@ThenElection's banner p




0 followers   follows 3 users  
joined 2022 September 05 16:19:15 UTC
Verified Email


User ID: 622



0 followers   follows 3 users   joined 2022 September 05 16:19:15 UTC


No bio...


User ID: 622

Verified Email

Just to psychoanalyze from afar, she (for some reason, the writing seems very feminine, and I'd bet good money that the writer is female--a male wouldn't bother writing this email, and if he did it would be written more as a string of curses without any attempt at politeness) is a not-particularly well-educated but generally well-meaning individual from a bad background who has probably done things mostly right (at least within her context) but nevertheless has accrued a lot of damage over her life and isn't particularly successful. The letter is half an outpouring of frustration at where she's ended up, and half an attempt at some kind of connection. What kind, it doesn't matter--the response could be an angry condemnation back, and it could be self-flagellation from the recipient. Doesn't matter. She just feels very lonely and isolated and desperately wants some kind of reassurance that she's seen. She makes her pain obvious by lashing out in the hopes that it will create some kind of response.

I'm a broken record on this, but Native Americans have no better standing. They ethnically cleansed the pre-Clovis peoples from North America, with the only difference being that they were smart enough not to leave any records of their existence (aside from traces left in their mtDNA).

I'm not aware of that issue in Stable Diffusion; the recent fracas seems to be about an inability to consistently generate images of people at all, not just white people.

If you mean more broadly, sure. But I meant with respect to SD3 in particular, not issues with unrelated products.

It's definitely an approach that many have been taking, and it's likely to be significant for the next couple years.

A major issue with it is that many times general intelligence needs to answer questions that cut across multiple domains. You route a question to one module, and it turns out another module would be better for it; or, more fundamentally, you need a module that incorporates elements of both, and either one individually (or both combined with a module meant to synthesize them) is inferior to the ideal result. By analogy, suppose you have a problem where you want to do sophisticated analysis on millions of x-rays: although you could probably get by with team of experts (e.g. a radiologist and a statistician), the results will probably be inferior to an expert radiologist who's also an expert statistician.

PC, as in wokeness? Not at all.

PC, as in not wanting to be labeled an adult content company? A whole lot. It means more legal hassles, higher credit card transaction fees, and limits your market to the consumer segment.

Reminiscent of Dr. Evil: "Here's the plan, we use AI to write home inspection reports, and we hold the market captive for... ONE MILLION DOLLARS."

Sure, do that a thousand times and you get to a billion. But this is all chump change. And when you're LoRA-ing your Llama, whose GPUs are you training on? Do you have a cluster in your basement, or are you calling some API hosted by a big provider? And when you're generating the home inspection reports, where are they actually coming from? Even if you're able to do that all yourself, can you do it faster than someone who takes some VC money and writes a big check to a big provider to GTM faster?

None of that is to say that there's not money to be made; there is. But a substantial chunk of it will accrue to centralized providers.


Most of the economic value of LLMs will come from enterprise, not consumers. For that, there's a variety of factors beyond cost-to-serve that will drive centralization to a few large providers. See the move to cloud in the 2010s. E.g. HIPAA, privacy, "no one ever got fired for buying IBM," reliability, access to large and growing proprietary data sets, custom hardware for specialized workloads, complex data residency requirements.

Google is much better positioned for this than many people give it credit for. OAI will have to learn to deal with a bunch of stupid shit ("what, the government of Saudi Arabia is demanding we set up a data center in the country if we want to operate there? And integrate with a bespoke IAM system to allow the government to view user data?"), and they're just starting to run into these issues. Their saving grace is their relationship with MSFT, which has solved all of these problems; otherwise they'd be dead in the water.

As for the wealth captured: more than they do now, but not everything.

I agree that Chollet is overly negative on LLMs, but what's the issue with the benchmark? It's a set of well-structured problems that any child could solve, but most current models fail miserably at. My only complaint about it is that he makes it impossible for leading LLMs to actually compete in it.

