disgruntledphd2

> The most charitable way I can read those paragraphs is that he's vaguely alluding to a nebulous conspiracy by a cabal of elites, which evokes a specific ideology that cannot be discussed constructively.

If the markets/finance people reward you for saying you are all-in on AI, people with much of their compensation dependent on the markets (i.e. executives) will claim that AI is driving amazing results.

Like, I find AI useful in lots of places and use these tools daily, but I don't see how the pure play model companies can deliver a good return for their investors.

Unlike ed though, I do think that we'll find good uses for all the compute and (hopefully) the energy infrastructure being built out for this compute.

> But the big labs have claimed they are profitable on inference.

As the accountants say, profit is an opinion, but cash flow is a fact.

Like, if these companies are profitable on inference (depends on how you count training expenses I suspect), then they shouldn't need to raise more money.

For reference, Anthropic appear to have raised about $130bn, which is a lot. Assuming that OpenAI have raised about the same.

Even to get a 2x return on this investment they'd need to start making about $50bn per annum (profit, not revenue) for 10 years. That's Google level net income, on a very very different business (google's business is much more capital efficient).

I personally find that very unlikely, but we'll see I guess.

> a detailed analysis of capex spend and amortization

This is kinda irrelevant unless they are making money (which the hyperscalers are, and the pure play model companies are not).

I get the impression that they'd prefer to sell as much of their TPU capacity as possible.

Which is definitely a risk-minimisation approach, which might leave them in a better place if/when this all collapses.

> t's not clear how you get an actual comparison of "effectiveness" from year 0 to year 4.

You'd probably need a few different measures of attention/executive function and you'd need to counterbalance them across people to prevent habituation.

Given that the regulatory bodies don't require studies like this, it's unlikely to ever happen.

Trump is wildly inconsistent, and his policies are mostly (apparently) driven by personality, so it wasn't that much of a surprise to me.

> Apparently the physiological effects of prescription ADHD meds will tend to still be ~same intensity for years and years.

I'm not sure how one would even measure this. Like, back in the halcyon days of my youth I took a bunch of stimulants, and experentially there was definitely less impact over time (habituation is a thing).

And given that ADHD is predominantly a psychological issue, then the only measure that matters is that of the patients themselves.

Do you have some of the sources you mentioned available, I'd be interested to understand the details?

> You could say the same about Trump, and no, this is not supposed to be a compliment.

That's true, both Mamdani and Trump are very effective politicians. One can disagree with either or both of their positions, but that doesn't mean that they're not good at it.

> ADHD medications do not lose effectiveness over time.

I mean, they're mostly stimulants, which definitely lose effectiveness over time. It would be really weird if they didn't as that's basically how our bodies work in general.

Note: I have been prescribed said medication, but have not started taking them yet so I might be entirely wrong in the practicalities (and how would one disentangle the psychological impact of medication from the physiological impact?).

> disproportionately so relative to their benefits to the world, compared to other big tech companies.

This is certainly what many people around here appear to believe, but there are lots and lots of people who get much more value out of Meta's products than those of any other tech company. Whatsapp alone is probably the most useful tech product for many, many people.

That being said, FB/Meta have done a bunch of awful stuff, but to say that they're worse than Google/Amazon/Microsoft is not necessarily obvious.

Look, that's an estimate (by a presumably biased observer). I don't really have a horse in this race, new large tech companies are fine by me, even if only to reduce the power of the current ones.

However, I find these numbers incredibly hard to believe, and most likely deceptive, given that they recently (like March) started making enterprises pay API rates, so even if they were profitable (and if they paid from Fable training from this post cost money) I would be sceptical that this will continue, given all the competition in this space.

tl;dr let's all wait for the S1 (it will presumably be soon, unless SpaceX declines get them to postpone).

> We most certainly have: https://www.morningstar.com/stocks/anthropics-gross-margin-i...

This is all un-audited speculation. Even audited financials often have massive weirdness, but unaudited revenue/margin numbers are basically garbage.

> The companies are growing so rapidly that citing their financials from several months ago is practically worthless.

I wish I lived in a world where this was true. Unless their inference margins are fat enough to pay for training and employees (including SBC) then their historical financials are really important.

