site banner
Hour Day Week Year All
Hot Bump Top Bottom Old Controversial Comments
Advanced search parameters (with examples): "author:quadnarca", "domain:reddit.com", "over18:true"

Showing 25 of 9554 results for

domain:anarchonomicon.substack.com

Is your "AI Assistant" smarter than an Orangutan? A practical engineering assessment

Well, okay, yes I will grant you that statistical awareness of anything more complicated than a surface level understanding of Pearson's r is more rare than it should be, so I can see that.

Helpful video link of Hinton, definitely illuminates a bit of where you're coming from (and I was pleasantly surprised to see him talk, essentially, about my exact point regarding memory when he talks about "time scale" as something he thinks is an area that could be fruitful; a memory approach the cartridge cache idea, while neat, isn't really, though we'd need to get into an extensive tangent about neuroscience, and likely a little more linear algebra than I'm up to, to tease out why). I will say that genius scientists having major realizations after they are old and their careers are over don't have the greatest track record, but he's still got a point. I think upthread I alluded to a group of scientists who think that what is happening behind the scenes in LLMs is true understanding, and Hinton as I understand it falls into that group (and thus has an above-average P(doom) to use the cliche'd term). But, as I noted then, that group is still the minority view in the field. So I think epistemically it's solidly in the realm of 'reasonable people can disagree'. I lean against, but I'm still a bit agnostic about it. From a statistics perspective, we're in a little bit of uncharted territory, where we have a bit of the scaffolding but the setups have some novel traits, so the lack of total clarity for "higher order" relationships doesn't really surprise me.

I do think you're a little too unkind to self-directed learning. The AR-Zero paper is exclusively looking at if LLMs can, essentially, teach themselves to code if they have access to a Python environment and its output... and already went through pre-training. That's just replacing the RLVR, amplifying a pre-existing skill in a closed, human-built, rules-ruled system, which is impressive as a practical matter but not really what I was getting at. And not very similar what most of the forms of learning we'd want would look like in practice, either.

He also touches on and as you elaborate (I enjoyed reading your thoughts about the numbers), yes there is definitely a difference in parallelism and throughput and metrics like that for brains vs LLMs, and this is part architecture but part the simple decisions of what the 'market' is investing in and exploring. I too wonder what might emerge, or might end up being impossible.

It's interesting to me that you think these differences on the whole aren't very relevant to an (abstract) understanding of intelligence, however. Do you think human perceptions of intelligence are essentially one big conflation because we happen to be "solitary mobile embodied agents" that are also, as a group, quite clearly different in scale and perhaps nature from the rest of the animal kingdom (thus we ascribe too much of this difference to intelligence)? I tend to think that these things cannot be easily divorced; intelligence only ever makes sense to think about in a human philosophical context, rather than a more abstract criterion. As you say, the human form of learning and intelligence is very intrinsic to us, why fight against the natural desire to evaluate it this way?

Culture War Roundup for the week of July 21, 2025

You can override WD, though. You're making the choice not to, because you think it might be malware.

Windows has gotten significantly worse about anti-user features since 10, but you can still run code MS doesn't like AFAIK (can't speak to Win11, as MS is on my hatelist since they started bankrolling OpenAI).

Culture War Roundup for the week of July 21, 2025

Meritocracy is, in some very real sense, "discrimination against dumb people"

And in that same sense, countries that grant their citizens broad liberties and freedoms discriminate against the stupid and virtueless.

"Ruining it for everyone" is the excuse to socialize your private virtue for those people.

Culture War Roundup for the week of July 21, 2025

The correct question is 'What rate of pay would Jeff Bezos and his wife have agreed to in return for her assistance?'

No thanks, Software As A Servuce is bad enough, I'm not entertaining Marriage As A Service.

but that's not actually relevant so long as the court would have forced him to pay her for her labor regardless of the success it engendered -- her compensation was guaranteed, so there should be no risk premium.

It's not about compensation and premiums, it's about ownership.