The myth of hypoagency rears its head again. Women are always victims: they might end up being part of something worthy of condemnation, but it's always because men forced them to as part of a plot to subjugate women. They're hapless objects who aren't capable of meaningful agency.

Popular feminism is infested with this ideology, which manages to be deeply sexist and reductive against both men and women.

In terms of revealed preference, the NYT-type stance would be inverted. I guarantee the vast majority of NYT male writers and readers regularly watch porn. And the vast majority of NYT readers and writers of either sex would be quite upset if their 18 year old daughter came home and announced that she was going to be a professional porn star and needed a couple hundred dollars for anal training devices.

Chollet has been skeptical on LLMs awhile (predating GPT-3); this isn't a recent thing for him, and he's not IIRC ever been an apocalyptic type. And Stable Diffusion's woes are the result of external regulatory forces, not inherent architectural limitations or costs.

So where are things? I think the frontier labs are in a consolidation phase: what exists now is entirely capable of having massive economic effects. It's not going to prove the Riemann hypothesis, but it can replace vast amounts of low-level intellectual work that's based on pattern matching and stitching together bits and pieces present in training data. This ranges from call center workers to the bread and butter work doctors, attorneys, and SWEs do. There's an extraordinary amount of money in even that. So that's what they focus on: making models cheaper to deploy and more consistent, product integration, and satisfying public opinion and regulators (and ideally regulatory capture to prevent upstarts from breaking into the market). Look at OAI's job postings, and this is clear (though to be fair, top level researchers are not being recruited via OAI's career portal).

There are lots of potential improvements even in published research papers; who knows if they'll scale. The leading labs will experiment with the most promising ones, but giant leaps in capability are not critical to success, except if a competitor lab stumbles on one before they do.

For those who worry about an eldritch abomination being released on the world, this is a good thing. If you're worried about a Great Replacement by AI slop, or if you were hoping for a rapid jump toward a post-scarcity utopia, not so much.

It's even simpler than that (and applies to both men and women). Dating preferences are heavily correlated. If you find a single person who matches all the desirable dating preferences, they likely have other factors that make them unsuitable for a long-term relationship; otherwise, they'd have landed and stayed in one. The old saw "all the good ones are taken" is true.

This is why widows/widowers tend to be really good potential partners, despite the baggage of their partner's death. They became single through no choice of their own, but still have all the traits that make them a good partner.

Your vote for no one doesn't matter: the only thing it does is create, in your mind, a relationship between you and "democratic processes." It cannot change the results of the election, and a burst in spoiled ballots isn't going to influence policy or governance.

The main effect of casting your ballot is going to be how it influences how you think and relate to the State, granting it more legitimacy than actually exists. It's the secular equivalent of hating religion, getting nothing out of it, and nevertheless insisting on going to church with a constant grimace on your face to register your objection.

(It's still better than voting for "the lesser of two evils," which has a tendency to drive people to adopt a tribe and shift their positions to make the lesser evil feel less evil or even actively good.)

I don't think male infants being more likely to be victims of infanticide is a strong argument about gender bias one way or another. It is a strong argument against the idea that today, in most societies, female children are more likely to be a victim of it than male children: the contemporary US is a society that exists today that has some relevance.

And, yes, there are strong ethnic trends contributing to it.

In the US today: depending on what you consider infanticide, when IVF is used sex-selectively, it's usually to select for a girl. And when children are being adopted, adoptive parents have a strong preference for girls.

For infanticide itself, according to the CDC, boys are more likely to be victims of it than girls, both in absolute terms and proportionately (8 boys per 100k person years vs 6.2 girls per 100k person years): https://www.cdc.gov/mmwr/volumes/69/wr/mm6939a1.htm#T1_down

I couldn't find concrete statistics on sex-selective abortion in the USA, but I'd expect it to follow the same trend.