Again, I could be wrong here but without the S-1 nobody really knows (unless they work for Anthropic, in which case they should really not be commenting in this thread).

> Anthropic has 80%+ margins on inference.

If you read the interviews where Dario said this closely, you'll find that he's talking in hypothethicals.

We will have to wait for the S-1 to get audited figures, until then, only insiders have any idea (and only the accountants can be sure).

> In America, ordering a taxi before rideshare apps was unreliable at best, if you were outside of central parts of major cities.

This was not the case everywhere. For instance, Ireland deregulated taxis in 2002, and service improved massively after that.

It's funny how many amazing tech developments are really just a reaction to the peculiar pathologies of American/SF culture.

> It's not necessarily about intellectualism. Science in modern times is plagued by publish or perish. Researchers just need to get things published to keep working, and negative results don't get published. So I think people often publish stuff without proper controls/missing confounding variables/inversion of causality/etc not out of ignorance, but because they know the conclusion is dubious - but if they don't publish, they're going to be out of a job.

This is very true. In my PhD research, I managed to perfectly predict the placebo effect using a random forest. Now, this was actually due to sampling variance, and the finding went away when I averaged over splits.

I often wonder if I'd still been in academia when this happened, if I would have been able to resist the siren call of a Nature/Science paper (and the consequent career benefits).

AI is a power tool/factory with no safety features. Expertise is required to get anything useful from them.

This makes the theory of incentives unfalsifiable, which tracks with how people talk about it.

Their big disadvantage is their terrible, no-good, awful product management (which has always been their issue).

> But I can't believe Dario, probably the most AI-pilled of them all, would under-estimate demand so severely.

To be fair, this mess up actually increases my respect for him, as he focused on making sure that they could continue to run the business under reasonable assumptions rather than YOLOing compute into existence like OpenAI.

Really? I immediately got a warning that the tokens would come from extra usage, and noped out immediately (so maybe pi lied to me?)

To be brutally frank, this was always going to happen when Western countries outsourced their manufacturing.

And all of the rationales were self serving garbage, it was about juicing the next quarters numbers.

Fundamentally though, it's the removal of capital controls that has driven a lot of this behaviour, and it's amusing to me that reintroducing them is so far beyond the Overton window that it never even comes up.

> I think the real punchline was that the value saved was on the order of cents!

If the agents can read less tokens in the future, and make changes more effectively then this would add up over time.

> Nuclear is a high-capex source of electricity. It doesn’t do well economically in a grid where power if essentially free on sunny or windy days but you need a lot of backup power for a few hundred hours per year.

Fair enough, but it's still incredibly helpful as a source of no-carbon electricity.

> My reading of this is that the threshhold for the USA government to withdraw financial (and other?) support for disabled people is too low, is that right?

In general, step functions (earn 200 get benefits, 201 no benefits) are really really bad from a public policy perspective. If you want to get people out of these dependency traps then you need to taper them more effectively.

Even tapering them is problematic as you end up with astronomically high effective marginal tax rates, but it's better than a hard cut off.

> Long game - lol. xAI lost/never gained traction. The talent all left, so Musk is trying to recoup some of the enormous investment in hardware.

This seems more likely to me, but we'll see I guess.

> investing in education has barely an effect on economic growth

Can you provide some evidence for this statement?

Like, imagine a country where 50% of the population was illiterate. Would one expect to see lots of white collar workers in this country?

Sure, but that's more of an indictment of SpaceXAI. All those resources, and they get upstaged by a bunch of much more compute constrained Chinese startups.

If you click on the address result (the button in the map), it will normally work.

Mostly tested on Firefox on Mac in the EU.

I too want to be an artist, but I work in an economy that requires me to be an artist mostly in my off hours.

And honestly, this is such a weird hill to die on. Lint rules when automated are basically pure win in most scenarios, and for your example, just exclude the rule for this block.

For me, at least, the overall design needs to come from using the code.

I think for refactoring, you're definitely right as if you give them a good high level sketch you can get them to do all of the more tedious implementation.

LLm generation of code is an absurdly high dimensional process, and incredibly sensitive to the starting conditions and any other interactions so in some sense it's expected that there will be wildly different outcomes.

Personally, I'm more on the LLM code is pretty bad from a design point of view, particularly when much of the code is already LLM generated.