tskj

I agree it does seem pretty plausible as some kind of iterated / timeless theory divorced from reality, but especially the assumption that time is irrelevant to the buyer seems completely fanciful. I am usually all for theories that make idealized assumptions and think there's merit to approximations, but this is so egregious--there will never be any market or situation that is anywhere close; and the theory completely breaks down if the assumptions aren't utterly fulfilled. That makes it not even of academic interest, it's completely useless even as a theory.

I'm curious if they're keeping up mostly due to distillation or how that works. Does anyone outside China know?

Yeah I guess it's fine to demand the situation actually conform to the assumptions of the theory, but these assumptions are so extreme they have literally never held for any situation, and they never will.

But this conjecture predicts that Steam prices will drop immediately to their final low price. So Steam is actually also a counterexample.

Idk the obvious answer seems to be that buying now vs buying later isn't the same? Seems like a preposterous assumption; or rather an assumption that obviously never holds for any market ever so this theory is unfalsifiable by empirical data (and also irrelevant to the real world).

I think basically economists aren't that interested in looking at real world, messy data.

First of all, yes cutting developers to fund AI spend budgets, is the entire operating idea behind these AI companies; and most companies would love doing that. I'm not saying this is a good thing, my heading is on the chopping block like everyone else's.

But isn't this like a Jevon's paradox thing, also? If I'm able to become vastly more productive, and that value produces more sellable output for my company, there's no reason to cut anywhere to fund it. This is the same reason a company like Microsoft can hire 80 000 developers, it's because each dev pays for themselves in value (on average). I guess the same can be true for AI spend?

No, but if you'd actually read what I wrote, you'd see that what I said (that Dario said) was: if you consider each model as its own company, they've all been individually profitable. However, they keep re-investing that (and then some) into even bigger, more expensive models each time, causing the company to look unprofitable.

That being said, Anthropic did report being profitable this review quarter (Q2), so it's not as unreasonable as you claim.

Also Google is a pretty major AI company, and they're _insanely_ profitable.

Unfortunately most of society and the economy is built on the foundational assumption of a growing (or at least stable) population. There will be some extremely tough challenges to solve with an actively declining population.

Interestingly Wind Waker's art style was its main detractor among critics when it was released, which is wild and incomprehensible to me now. One of my favorite games of all time.

I don't know man, people basically aren't having sex already. The fertility rate is way below replacement in all modernized countries, half of gen Z has never had sex at all, contraception is free, people struggle to get pregnant the more chronically ill the population gets, and the people who actively do want to get pregnant are going to want to use whatever technology there is to improve things.

Thanks, that's an interesting analysis. I'm not a Christian, but I definitely get the sense that we're playing with an extreme danger here, and that we're not being sufficiently humble / cautious / in awe of the sacred -- and that such hubris might literally lead to human extinction, or close to it.

Well that's the point about unknown unknowns though, we actually have no idea what the probability is. But we do know it's not low, that's the only unlikely scenario.

We only know it's a very complicated system with many interlocking dependencies, and based on what we know about complicated systems in general, as well as biology in particular, is that if any one of these unknown dependencies break, the whole system can fail catastrophically.

Therefore the probability that something will go wrong is very high, and the damage could easily be irreparable damage to the species, if not extinction. Does that not give an intolerably high risk?

I think the case for this is pretty strong actually. Last year my company was maybe willing to pay $100 a month to Anthropic (per developer). Today we're all on the $300 plan without any hesitation. If Fable ever becomes available as the default model, I imagine my company would be willing to pay in the $500-$1000 range per month per developer.

Well yeah obviously they have to stop reinvesting more than they make at some point to become profitable. To be clear, I think what Dario was saying was that if you consider each model training + deployment as a company, meaning all expenses and taxes, it was still profitable.

Whether he's lying is another question, but seems unlikely.

I guess I'll chime in as someone who thinks LLMs will be earth shattering, and specifically don't think it's a net positive for anyone but those whose power will be consolidated.

This seems distantly impossible right now, but for this reason, I predict that any species that survives this kind of "great filter" effect of accidentally messing up their genome long term, will develop a strong taboo against fertility treatments and treatment of genetic diseases.

