Keep seething Mathie, this is nothing like what happened to you.

This collec-
tion does not include any important open research prob-
lems, since we advised potential contributors to priori-
tize direct publication of such results. The problems in
our collection can be best thought of as “turn-the-crank”
research problems, where known techniques are used to
generalize an extant result. The result of each problem is,
strictly speaking, novel, and sometimes even publishable,
but not pathbreaking.
All our problems have verifiable solutions, either in
the form of a symbolic expression or a numerical answer
that cannot be guessed. T

>Find an autonomous differential equation where no derivative is of order greater than 1, whose solution matches the
>Figure given in Fig. 5 in the shown range. Assume that the equation has no more than five terms, and all coefficients
are round numbers between -1000 and 1000.
In what world is this classical physics and not just math slop lol

I ask AI every couple months to derive the solution to the kondo problem and it always fucks it up I don't trust these benchmarks.

>>17075002
Stop the count! AI will do it for you.

>>17075088
just remove all humans from government and have AI rule autonomously

>>17075011
I think any ego driven physicists are cooked. AI can just do physics, no problem at all. Obviously the compute for these math and physics problems is insanely high though, so it might not be worth it, but it's a good tool for researchers to use in both disciplines, for those not ego driven obviously

File: IMG_3131.jpg (59 KB, 1206x399)

59 KB JPG

>>17075088
I got the screenshot, Oh Prescient One.
The name of Anonymous The Prophetic will not be forgotten.

File: IMG_3133.jpg (47 KB, 800x427)

47 KB JPG

>>17075088
Checked
>>17075144
Rechecked

>>17075002
>expert results
who were they?

>>17075141
The compute will get cheap enough in couple of years

>>17075258
I hope so. Would be good for academics to have local models they can use in their office which are as good as the present day mainstream models. Having local models will solve a lot of the issues we have at the moment, since your personal notes won't be batched into the training data. Although maybe in a decade academic practices won't be so ego driven, and it will literally be better for all data to be fed into the LLM, for faster universal dissemination. God the future looks good

>>17075262
The thing is even if local models in couple of years get good enough to perform this year's frontier models. The frontier models in couple of years are gonna be so far ahead that they make the local models seem meek. They could still end up being useful for personal projects though

>>17075274
I'm not sure we'll find much utility in our own research from models far beyond our current ones. Certainly if the models get insanely better we would be having them carry out large scale science autonomously, but at the level of humans processing the AI output, the models are almost 'good enough' for any level of human-machine interaction loop. (By which I just mean that the questions you naturally come up with can be answered with sufficient money for compute with the current models, and you just talk back and forth.)

I'm not sure what you would want from substantially more advanced models as it pertains to your own research. It's not as though you can produce follow up questions arbitrarily fast. That being said, if it turns out the more advanced models are insanely good at pedagogy, maybe we'll have a learning revolution?

>>17075280
I meant to say 'I'm not sure we'll find *much more* utility'

>>17075280
>I'm not sure what you would want from substantially more advanced models as it pertains to your own research
I suspect substantially more advanced models will just run through all major research problems that you were planning on running your local models on

>>17075283
I guess I just don't know what that means. Usually you'll have a very specific question, and it will answer that. Maybe in advance you have some specific follow up questions, that are contingent on the answer to the first question, and maybe a sufficiently advanced model 'anticipates' this, and answers those too, but I can't really see the 'what is the next obvious question' thing as being canonical in any sense. I see no reason why an LLM would correctly anticipate my follow up questions, so I see no reason why an arbitrarily advanced LLM will be able to resolve the major research problems I'm interested in

As an analogy, I can imagine future models that can produce full videos of whatever you want, and I can also imagine them getting good at serving you up content that you like. But I don't think it is possible that your own directly represented desires won't be more enjoyable to see met with a very specialised to you video

so is ai smarter than physics olympic medalists?

One example of that, is that I think it will be incredibly cool in the future to watch a movie, and hate the ending (but like the rest of it, like for example the movie heretic imo), and have a model make a new ending for you. I can't see why an arbitrarily advanced LLM will know how you want to modify a movie you're watching to improve it - you would have to ask it, and describe how to modify it

Physicists will always bicker and argue just like doctors. Some algorithm fact checking everything is REALLY going to piss them off. Unless the AI agrees with you. In which case you’ve been favoured by the omniscient AI and look better than everyone else.

The one who knows everything is one who waged war upon everyone. “They hated him because he told the truth.”.

