gradient descent into madness

Reinforcement learning guy, #1 mirror descent enjoyer, game design enthusiast, and haver of many bad opinions

[bayesic-bitch]

I have got to get an RLVF job stat

[prokopetz]

Soulslike boss fight that riffs on the Black Knight sketch from Monty Python's Quest for the Holy Grail where every time you hit a new phase you hack off a limb, and somehow it just keeps getting harder. Most guides recommend cheesing burst DPS to skip the hopping-on-one-leg phase because it's just plain unfair.

[eka-mark]

This week, I get to show somebody one of my favorite math book figures, illustrating Liouville's theorem in Hamiltonian mechanics:

[sequred]

ime pure sociopaths get put out of doors pretty quickly, the people who float to the top are people with very active yet very mercenary emotional lives, thats the new meta

[sequred]

once in a while you see someone whos overcome with pity and horror-at-all-pain the moment their friend is getting shit, whos swamped with tics and trembles whenever they need to look weak, who relaxes into smooth /b/tard cruelism whenever someone inconvenient needs to go. and never hesitates and never gets self-conscious and never feels their opponent threw a nice punch. and look as long as you like these feelings are all completely authentic. every chord of human experience is there but here the instrument's in tune

[yorickless]

blogger basically making the argument that the cultural revolution is what allowed china to later experience such meteoric economic growth. it's nice to see contrarian opinions like this. I've played enough victoria to be sympathetic to this idea that the destruction of the peasant is the ultimate good, but you know. none of this is horribly rigorous

[beesmygod]

HARMFUL GAY STEREOTYPES EMBODIED BY MOHG

lives in sewer

kidnaps children actually he was cleared of this one

blood magic

[hexdrake]

never stops being funny

[bayesic-bitch]

https://arxiv.org/html/2604.03071v1

I think we might actually be there for math formalization now. It's not to the point where it's super cheap to do whole textbooks, but we're getting to where if you have a specific theorem that you need formalized to check your paper, you can pull the supporting lemmas and get it formalized with a few days and a Claude subscription. And then you can have it formalize your paper from there

[bayesic-bitch]

It's insanely lucky that there were these two research areas, type theory for formal verification and reasoning AI models, each of which was of very limited usefulness to math research on its own, each happened to be reaching maturity at almost the exact same time. Math has such a high standard of rigor that can basically never just use an ai generated proof. And proof assistants like Lean are so incredibly arduous to work with that nobody wants to actually use them. But slap them together and you get a closed feedback loop where the model gets checked against an extremely rigorous formalism, and the model abstracts away enough that you don't have to learn homotopic type theory to use it. And you end up with something that is both easier and more rigorous than doing the proof fully by hand. And both of these went from deeply underpowered to pretty useable since 2023, just in time to be practical when paired together

[bayesic-bitch]

Update on this. I've been working on a bandit theory paper, and I was using Claude to typeset it in latex. When pinning down some of the lower bound arguments, it comes back and tells me that it found a log(t) improvement on the lower bound while writing it up. That's great, but I still don't have a good intuition for the bandit lower bound arguments in the first place, so while I can check it's proof, I don't feel 100% confident that I'll catch the more subtle errors it could have made.

So I set it up in a lean environment and tell it to formalize the argument. The only problem is that MathLib is ofc not at all focused on bandit theory. There's a GitHub repo which has like two very basic results in it, and that's it. No matter. Three days later, and it's formalized a bunch of information theory background theorems, Bretagnolle-Huber, the divergence decomposition lemma, the minimax lower bound, and all the lower bounds in my paper. I guess the argument goes through!

[not-so-impossible]

How easy is it to know that there aren't any errors in the lean formalization? I know that lean guards against logic errors but that doesn't prevent errors in problem statements or in others' theorem statements? Not at all my field so I'm curious

[bayesic-bitch]

Basically it just comes down to checking the theorem statement and the supporting definitions, which is pretty easy bc that's the part where Lean goes out of its way to make sure it looks the same as the mathematical statement. Its simple enough that as long as you ask the model to formalize the statement first and then follow the proof, the big models (ie, not aristotle) don't seem to make many errors. Might be different in something like geometry that requires a visual sense they're not good at.

[bayesic-bitch]

https://arxiv.org/html/2604.03071v1

I think we might actually be there for math formalization now. It's not to the point where it's super cheap to do whole textbooks, but we're getting to where if you have a specific theorem that you need formalized to check your paper, you can pull the supporting lemmas and get it formalized with a few days and a Claude subscription. And then you can have it formalize your paper from there

[bayesic-bitch]

It's insanely lucky that there were these two research areas, type theory for formal verification and reasoning AI models, each of which was of very limited usefulness to math research on its own, each happened to be reaching maturity at almost the exact same time. Math has such a high standard of rigor that can basically never just use an ai generated proof. And proof assistants like Lean are so incredibly arduous to work with that nobody wants to actually use them. But slap them together and you get a closed feedback loop where the model gets checked against an extremely rigorous formalism, and the model abstracts away enough that you don't have to learn homotopic type theory to use it. And you end up with something that is both easier and more rigorous than doing the proof fully by hand. And both of these went from deeply underpowered to pretty useable since 2023, just in time to be practical when paired together

[bayesic-bitch]

Update on this. I've been working on a bandit theory paper, and I was using Claude to typeset it in latex. When pinning down some of the lower bound arguments, it comes back and tells me that it found a log(t) improvement on the lower bound while writing it up. That's great, but I still don't have a good intuition for the bandit lower bound arguments in the first place, so while I can check it's proof, I don't feel 100% confident that I'll catch the more subtle errors it could have made.

