dlkf

If you don’t quantify these things and perform cost-benefit analysis, it’s like saying we shouldn’t drive cars , because in the worst case you burn alive.

SSRI side effect prevalence is a hard thing to measure for a large variety of reasons. But in [1] only 1/4 patients report the side effects as ”very bothersome.” Given that 100% of patients are so anxious/depressed that they want to try psych meds, it seems like SSRIs are a great treatment option for many.

[1] https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2719451/

Parent claims that because we don’t eat meat plain, this shows that it isn’t the meat that ”tastes good”, its the seasoning. The fallacy is that we could have applied the exact same logic to the seasoning: we don’t eat steak seasoning plain, therefore it isn’t the seasoning that tastes good, its the meat. By parent’s logic, meat both does and does not taste good.

The resolution, which has always been obvious to everyone (except hackernews, apparently) is that flavour is the result of the combination of ingredients. That you typically eat X and Y together doesn’t mean X is tasteless and Y is tasty (or the reverse). It just means X and Y combine to make a good dish.

The fallacy here is pretty transparent. On the one hand, you’re saying meat ”doesn’t taste good” because it is usually not served in isolation. But on the other hand, you claim that seasoning tastes good, despite the fact that it is never served in isolation.

I love that someone has this take, but I think you should elaborate. In what sense are they terrible? Why do they experience sustained popularity? What does this tell us about Lindy?

The only detailed example Doctorow draws on (ODF) shows the free market converging on a good solution without government intervention. It’s an interesting case study, but it doesn’t illustrate what he thinks it does.

Like what? The lack of indoor plumbing? Or perhaps the high infant mortality rates? The only way to believe this is to live in total ignorance of history.

> We have reached the point of diminishing returns when it comes to technological development.

Every generation says this, while at exactly the same time regarding previous ways of life completely unthinkable.

Martin Newell is an English singer and poet who probably doesn’t get the credit he deserves.

I encourage you to research whether flying has gotten safer or more dangerous since 1970, and then consider why that is.

Those things require energy. Incremental efficiency improvements will not make a big dent. Either we continue to use the same level of energy, or we start giving things up. It’s simple arithmetic.

Your comment, along with cpgxiii and n2d4’s are all really good. I have a question: suppose training and inference of an LLM were made to be deterministic at the cost of performance.

Would the cost be “everything will take twice as long” or would it be more like “inference will take a week and training will take a couple lifetimes”?

If it’s the latter, then it seems disingenuous to call this a “bug.” It’s like saying F1 cars could be horse drawn, and they only use internal combustion for “performance reasons.” If its the former, then maybe there is a more interesting discussion to be had about the potential benefits of determinism? (That said, I agree with n2d4 that it’s stupid to insult the authors. Talk is cheap and building is hard.)

> the "rocket fanatics" peaked in 1969 and humans haven't gone back to the moon, let alone conquered space

What’s the point of going to the moon? Meanwhile, we have thousands of satellites in space that help predict the weather and facilitate instant long distance communication.

When you start using eg meth, you open up, with reasonable probability, a whole host of other externalities beyond impaired driving. (The same is true of alcohol, except with far lower probability, which is a key disanalogy.) Hard drug users commit property crimes and violent crimes at a much higher rate than people who don’t use. Crime increases during binges, suggesting causality. But you really don’t really need academic study to demonstrate causality here - it’s obvious that people steal to fund addiction.

And yes, you can prosecute them for whatever property or violent crimes they commit, but for the victim, it’s too late! They have already suffered. A better outcome is if we decrease the amount of usage of drugs like meth and fentanyl via law enforcement, and subsequently decrease the externalities that society has to bear.

But it allows for an increase of the frequency at which those crimes are committed - at which point it’s too late. Sure, can punish the criminal afterwards, but the victim has already sufferred.

> In case you aren't, being able to do what you want with your own body does not include a right to harm others

You have to make the argument that drug use doesn’t harm others. In a libertarian thought experiment maybe it doesn’t, but in the real world it does.

In particular, we tend to criminalise behaviors that increase the risk of harming others. Consider drunk driving. Nobody is harmed if I don’t have an accident. But the risk of an accident increases, so we criminalize it. The same is true of hard drugs. The probability I will harm others conditional on meth addiction is higher than the background probability.

> It's your body, you can do what you want with it.