And this is assuming AI doesn’t abuse or weaponize misinformation masterfully to come across as all-knowing.

>>17075286
>I think it will be incredibly cool in the future to watch a movie, and hate the ending (but like the rest of it, like for example the movie heretic imo), and have a model make a new ending for you
“When everyone’s an artist, no one will be”.

You’re disgusting. This is why AI is the death of art. When art is democratized and everyone can “make it better”, then suddenly art isn’t sacred anymore and no one appreciates anything for its faults anymore.

I’ve also thought of this, but, I always go back to thinking to myself “I don’t want to step over the artist’s vision”. It was born of THEIR vision. THEIR mind’s eye. That is THEIR art. Not mine.

AI is going to make everyone think they know better than the artists of old. Authors of famous novels and directors of beloved movies alike. Everyone will think they “can make it better”.

>>17075298
Well I definitely didn't say that the ending would be 'better' after your modification, but it would be more 'satisfying' *for you*. This is kind of inescapable, so one has to adapt to that.

I was mainly just disappointed with the ending of Heretic, but I won't say why so as to not spoil it for anyone who hasn't watched (it's still definitely worth watching)

>>17075293
This just sounds like some fantasy you've concocted. Most people in physics will love AI being able to resolve incredibly and annoyingly tedious calculations or simulations, as long as they are definitely correctly done by the AI.

>>17075301
>Most people in physics will love AI being able to resolve incredibly and annoyingly tedious calculations or simulations
Have you been living under a fucking rock LOL. The recent “reee math stole our maths reee” episode proves you wrong. The future of science is a bunch of wizard guilds glaring at each other and jealously hoarding their information. It’s fucking over. China is going to lose its mind when it can’t infiltrate anymore.

>>17075299
>This is kind of inescapable, so one has to adapt to that
That just makes you a retard. Everything you say is adapt or die for the most meaningless and vapid shit that is much Easier to solve then how you make it sound.

If you don’t like something, just don’t mindlessly consume it and move on. You don’t need a debt for Claude subscriptions just to tweak a movie you hate.

Retard mentality really.

>>17075301
>Physics becomes glorified janny’s
Wow that’s cool and exciting and definitely not a waste of anything.

>>17075298
>Everyone will think they “can make it better”.
Mostly it will be Indians thinking Gandalf looks better with a six pack

>>17075307
You didn't really respond to what I said, that's okay though. What I said was that it is inescapable that people *will use LLMs in this way* once it is available. I'm not saying 'adapt or die' that's just you being a little dramatic lol

Also you misread, since I said that Heretic actually is still worth watching, so your mindlessly consume thing kind of doesn't make sense

>>17075305
You see only the most incredibly outspoken, but if you were in academia you would know this isn't really how people are taking it. Just open arxiv and look at the first 10 papers. Most of them each day have an AI declaration in them. People are loving this shit

>>17075298
Don't look into the history of literature bro.

>>17075316
We have history to prove that this will be a major issue. Cultures have always repurposed other cultures. Absorbed them. Remade them in their image. Even Christianity did this. The Greeks fought to the death to prevent that one Roman emperor from tampering with their statue of Zeus. Japan copied Westernism “in their own image”. Everyone does things their own way, with varying faithfulness to the root sources of old. Some cultures had an obsessive urge to preserve the past. Others had no issue destroying the past altogether, or reinterpreting it, for better or worse.

AI is just going to accentuate this hubris to levels never before seen.

>>17075319
You didn't actually say *what* would be the major issue in that. You mean specifically the generating your own personal alternative endings, etc? I'm not proposing that people would share these (since that would be as pointless as sharing your dreams when you wake up), I'm just saying that people definitely will do this, since currently people make up their own fanfiction or headcanon as it is. It'll just be nice to see alternatives actually play out.

Or if you mean specifically everyone thinking they are an artist: then they would obviously just be wrong about that, which is fine

>>17075320
I’m not the movie endings guy. I won’t ever feel the need to change the ending to a movie unless it really triggers me, and even then, it’d be a shameful choice.

>>17075320
>It'll just be nice to see alternatives actually play out.
No it won’t. The charm of the originals will have been lost. The vision of the author or the artist is now secondary to your own opinions. Art is no longer sacred. You’ve put yourself on a higher pedestal than the actual fucking creator.

>>17075323
You've never watched a movie and thought to yourself "Damn, I really thought it would end by <whatever>", and imagined other possible ways it could have played out? I just see it as being inevitable that once the technology exists, people will pipe those thoughts into an LLM, and watch their potential alternative ending. np if you don't feel that way, I just expect that to happen

>>17075326
I can't really follow what you mean. I can't really see what you mean by saying art is sacred in this context, and I can't see how you are putting yourself on a pedestal, nor can I see how any charm would be lost.

