programjames

> And in those other cases? You have a rigorous definition of comedy?

There's surely more to comedy than subverting expectations. Someone else who cares more about comedy in particular can figure that out for themself, but surely I gave enough of the general idea to make it clear how you could go about measuring the intelligence necessary for comedy.

> A story is believed because a listener "wants" to believe it. Some listeners have more or less complex criteria for acceptance.

Yeah, that's the sense of "better" I was going for. I could have been more clear here, so I'm glad you figured out what I meant.

> Couldn't I say the same about your use of Introspective? Surely a more detailed phrase exists to describe what you mean.

It was a not-so-kind way of saying, "don't point at vague ideas to obscure what you really mean and make it difficult for others to understand what you mean to keep your opinion unassailable."

> Why exactly do you say this skill [resolving conflicts] is not intelligence-based?

Most people have more time to think than they actually use during conflicts, so I expect most of the time conflicts come from people preferring to not think than because they lack the ability. That or a fundamental value difference (you want my food, I want my food).

I purposefully do not give out methods to measure intelligence, because people can train on them. I knew you wanted that, but that does not mean you get what you want. I also find it strange how you expect me to be empathetic in a way that makes you feel good about yourself, when you deserve no such compassion after pulling the dark arts on me.

> How do you measure the capacity for improvisational comedy?

What makes something funny? Usually, it's by subverting someone's predictions. You have to be good at predicting other's predictions to do this well.

> How do you measure a talent for telling convincing lies?

You have to explain a phenomenon better than the truth to convince someone of your lie.

> How do you measure someone's capacity for innovating in a narrative medium?

As in, world-building? That is more of a memory problem than an intelligence problem, though you do need to be good at compressing the whole world into what is relevant to the story. People who are worse at that will have to take more notes and refer back to them more often.

> How do you measure someone's ability for psychological insight and a theory of self?

They are better at explaining a phenomenon (their self).

> How do you measure someone's capacity for understanding complex, multi-faceted irony or picking up subtle social cues?

Refer to the above. Also, using the adjectives 'complex, multi-faceted' is lazy here. Be more introspective and write what you really want to say.

> Or for formulating effective metaphors and analogies, or boiling down concepts eloquently?

Compression = finding short programs that recover the data.

> How about for mediating complex, multifaceted interpersonal conflicts effectively?

Quite often not an intelligence problem.

> How do you measure someone's capacity for empathy, which necessarily involves incredibly complex simulations and mental models of other people's minds?

"incredibly complex simulations and mental models of other people's minds," however will you do this? Oh, I know! Your brain will have to come up with a small circuit that compresses other people's brain pretty well, as it doesn't have enough capacity to just run the other brain.

> Do you think excelling in any of these doesn't require intelligence? You sound like you consider yourself quite intelligent, so are you excellent at all of them? No? How come?

I am actually pretty good at pretty much all of these compared to the average person.

You're quite literally babbling. If a word has no good definition, it ceases to be a word. All you really mean is you use the word "intelligence" very loosely, without really knowing what you mean when you use it. You just use it to point at a concept that's vague in your head. That does not mean you could not make that concept more precise, if you felt inclined to be more introspective. It also does not mean that the precise idea I think of when I use the word "intelligence" is the same as your idea. But they'll often be close enough or even equivalent mathematically, as long as we both have precise definitions in mind.

> What I deeply believe is: We're never going to invent superintelligence, not because its impossible for computers to achieve, but because we don't even know what intelligence is.

Speak for yourself, not all of humanity. There are plenty of rigorous, mostly equivalent definitions for intelligence: The ability to find short programs that explain phenomena (compression). The capability to figure out how to do things (RL). Maximizing discounted future entropy (freedom). I hate how stupid people propagate this lie that we don't know what intelligence is, just because they lack it. It's quite convenient, because how can they be shown to lack intelligence when the word isn't even defined!

I think the issue with the irrationality is that it isn't random. If some people made a mistake 20% of the time, completely randomly, when deciding between two products, it doesn't matter too much, and the dollars flow to the person who trembles less. However, humans can be exploited to make more mistakes. Gambling companies are a clear example, but I think run-of-the-mill advertising optimizes to be exploitative more than informative as well. Thus, all the dollars get sucked in by people who are actively anti-social, instead of those that offer a better product or make fewer mistakes.

"I like to think I'm generally intelligent, but I am not robust to adversarial attacks. I'm trying arc-agi tests now and it's looking bad for me."

One man's modus ponens is another man's modus tollens.

"I'm trying arc-agi tests now and it's looking bad for me. I am not robust to adversarial attacks. I think I'm not generally intelligent."

(1) You don't see the problem with people who are better at manipulating and lying to others through deceptive advertising getting a bigger market share? That's just so obviously bad for society.