I can’t hit people with it, I can’t transport it onto private property, I can’t use it to yell all night, I can’t sell organs, the list is endless.

Anything we do, we do with our body, and every law ever applies to what I do with my body. The question is: do we think the costs of law the restriction outweigh the benefits? You may think so, but you have to argue for it, you can’t just invoke a principle which has never existed.

> the applications of these sort of reasoning atomic configurations I think are unlimited

They are limited to applications in which the latency slo is O(seconds), knowledge of 2021-present doesn’t matter, and you’re allowed to make things up when you don’t know the answer.

There are, to be fair, many such applications. But it’s not unlimited.

> They’re hugely successful corporations. Saying they’ve “lost” is a tad ridiculous.

Let’s quantify it so it’s not ridiculous. Since getting their head start, they have ”lost” about 90% of total market share to American companies who are doing a better job. And this is before the new transfer framework that cane out this month. What is the value prop of OVH without regulatory capture?

> Europe is also a hotspot of AI research

This is absolutely true. Amsterdam for example produces outstanding ML research. But where do their freshly minted phds go? American companies!

> I really do get fed up with how some Americans believe it’s impossible that any other country could be successful.

For what it’s worth, I live in northern Europe. As you point out, European universities are great. They are an AI research hotbed and they produce highly capable undergrads. But this clearly isn’t translating into world class tech companies, and we should ask ourselves why.

> OVH and BT for example. Both of which predate AWS and GCP.

How is this not an own goal? These companies had a head start and they still lost. Not by a little, but by such a staggering margin that AWS does two OOMs more revenue.

> I literally just said I work for an AI tech company in Europe.

The argument is not ”there are no AI companies in Europe.” The argument is ”there are no competitive AI companies in Europe.” The only one I can think of is DeepMind, and they are A) English (no longer EU) and B) were acquired by Google ten years ago.

What are these European tech companies that are competitive with Nvidia, OpenAI, Google, etc? I’d like to learn more.

> You do realise that London, Berlin and other European tech hubs have their fair share of talented engineers who have moved to that country for work too?

A lot of them work for subsidiaries of American companies.

> creating art or music or playing go

One of these things is not like the others

How does the issue of plagiarism bear on its utility? That’s like saying it’s impossible to get value out of listening to a torrented mp3.

I’m not sure understanding is an inherent issue either. When I Google ”weather lisbon,” Google spits out a number. The app doesn’t ”understand” anything, but it still helps me plan my day.

> But analyzing his mental condition is sure as shit not up to me or to anybody else.

Why not? Am I unallowed to suggest that Howard Hughes had OCD? That Kurt Gödel was paranoid?

Submitting a hypothesis about some’s mental condition is a reasonable thing to do, given that you have a substantial body of evidence in the form of behavioural reports.

I don’t find the above argument convincing, but only because it’s vague and lacks specific examples - not because it’s ”not okay.”

By ”the Pepsi critique” I specifically meant the exposure of both variants to a single user.

I agree with you that having users take a single sip is also a problem for construct validity. But it’s not clear from your post how this would this would happen in a properly run AB test.

Are you thinking of the case in which we ship a change that confers short term wins in user spend or churn, but comes at a long term cost that isn’t visible within the window of the test? This is a common problem, but it’s not clear to me that it’s the same problem Pepsi had. Here the problem is with choice of response variable, whereas Pepsi’s problem is choice of treatment.

That isn’t how AB tests work. In a (properly-run) AB test, a user will only see one variant. AB tests have shortcomings, but the Pepsi critique isn’t one of them.

Da Vinci was a renaissance man?

Your argument would apply equally to a drunk driver who didn’t happen to crash that night.

The problem is that taking pcp significantly raises the probability the user will commit an act of violence, by which point it’s too late.

> Even if you just make GPT-4 say 33% smarter

What is your unit of intelligence?

> It's a massive, massive advancement in the field of statistics and learning from data, but doesn't seem to map to my conception of intelligence at all.

You think ”learning from data” has nothing to do with ”intelligence” ?

I’m not talking about a single paper, I’m talking about a whole research programme. But Rumelhart, Hinton, Williams 1986 is very improtant. A lot of the foundational work is collected in the PDP report.

The idea that knowledge can be expressed in distributed representations rather than sequences of symbols is a huge advancement in our understanding of intelligence. Obviously we haven’t ”solved intelligence” but now we at least have some idea of what the right questions to ask are.