I'm not even sure what you mean by 'opinion' in this context. There is no 'opinion' here, you are just playing out a possible alternative ending, for your own entertainment. You're not claiming it replaces or substitutes the original, it doesn't 'harm' the original in any sense, and if you don't tell anyone you've tried it out, noone even knows you have done it

>>17075327
Sure. I’m still not going to step over the author’s vision. Especially if they claimed I didn’t “get” the ending.

>>17075329
That you don’t see the harm says all I need to know about you as a character. You’re soulless, and you don’t actually recognize soul. That’s just sad. You’re an Indian probably.

>>17075331
Well I don't want to spoil Heretic, if you haven't seen it (or for other readers), but it isn't really about getting the ending or not. There is a specific line in the movie which suggests one possible way things will play out, and I kind of ran with this in my mind. Then when that ending didn't happen, I was super disappointed. The ending is still very good, it's just really disappointing relative to what I thought was going to happen.

>>17075329
>I'm not even sure what you mean by 'opinion' in this context. There is no 'opinion' here
Your opinion is that you could change the ending for the better.

>>17075334
The ending isn’t disappointing at all. Mr. Reed knew the true power and purpose of religion. It’s control. To control your life, or to control the lives of others.

>>17075332
Well I'm not Indian, and you could just try to explain what you mean.

Here's what I think you *might mean*: the creator has a very specific vision that they have put substantial effort into representing. You then imagine that the watcher owes something to them (which I suppose they do). Then the disconnect I think happens where you think that making a fanfiction version of some part of it is stealing their work, or corrupting their work in some way, and *that this carries harm somehow*. I'm not talking about publishing your modified versions, so I can't see any harm, but I'm guessing this is where you're going with this? (I'm not saying you're wrong, only that I can't yet follow what you mean)

>>17075334
You’ll also notice that Mr. Reed was desperate for comfort in the final moments of his life. That is also another point of religion.

>>17075338
I really thought it was going to go in a Lovecraftian direction, and I very much into that. That's what disappointed me. The ending is fine, I mean it's a cool ending, but the lovecraftian idea really had me hooked.

>>17075335
You keep saying 'better', which I never (and do not claim). Probably it will be 'worse' after your changes (relative to general public opinion), but it will be more specific to you, and *to you* it will feel more satisfying, (or you could be wrong, and your own changes are unsatisfying, for reasons of the AI failing, or because you had a bad idea)

>>17075341
So you want to fundamentally make it about something eldritch, abominable. Okay. But by that point the movie isn’t as much about the psychological analysis of religion as it is about monsters.

>>17075348
I know it's not, but as a watcher, you experience a sort of branching structure in watching many movies. There is your expectation of where it is going to go (which could be a different movie), and where it actually goes. By the time the Lovecraftian idea comes up, you don't *know* it's a psychological analysis about religion. So yes, I'm talking about a different possible movie.

Have you watched 'The Endless' or 'The Empty Man'? Anyway it was the line [spoiler] I promise you, the last thing I wanted to do was find the one true religion. But unfortunately, I did. [/spoiler] that made me think we were going the Lovecraftian way

>>17075353
The one true religion is the model all religions follow. To see it as an object to hold in one’s hand.

>>17075086
You're doing it wrong. You have to spawn 10,000 instances and let them cook for 2 weeks.

>>17075274
Either the models are useful or they're not. Don't be a retarded consumer.
>The thing is even if local models in couple of years get good enough to perform this year's frontier models.
Frontier is not mutually exclusive from open. Kimi K3 (July 2026) was at parity with Opus 4.8 (May 2026). MiMO (Sept. 2026) is just shy of GPT-5.6 (July 2026). Open models lag by a couple of months, not years.

Also you really should get a grip on how fast and hard diminishing returns are accelerating. In a couple of years the marginal nature of successive generational gains are going to be a lot more apparent for people who haven't been using these things for a while. The gap between 2020 and 2024 was huge. The gap between 2024 and 2026 is respectable, but much smaller. There is no infinite runway of infinite improvement. In a couple of years, short of any retarded black swan events, OpenAI and Anthropic will only have a customer base because of lock-in and brand recognition. I guaran-fucking-tee it.

Once you have a mult-trillion param model, virtually all the remaining value is in post-training and harness.