(2) Suppose you need insulin to live, but it's suddenly become a meme to start snorting insulin and all the stupids make the price shoot through the roof. That's what stupid people "making prices irrational" looks like, and it happens with fads, or inferior products, or even allowing actual scams to be posted in online marketplaces. Happy smiles on paid actors should not be enough to make your product more appealing than your competitors', and yet the stupids will drive out of business people who don't engage in such pathological behavior.

I see a problem with it. It

(1) makes companies market (lie) more aggressively, because it ends up working out.

(2) makes prices irrational, because if a bunch of stupid people will buy your shit product, why would you care about the 1% who actually do their research?

> if you purely use policy optimization, RLHF will be biased towards short horizons

> most RL has some adversarial loss (how do you train your preference network?), which makes the loss landscape fractal which SGD smooths incorrectly

If there isn't a path humans know how to take with their current technology, it isn't a solved problem. It's much different than people training an image model for research purposes, and knowing that $100m in compute is probably enough for a basic video model.

A good start would be:

1. Robust to adversarial attacks (e.g. in classification models or LLM steering).

2. Solving ARC-AGI.

Current models are optimized to solve the current problem they're presented, not really find the most general problem-solving techniques.

Large latent flow models are unbiased. On the other hand, if you purely use policy optimization, RLHF will be biased towards short horizons. If you add in a value network, the value has some bias (e.g. MSE loss on the value --> Gaussian bias). Also, most RL has some adversarial loss (how do you train your preference network?), which makes the loss landscape fractal which SGD smooths incorrectly. So, basically, there's a lot of biases that show up in RL training which can make it both hard to train, and even if successful, not necessarily optimizing what you want.

Far too much marketing speech, far too little math or theory, and completely misses the mark on the 'next frontier'. Maybe four years ago, spatial reasoning was the problem to solve, but by 2022 it was solved. All that remained was scaling up. The actual three next problems to solve (in order of when they will be solved) are:

- Reinforcement Learning (2026)

- General Intelligence (2027)

- Continual Learning (2028)

EDIT: lol, funny how the idiots downvote

This is only for review/position papers, though I agree that pretty much all ML papers for the past 20 years have been slop. I also consider the big names like, "Adam", "Attention", or "Diffusion" slop, because even thought they are powerful and useful, the presentation is so horrible (for the first two) or they contain major mistakes in the justication of why they work (the last two) that they should never have gotten past review without major rewrites.

No, they aren't. Locally, the person with the most esoteric knowledge is probably a weird nerd. It's mostly an accident that they chose to invest time in things typically associated with smarts. But globally, the best wizards got there by making it their profession. So maybe at your middling university, the people who could land a job at a frontier lab were nerdy wannabe frats, but at decent universities like MIT or Tsinghua, they're usually just better in every aspect of their lives. E.g. MIT has "math olympiad fraternities" all the cool kids join.

Update 3:

Okay, so my PixelCNN masking was wrong... which is why it went to the mean. The earth-mover did get better results than negative log-likelihood, but I found a better solution!

The issue with negative log-likelihood was the neural network could optimize solely around zero and one because there are poles there. The key insight is that the color value in the image is not zero or one. If we are given #00, all we really know is the image from the real world had a brightness between #00 and #01, so we should be integrating the probability density function from 0 to 1/256 to get the likelihood.

It turns out PyTorch does not have a good implementation of Beta.cdf(), so I had to roll my own. Realistically, I just asked the chatbots to tell me what good algorithms there were and to write me code. I ended up with two:

(1) There's a known continued fraction form for the CDF, so combined with Lentz' algorithm it can be computed.

(2) Apparently there's a pretty good closed-form approximation as well (Temme [1]).

The first one was a little unstable in training, but worked well enough (output: [2], color hist: [3]). The second was a little more stable in training, but had issues with nan's near zero and one, so I had to clamp things there which makes it a little less accurate (output: [4], color hist: [5]).

The bits/dim gets down to ~3.5 for both of these, which isn't terrible, but there's probably something that can be done better to get it below 3.0. I don't have any clean code to upload, but I'll probably do that tomrrow and edit (or reply to) this comment. But, that's it for the experiments!

Anyway, the point of this experiment was because this sentence was really bothering me:

> But categorical distributions are better for modelling.

And when I investigated why you said that, it turns out the PixelRNN authors used a mixture of Gaussians, and even said they're probably losing some bits because Gaussians go out of bounds and need to be clipped! So, I really wanted to say, "seems like a skill issue, just use Beta distributions," but then I had to go check if that really did work. My hypothesis was Betas should work even better than a categorical distribution because the categorical model would have to learn nearby outputs are indeed nearby while this is baked into the Beta model. We see the issue show up in the PixelRNN paper, where their outputs are very noisy compared to mine (histogram for a random pixel: [6]).

