a less-accurate but faster simulation.
The really cool thing is that sometimes, these models are better than the more expensive physics simulation, probably because real physics is logical and logical things are easier to learn.
We've done things like this for ages. One way we can improve them is by giving them multiple time steps. Unfortunately they kinda suck at seeing connections over time, so this is expensive. Luckily, transformers were invented! This is a neural network shape that is really good at seeing connections over one dimension, like time, while still being pretty cheap and really easy to do run in parallel (which is how you can go fast nowadays).
With a bunch of extra wiring, transformers also become GPT, i.e. text-based AIs. That's why they suddenly got way better; they went from being able to see connections with words maybe 3-4 steps back, to recently a literal million. This is basically the only relationship with "AI" this has.