Small-Scale Question Sunday for July 27, 2025

Free by Lea Ypi, a memoir of the author's childhood in Albania as it transitioned out of the USSR out of socialism and into a liberal democracy. Unsurprisingly, it presents living in a socialist country in a very negative light. Comparisons with My Brilliant Friend are apt (Albania in the late 80s/early 90s seems about as economically deprived as Naples in the 60s), even though this one is marketed as non-fiction. Very readable, and I'm glad the focus is mainly on the politics and the disruptions the author's parents had to cope with, rather than endless trite anecdotes about the author's interactions with her primary school classmates or whatever.

Culture War Roundup for the week of July 21, 2025

then he has a specific threat given that Pinochet was suppressing them

then it could be "I am being persecuted under pretext of being communist", this does not require someone to be a communist

(rather than "I live in a terrible country and it can happen to me")

Culture War Roundup for the week of July 21, 2025

because basically no one walks in the area

oh, I keep forgetting about that USAsian weirdness.

Culture War Roundup for the week of July 21, 2025

Otherwise, we kind of move towards a world where everyone dons a disguise out in public just to maintain some semblance of anonymity.

Perhaps the political valence of wearing a facemask in public spaces will do a complete 180. Or better yet, burqas for everyone, not just women.

Culture War Roundup for the week of July 21, 2025

The labor theory of value is wrong, yes. I think you're missing a step or two between that and the Washington State Divorce Court being the proper way to assess that value. The correct question is 'What rate of pay would Jeff Bezos and his wife have agreed to in return for her assistance?' Which is unfortunately impossible to answer given that no such negotiation took place.

I suppose you could argue that he married her with the understanding that, should they divorce, their assets would be divvied up according to that process? That's technically valid, but it'd be just as valid if that process were anything else, provided those terms wouldn't have prevented their marriage; also impossible to say, I suppose. Still, I think this is the best supported position.

On the other side, one can consider what he'd have had to have paid someone else to fulfill those same responsibilities -- certainly far, far less than he ended up paying her, even if he'd had to take out a loan to do so. It's certainly possible she did something for Amazon no one else could have done, but neither accounts nor packing orders meets that bar. He likely wouldn't have taken out a loan to pay someone else to do those things (at least not very early on), but that's not actually relevant so long as the court would have forced him to pay her for her labor regardless of the success it engendered -- her compensation was guaranteed, so there should be no risk premium. But that's not what the court would do, and they both knew that at the time, so maybe a risk premium is fair.

Small-Scale Question Sunday for July 27, 2025

Eagles are probably Switzerland.

I would say Australia is a very good contender for this. The wedge-tailed eagle has a massive wingspan and length and it is endemic to the Australian continent. They are often seen here and are in fact the most common of the world's large eagles. IIRC Australia also has higher median wealth per adult than Switzerland, though also lower average wealth (I suppose Switzerland's average is pulled up by a small percentage of really high net worth individuals) so I think it fits well here.

A possible runner-up is Japan (probably features third behind Australia and Switzerland because it's not super wealthy, and it represents the edge of the habitat range for the species in question). The Steller's sea eagle is one of the heaviest eagles and can be commonly found overwintering in Hokkaido (they are also found in South Korea and China but in smaller numbers, so depending on your definition of marginal you could count them or not). The actual core of their habitat is in Russia, but that country definitely isn't wealthy.

Big snakes are, I’m guessing, either Malaysia or Singapore.

Singapore definitely wins this, they have the reticulated python. This alone doesn't make them unique - many other countries have large snakes, but what really wins them the title is that they are also very rich.

But I would include Australia before Malaysia in that list. Northern Australia in particular has its fair share of large pythons like the Australian scrub python (which is one of the world's largest pythons, capable of preying on wallabies) and carpet pythons, which can get large: example 1, 2, and 3. Also here is an olive python swallowing a crocodile in Queensland. You're welcome.

I realise this reply is very Australia-heavy but I think people underestimate just how much actually gigantic wildlife there is in the country. They definitely win the "large marsupial" category with red kangaroos, too.