Mass shootings (non-gang related), serial killers, and stranger rape all of have one other thing in common: they're extraordinarily rare. Most incels just descend into listless stagnation; they're not going to be agents that fundamentally change society.

It was tongue in cheek (not obviously enough so, apparently).

Those latter groups can easily be ignored and shuffled off to other physical areas for other people to deal with, on a day to day basis. Sexless men seem to sprout up everywhere and are encountered everywhere on the Internet, so they're harder to avoid and need a more forceful hand.

ETA: tongue in cheek

Access to sex is a major incentive for being economically productive for men. If the distribution of sex becomes more unequal, the marginal utility from additional work or career growth will decrease for men for those under the xth percentile (imagine the derivative of the Lorenz curve; depending on the starting and ending Gini indices, that break even point is around the 65th percentile). Those men will put less effort into moving up the curve through dedicated everyday effort and self-improvement and instead focus on consumption of other goods (video games, porn) and wild unproductive bets that would move them far up the curve in a single fortuitous event (gambling, speculative crypto, GME).

Which means a poorer, less dynamic society. Which is definitely a society that can survive, but I'd prefer a different one.

I would question whether it's really the young and attractive in particular; if you think about the Longhouse discourse, it supposed that it's the older matriarchs who wield power.

A model I've been playing with: imagine society as a graph, with individuals as nodes that create information. Information flows through edges in the graph, with more information flowing through those edges with greater weights. The nodes are labeled with either M or F, about half and half. The F nodes are more highly connected and have edges with greater weights (in this hypothetical, at least), while the M nodes are more sparsely connected and dominated by a single strong edge to an F node.

With this structure, although M and F nodes create approximately the same amount of information, the information that actually reaches any particular node will have passed through far more F nodes than M nodes. And, if you made that graph more highly connected, it makes the ratio even more disproportionate.

To be clearer, I'm suggesting that it's not useful to defend actually-existing capitalism with an appeal to ideal capitalism, just like it's not useful to defend actually-existing communism with an appeal to ideal communism.

The USSR wasn't Real Communism. It suffered from a lack of communism. It was only because wreckers not devoted to true socialist ideals corrupted the system that it failed.

I know a couple actual women in AI (not in the "I can call an LLM through an API" sense but in the "I was a listed author on the GPT-4 paper" sense), and the funny thing is that they despise things like @WomenInAI, who constantly spam them begging for a crumb of interest or validation. Their general attitude seems to be "why bother to spend time with a bunch of wannabe clout chasers whose only exceptional quality is being a mid-tier woman adjacent to tech, when I could instead spend the time helping create a God?" Which is, at the least, an attempt to produce something real.

I wonder how much age plays into it: if older people disproportionately serve on juries, then women would also disproportionately serve, even if serving rate conditioned on age was equal.

This struck me, from the Irish Times article:

In 2009, Irish academics who studied 108 rape trials found that male-dominated juries had the highest conviction rate. There was not a single [rape] conviction in the 17 cases which had female-dominated juries.

What's going on here (assuming that rape convictions aren't so rare that 17 non-convictions are possibly noise)? Maybe when there are men in a group, people fall in line behind a leader, but if there are not, you end up with egalitarian jockeying for social position that manifests as hung juries?

Probably need more actual studies to be able to productively speculate, though.

Many of these loans were taken out with the expectation of securing a valuable job.

Were they?

I sincerely doubt that CS or EE majors are anything but underrepresented in the recipients of these bailouts. Which does point to a third possible party who has at significant share of the blame: universities themselves. They encourage students to take massive loans of dubious benefit with false advertising. More economically advantageous majors are often even limited in enrollment.

I don't hold the student loan debtors too responsible for the issue. They're young idiots who are told by pretty much everyone that it's a good idea to go on what's effectively a four year party vacation because they'll automatically get a good job afterwards. It's akin to someone who doesn't understand how interest rates work taking out a payday loan. But we really should be punishing the institutions that put them in that position.