You know, I think this totally doable (though I probably wont have it close the window). I think I've found my next hobby project. Not going to be at the same quality of writing, obviously, but Qwen 3 can be prodded into being deranged enough to get the right mood.

I guess that would be a fitting job for AI to completely obsolete.

There is another point here about subjectivity, and LLMs having no direct apprehension of the world. That's true! It's really interesting what that could mean for concepts like "knowledge". But saying we know *conclusively* what that means is pretending we've solved Philosophy of Mind.

I'm not even saying that's wrong. Maybe understanding has some Qualia component an LLM can never achieve. I will happily read someone's publication arguing that. But I also want to hear how they know *people* have this "understanding." Instead, it's usually "it can't understand because it's random."

We can say "yes, it contains a map, but it doesn't *understand* it!" Ok, sure, but if "understand" doesn't encompass "an abstract representation of the concepts in the text it was fed which it can use in novel contexts to solve problems," I'm going to need to know what we *do* mean by it.

It feels a lot more parsimonious to say training the model has produced a structure somewhere that is like a map. For models where you can get the full reasoning trace this is even more compelling. The answers start by noting states in that general region, then check them individually.

All of the big models released this year can answer this question well without tool calling. We can try to tell ourselves that this is just because the model has memorized every possible description of geographic relations between states, though this isn't a common benchmark or thing people discuss

I'm pretty sure I'd barely have an advantage over chance identifying LLM generated ML papers mixed with real ones from 6 months ago. There is very little signal to use there beyond memorizing what papers are real exhaustively, which will always be weak unless they come up *a lot* in the corpus.

There was a website ca. 2010 that used a CFG to generated High Energy Physics and Mathematical Physics paper titles ("Super-paramagnetic Black Saturn: A Review"). Even as a grad student in the field it was hard to tell the fake from the real most of the time. I think you're on to something here.

It's possible, but most academic ML / SWE training paths really don't give you the stats / applied math knowledge you need to do the science role well. You *can* learn that yourself, though. It's going to be quite a bit of work for most people.

The path into the positions is often different too. A lot of "engineer" roles come up through SWE and don't have significant stats or applied math training. A lot of the "science" side comes out of math / stats / physics and can be pretty shaky on SWE concepts.

This really is a lot of it. It's also not a "soft" PhD position where you aren't really going to use a lot of the training. The difference in amount of math you need between "making a novel architecture that is even semi-functional" and "implementing the prod version of that architecture" is huge.

They differ in training a little, but also less than you’d think (less RL for vision). There’s been broad convergence in how these models work for the last 5 years.

A lot of this is also the collapse in meaning of “hallucination” from technical term-of-art for gen AI into popular meaning of “an AI is wrong”. But also modern vision networks use Transformer (SWin and its like) past the first few layer or two just like LLMs. Attention is too good to ignore.

I had a student come up after a talk ~2 years ago and tell me that "[Prompt Injection] is easy to solve---you just parse the grammar and determine what is a quote and what isn't." I love that LLMs have un-siloed semiotics from abstract thought experiments into direct engineering relevance.

The relationship between language, meaning, and thought is philosophically... controversial. You could easily spend a whole PhD doing nothing but arguing for *or* against every clause you have written here. That's without even mentioning LLMs.

I’ve seen a lot of signs but definitely feel the pride vibe too. Not complaining though—just hoping the weather keeps holding

Spectra also can’t be red all the way to zero, but it’s very common for them to be red down to the lowest frequencies sampled. Images in particular that have increasing power for a broad part of the sampled frequencies look *weird*

If you had some domain with spectra that were very blue you’d see the reverse. The reason reason real data usually aren’t is separate and *also* fascinating. Obv. finite variance signals can’t be blue asymptotically, but there are very few where they are even over an interesting chunk of freq. space

The reasons diffusion does this for images and time series are really interesting. It comes out of the stats of the signal that’s being modeled, not the method. Image power spectra are usually pretty close to pink. Correctly sampling the low frequencies first is optimal. It doesn’t have to be so.

Exactly! And the movie hates them for it. The ways it's a bad movie are so much more "predictive" than the world it shows.

So much of what was detestable about comedies of that time seemed to have burnt its way out of Hollywood in the 2010s. It's a pity that cruelty, cynicism, and contempt for virtue have exploded back into vogue again---like the villain at the end of a slasher movie. It was behind us the whole time.

I think this is ironically due to the way the movie is so problematic / wrong, not what it "got right". The mean-spirited, openly misogynist, weirdly eugenics-tinged vision of a future ruled by "idiots" *does* have surprising currency with the present. Just mostly not how it thinks it will.

YouTube video by Sarah Z No, Idiocracy Is Not A Documentary

Heard someone say (again) today that "Idiocracy was a documentary," and I thought back to the excellent @sarahz.bsky.social's video on it: www.youtube.com/watch?v=o52z...
I don't disagree with her. I think her points hold up. But the current moment and the movie still seem to have "resonance".

Not per capita. In total. It’s embarrassing that so many otherwise reasonable people keep saying this. You can keep hating AI! We can have that conversation another time. But repeating the water use thing is like conservatives complaining the NSF spent $40 k researching frog genders. It is *tiny*

People seem to think the point is that banning ChatGPT wouldn’t reduce water use as much as all of of us deciding to stop eating meat. It’s not. Banning AI entirely would have the same impact as a single school not using meat any more. Banning new training is like *just you* deciding not to eat beef

People *vastly* overestimate the cost—in water *and* power—of this industry. That’s the point. GTA 6 is almost certainly going to cause more environmental harm than ChatGPT.

There’s this incredible, persistent innumeracy in the replies to this and the original where people say “but I can eat beef! LLMs are useless!” Or “as if we shouldn’t just stop eating beef.” *Training* one of the mostly costly models to produce to date used maybe *1* entire cow.

ACAB includes defense attorneys advising you not to talk to the cops. Anyone who limits my freedom is a cop. Anyone. ACAB even includes the person who scrawled the “Don’t Open Dead Inside” sign on the door holding the zombies back.
You may not like it, but this is what peak Collectivism looks like.

This was amusingly *easy* compared to my previous attempts at tone forcing this way. It’s almost like there’s a vast corpus of this sort of “critique” to draw on.
The harder part was getting it not to do *overt* parody. Models need a lot of convincing to double down on “Pydantic and the Death Drive”