Like it seems horrible not to help the individual, when we have the technology to; but it's also horrible to hurt your species by selfishly propagating faulty genes. And this seems like the kind of problem cultural taboos are good at solving, and I don't really see any other mechanism by which a species can avoid this filter trap.

I think this is decisively the wrong way to think about it. Yes, layering hypotheticals like that means that any one scenario is extremely unlikely to be the thing that gets you, but that doesn't mean the shape of the problem is wrong.

It's like arguing with someone who doesn't believe in using seat belts when driving. "Why should I put them on?" they say, and when you try to explain what might go wrong they won't listen to any explanation that isn't a hyper-concrete hypothetical. So finally you give in and say, "Well, when we get onto the highway, a truck might lose control and hit us", and their response is "I don't think that's very likely, it seems highly improbable that today we will be hit by a truck when getting on the highway".

I agree with OP that this seems like the kind of thing where the unknown unknowns are so great that the correct approach is serious caution, and that any demand to know exactly how or why it will go wrong, falls in the trap where every specific example is very unlikely to be the thing that goes wrong, but still in total there's like an 80% chance that it goes horribly wrong. I don't know if we have the terminology to talk about this kind of failure mode. "You shouldn't play God" maybe? At least you shouldn't ask for specific examples of how things could go wrong, if you're going to turn around and claim each one highly improbable.

Why is it a designer's dream to hijack and control my scrolling experience? The scroll they've implemented is slow to respond, and has a weirdly low capped max speed. I don't understand why that's what a designer dreams of doing to me. I like my scroll (and other computer interactions for that matter) to be responsive and fast. You know, the kind of thing that puts me in control, not the designer.

That being said, the scroll was as smooth as regular webpage scrolls. Usually these JS scrolls aren't able to avoid dropping frames or otherwise introducing judder, but this one does appear to run at a consistent and high framerate, which is technically impressive.

You didn't really have to use it more than a day honestly to tell what kind of shocking paradigm change it was. Man do I miss it.

Dario has publicly claimed each model has been profitable, even accounting for its training costs; it's just that each new model is exponentially more expensive to train than the last, so the income lags and it looks like the company is losing money overall.

Now, we can't know if this is true unfortunately, but it's not directly contradicted by anything that's known publicly at least. I thought it was an interesting way to frame it and makes the whole situation look marginally less bad.

You might like Casey Muratori's "The Big OOPs" talk about this, if you haven't seen it yet: https://www.youtube.com/watch?v=wo84LFzx5nI

It's my favorite deep dive into the subject, it's super thoughtful and well structured.

I'm glad you like it. I feel validated not being the only one!

I believe laws were broken, even though no one contested it with the statute of limitation.

Saying they don't posses any level of intelligence is wild.

Because you can have it review itself and iterate on its own work before showing you. If you insist on reviewing its one-shot output you'll be disappointed, but if you consider its internal work private and only consider its final output, it's different.

Also we're still in the middle of the transformation, clearly the AI we'll have in 5 years will be radically different and better (by some definition of better) than what we see today. It's kind of weird that you'd be disappointed that the world will only be totally transformed in ten years, and not five.

The universe.

Anyway, I accept this usage of the word "lesson", so I also accept "learnings". My point was one of hypocrisy, not policing people in how they can use the word "lesson".

Hmm yes, I had it backwards. I agree this is very statistically significant, but the effect size is tiny. Up from random chance to a mere 60%, means that Scott proved with high statistical significant that people reliably cannot tell.

Also I'm not that worried about the adversarial conditions, any real life conditions are likely adversarial in the relevant sense. No one is one-shotting generation and serving you that, obviously output has been selected for quality. I would call Scott's test not adversarial, but fair.

I'm not sure I understand your argument. If you want an exponentially more expensive training run for each iteration, obviously you need to raise investments even if each training run is profitable. Now I'm not saying that's a good idea, or makes sense, but I am saying that "raising money" doesn't disprove neither that each training run makes money, nor that they're re-investing all that money in the next run.

To give a simple example: if each run simply makes a 10% ROI, but you want to spend 2x as much money on the next run, you still need to raise 90% of the previous run's expenses to have enough capital.

Thank you, this is obviously where we're heading. People who think in terms of "will it ever be profitable to sell tokens" are thinking in the wrong framework entirely. The correct framework is "will it be profitable to sell knowledge work", and the answer will clearly be "yes".