Small-Scale Question Sunday for July 27, 2025
  1. <1 km
  2. 3km
  3. 1km, apples
  4. 5km (not Amtrak)
  5. 4 km (local equivalent)
  6. 220km
Is your "AI Assistant" smarter than an Orangutan? A practical engineering assessment

I find the discussion about intelligence worth having and exploring, to discover the edges and weird applications of current terminology. We're not doing warfare and abusing language on purpose or anything, trying to pick apart both meaning and connotation is partially the point, no? But Google is a poor example, even given some wiggle room about word meaning, and I will continue insisting on that. Assembling Google, as a glorified directory, is not knowledge work, even if you consider the slightly more complex algorithms that go into its modern iterations. Full stop. And I think you'll find you're in the minority if you think humans are "merely information processing algorithms", do you really think this?

There's a bit of overlap, but sometimes in education you might see the word "synthesis" thrown around. It seems to me that "processing" implies analyzing, sorting, and reacting to information, and "synthesis" implies combining, connecting, and distilling different ideas. LLMs are neat because they don't just categorize or recognize, but are able to do recombination and produce novel outputs from a broad context base. Thus my example about how LLMs summarizing a document effectively is a higher-order work task. Maybe that's a slightly more helpful example of what I meant when I said "judgement" is applied to context.

Now, personally and in practice, i.e. outside of philosophical discussions, of course I think it's better just to say "LLMs display jagged intelligence" or "spiky intelligence" or something of that nature. I think it minimizes confusion. Clearly, the implication is that the intelligence is not generally applicable, and has major failure states, and that matches their real-world performance thus far pretty well, and it doesn't raise any excessive linguistic red flags. But I'm still comfortable calling in a qualified type of intelligence, qualification mandatory. Also, partly why I default to calling them LLMs instead of Artificial Intelligence (or at least I try to) despite them being somewhat interchangeable in practice, though I'm not perfectly consistent there. I'm too aware that until ChatGPT, AI meant something different, usually another confusing reference to "deep learning" which in turn was a poor alias for neural networks. Buzzwords, sigh. Anyways, no real test needed, and I view the need for qualification to be at least partially self-evident.

Culture War Roundup for the week of July 21, 2025

"I'm sorry sir. I just have to scan the ID of anyone who looks under 30" (guy looks 17)

Despite the presence of a posted sign that says that we only have to check the ID of anyone who looks under 40, management requires that we check the ID of everyone period. I'm honestly not sure how much of this is that management thinks we're too retarded to estimate people's ages and how much of it is management figuring that it'll offend people (mostly women) for us to estimate that they look over 40 regardless of accuracy.

Culture War Roundup for the week of July 21, 2025

But 'make lots of money' is only imperfectly correlated with 'the company I work for makes lots of money.' And, indeed, the correlations between 'doing my best to make money for my employer' and both 'make lots of money' and 'the company I work for makes lots of money' are very imperfect. In practice, generating maximum value for the company is only really the optimal path for 1. the owners 2. people in roles with very clear metrics (e.g. sales) -- and then only to the extent those metrics can't be more easily gamed, and 3. those with both a great deal of control over the company and a lot of their compensation tied up in stock options/performance bonuses/etc. (i.e. a handful of executives). Some other roles (e.g. security, compliance) have strong incentives not to lose the company an enormous amount of money (in certain specific ways)... Which isn't actually the same thing, as becomes abundantly clear if you ever have to interact with these people: they'd really much rather nothing gets done if it makes the particular sorts of incidents for which they'll be held responsible slightly less likely.

Everyone else is one or more principal-agent problems away from those incentives. Expecting corporations to actually maximize profit is only slightly less naive than expecting command economies to actually optimize for the public good. Their owners want that, but only a tiny minority of the decisions are actually made by the owners, or by executives, or even by directors. The vast majority are made by bottom-level employees and their direct superiors, which, in large companies, are very detached from the company's actual profitability. They'll lose their jobs if it goes under, of course, but it's not like their personal efforts can do a lot to prevent that or bring it about -- there are a lot of these people.

The incentive is to keep your boss happy enough with you and otherwise do as you like, which might mean slacking off, or using your position to push your morality or politics, or maybe even doing a good job for the simple satisfaction of it. But it's a mistake to assume profitability is the overwhelming incentive, or even a particularly strong one, given how difficult it is for the people who really want that to enforce their will over the entire organization.

Small-Scale Question Sunday for July 27, 2025

The Newbery Medal and the similar Carnegie Medal provide lists of children's books decades long which have been critically acclaimed for their quality. As long as your friend stays away from anything after 2010, the titles should be non-woke.