[1]: https://ir.cwi.nl/pub/2294/2294D.pdf [2]: https://imgur.com/e8xbcfu [3]: https://imgur.com/z0wnqu3 [4]: https://imgur.com/Z2Tcoue [5]: https://imgur.com/p7sW4r9 [6]: https://imgur.com/P4ZV9n4

Update 2:

After another 24 hours of training and around 100 epochs, we get down to 4.4 bits/dim and colors are starting to emerge[1]. However, an issue a friend brought up is that log-likelihood + beta distribution weights values near 0 and 1 much higher:

     log(Beta likelihood) = alpha * log(x) + beta * log(1-x)
                                      ^
                               log(0) --> oo

This means we should see most outputs be pure colors: black, white, red, blue, green, cyan, magenta, or yellow. 3.6% of the channels are 0 or 255, up from 1.4% after 50 epochs[2]. Apparently, an earth-mover loss might be better:

    E_{x ~ output distribution}[|correct - x|]

I could retrain this for another day or two, but PixelRNN is really slow, and I want to use my GPU for other things. Instead, I trained a 50x faster PixelCNN for 50 epochs with this new loss and... it just went to the average pixel value (0.5). There's probably a way to train a mixture of betas, but I haven't figured it out yet.

[1]: https://imgur.com/kGbERDg [2]: https://imgur.com/iJYwHr0

The human traffic code is also written in blood. But humans are worse at applying the patch universally.

The model seems pretty shitty. Does it only look on a frame-by-frame basis? Literally one second of video context and it would never make that mistake.

What dates? 2000 August–2025 August gives 3.18%/year for the consumer index and 7.12%/year for egg prices. Even if you assume egg prices will come down to $3/dozen, it's still 6.24%/year.

Update 1: After ~12 hours of training and 45 epochs on CIFAR, I'm starting to see textures.

https://imgur.com/MzKUKhH

Hmm, I think a mixture of beta distributions could work just as well as cateogrical here. I'm going to train it for PixelRNN, but it's going to take hours or days to train (it's a very inefficient and unparallelizable architecture). I'll report back tomorrow.

I would prefer you refer to it as "courtesy" or "consideration" rather than "freedom".

I don't think schools have always been like that. I especially don't think schools have been like that for the top 10% of students until very recently (<20 years).

And paper money is a Chinese invention. Doesn't mean it's worthwhile to spend two weeks in an anthropology class talking about how much awesomer they are.

That isn't a comparison to the state of the art, just a naive quantum clock.

I believe the issue is measures of 'achievement' and 'excellence' typically use the 20th and 50th percentiles, respectively. For example, most education studies focus on what improves graduation rates, passing rates, or the bottom 10–20% on standardized assessments. Most school districts' financial incentives rely on these metrics as well. Occasionally you'll get a study that looks at the median (and even more rare, the 90th percentile), but no standardized assessment even showcases the bellends of the distribution. This is why you get claims like this

> High-achieving kids are doing roughly as well as they always have, while those at the bottom are seeing rapid losses

in the article, when they're actually doing much worse as well, which can be seen from the median (or top) scores in STEM competitions. All the financial incentives are for schools and teachers to focus on their bottom 20% of students, and even if they cared about their top students (culturally, most education departments do not), it's not even feasible to collect that data, short of signing everyone up for the AMC 12. So, naturally, the percent of elementary schools offering gifted programs has declined by fifteen (twenty?) percent in the last twenty years, and the general standards and curricula have been lowered to teach to the 20th percentile. This also creates a cultural issue where many students recognize they're bored, and that schools are not trying to teach them, so they become disaffected and stop caring. My friends in high school joked that school was for socializing, and frankly, what other purpose does it serve anymore for most kids?

It's a little easier to see what's happening if you fully write out the central flow:

    -1/η * dw/dt = ∇L - ∇S * ⟨∇L, ∇S⟩/‖∇S‖²

We're projecting the loss gradient onto the sharpness gradient, and subtracting it off. If you didn't read the article, the sharpness S is the sum of the eigenvalues of the Hessian of the loss that are larger than 2/η, a measure of how unstable the learning dynamics are.

This is almost Sobolev preconditioning:

    -1/η * dw/dt = ∇L - ∇S = ∇(I - Δ)L

where this time S is the sum of all the eigenvalues (so, the Laplacian of L).

I think the take is: If 100k people watch the episode, spending $200 more for higher quality subtitling comes out to... a whopping 0.2 cents per person (per episode). Let's just say that would cost an extra $1/month per person. Are they price sensitive enough that they won't go to a competitor that's a few dollars more expensive per month if it has better subtitles? I don't know, but maybe some manager believed they were, and thus it was worth it to make the subscription a little cheaper.