So I set it up in a lean environment and tell it to formalize the argument. The only problem is that MathLib is ofc not at all focused on bandit theory. There's a GitHub repo which has like two very basic results in it, and that's it. No matter. Three days later, and it's formalized a bunch of information theory background theorems, Bretagnolle-Huber, the divergence decomposition lemma, the minimax lower bound, and all the lower bounds in my paper. I guess the argument goes through!

[trekinist]

> The Stalker says he can take me to the Zone

> I ask if the Zone is creepy or wet

> He doesn't understand

> I light a cigarette and make my speech about what constitutes something being creepy or wet

> He does not laugh and says "The Zone demands respect"

> it's creepy AND wet

[bayesic-bitch]

https://arxiv.org/html/2604.03071v1

I think we might actually be there for math formalization now. It's not to the point where it's super cheap to do whole textbooks, but we're getting to where if you have a specific theorem that you need formalized to check your paper, you can pull the supporting lemmas and get it formalized with a few days and a Claude subscription. And then you can have it formalize your paper from there

[bayesic-bitch]

It's insanely lucky that there were these two research areas, type theory for formal verification and reasoning AI models, each of which was of very limited usefulness to math research on its own, each happened to be reaching maturity at almost the exact same time. Math has such a high standard of rigor that can basically never just use an ai generated proof. And proof assistants like Lean are so incredibly arduous to work with that nobody wants to actually use them. But slap them together and you get a closed feedback loop where the model gets checked against an extremely rigorous formalism, and the model abstracts away enough that you don't have to learn homotopic type theory to use it. And you end up with something that is both easier and more rigorous than doing the proof fully by hand. And both of these went from deeply underpowered to pretty useable since 2023, just in time to be practical when paired together

[bayesic-bitch]

https://arxiv.org/html/2604.03071v1

I think we might actually be there for math formalization now. It's not to the point where it's super cheap to do whole textbooks, but we're getting to where if you have a specific theorem that you need formalized to check your paper, you can pull the supporting lemmas and get it formalized with a few days and a Claude subscription. And then you can have it formalize your paper from there

[luntfontanne]

damn he really is an all-time poster

[stumpyjoepete]

Duality in optimization theory seems like a magic trick in that way that math sometimes does, but the max-flow / min-cut duality specifically is *chefs kiss*.

The Soviets are trying to optimize shipping on their rail network and the US is trying to optimize potential plans to bomb that same Soviet rail network, and it turns out the problems are mathematically dual to each other. Ultimately both sides fail to do either of these things, and in the end, the math primarily benefits petty capitalists trying to maximize quarterly profits from production of gizmos.

God really went above and beyond in making the joke dialectical.

[bayesic-bitch]

> the math primarily benefits petty capitalists trying to maximize quarterly profits from production of gizmos

Which turns out to be dual to the communist strategy of maximizing production under linear constraints

[bayesic-bitch]

Was Nietzsche's writing good philosophy, both in the sense of producing technically correct arguments where the conclusions follow from the premises, and in the sense of "love of wisdom", of caring deeply about the truth and understanding it? no, he was an anti-rationalist and dismissive of the value of truth, and most of the concrete factual/philological claims he made in his work do not hold up to scrutiny. But did he have a lasting and positive influence on philosophy as a discipline? also no, the loudest part of his philosophical legacy was carried by hacks like Derrida and Deleuze. But did his writing help us build connections and make us better people? No, outside academia his greatest influence on the world is through people like Hitler and Ayn Rand.

But that's not what's really important when judging Nietzsche's success. It's not what Nietzsche himself cared about, and he should be judged on his own merits, not on the merits of writers, philosophers, and moralists he openly despised. What's really important is whether rejecting morality would give us the space to be healthy, vibrant, and self-affirming. Whether it would fill us with a zeal for life that spills out to everyone around us, let us face honestly up to hardship and make something beautiful out of suffering, leave us so in love with being alive that we'd have no regrets and want to do every last bit of it over again. Did his writing do that? did he himself live a life that he was proud of and embodied his ideals? lmao

[bayesic-bitch]

the problem with most "philosophical" fiction is that the skillset required for writing good fiction is almost completely orthogonal to the skillset required to write good + novel philosophy. so very frequently you get a decent-but-not-great story that just name drops sparknotes summaries of freud or camus and everyone acts like this makes it deep, despite having absolutely no original ideas of its own.

[discoursedrome]

I am informed that the Office Québécois de la Langue Française has endorsed "contenu dégénératif" as its official translation for "AI slop", which definitely says something about the comparative approaches of the two languages

[kushblazer666]