In particular, the early Newbery Medal winners and runner ups (called Newbery Honors) are entering the American public domain. Thirty-one of them are currently there, and another nine will join them on January 1st, 2026. So if your friend is willing to give his niece an e-ink reader, he can just download several of them for free from Project Gutenberg.

Alternatively, he could try a long running series such Animorphs, Goosebumps, or Encyclopedia Brown.

Small-Scale Question Sunday for July 27, 2025

Considering taking the grill-pill: Does anyone here have any experience or recommendations on pellet smokers? I've had some success in the past smoking meat on a basic kettle grill with charcoal, but it's a bit of a pain to setup and baby sit all day. I'm thinking it'd be interesting to get something a bit more turn-key to make it easier to do more frequently, and pellet smokers (Traeger, et al) seem frequently recommended these days for both traditional low-and-slow barbecue and occasionally for other outdoor cooking (pizzas? burgers?). Are they actually as good as advertised? There are tons of models at different price ranges, but which features (and sizes) are actually useful?

Culture War Roundup for the week of July 21, 2025

You're not wrong. I want to argue that viewership by hour is not a good metric because because men usually, from my experience, stream about 1.5 as often as women do in a similar period. Or peak viewership is kinda just decided by twitch itself based on who they put on the front page (and people covering e-sports are going to get that over women who just simply don't cover e-sports). And who knows what the real numbers of any of this are because of how botted everything probably is.

But then there's the twitch payout leaks and they're pretty much the same thing as those lists. 99% men. But "top talent" is pretty reaching, it's just internet ratings, or are we prepared to say that television's top talent is Shonda Rhimes?

I suspect its similar to the amount of people that want to play a male vs female character in a videogame when they have the choice, apparently the vast majority pick male every time. Men are probably the largest demo here and prefer to play as a man and watch men. I remember hearing Northernlion say a few months back that 3% of his viewing audience is female on Youtube.

Though, I don't think it's that hard to have a good number of women to follow on either platform though like I said before they put out less content and also drop out way more often whether retiring, maternity leave, or simply stopping streams apropos of nothing. It's kinda like how women have three set matches and men have five in tennis. If women's matches were five sets then there'd be like five women in the world who would be capable of competing. I have exactly two women I follow who consistently put out content and aren't going offline for weeks or months at a time for maternity, vacations, or mental/physical health breaks or just in general being flaky. And I'm not saying that it'd be better if they did because women are generally better at communicating with the audience and you don't get the summit1g playing a game for 20 minutes of complete silence then dying and saying "aw damn" and going back into the queue in complete silence but maybe you would if they tried as hard as guys do.

EDIT: To give a more concrete example of why I think numbers are botted look at Rifftrax and MST3K in the leaked numbers. For as long as I can remember MST3K had at least 100 more viewers than Rifftrax averaging around 600-700 whereas Rifftrax had 400-500. When the latest MST3K kickstarter happened the numbers went up to 1000 and stayed that way for like six months before dropping back to 600 until this year when they finally dropped below Rifftrax and now the numbers are about the same for Rifftrax and MST3K has about half that. The payout numbers make it seem like Rifftrax is 5x more popular than MST3K and as someone that switches between the channels it's easy to notice that Rifftrax's chat is about 5x faster/more populated than MST3K's and has been even when it had double the amount of viewers.

I don't even think it was nefarious on the channel's part I think someone just wanted to support MST3K by paying for bot viewers. I also wouldn't be surprised if people were paying for bot viewers on Rifftrax as well but it's just been more consistently the same. It's hard to find an apples to apples comparison for what viewers are willing to pay so it'd be hard to make a similar comparison to other types of channels but this is probably as close as you'd get.

Friday Fun Thread for July 25, 2025

Look, I’m not saying there’s anything wrong with Anya Taylor Joy, I’m just saying it’s tacky and historically inaccurate to have an... “Extraterrestrial-American” actress in a movie about Vikings. It’s probably for propaganda purposes, even if she is a good actress.

Friday Fun Thread for July 25, 2025

Yes and no, actually a bit of hard question. Let me say this. Factually, world-building-wise, the world absolutely is not in medieval stasis.

In practice, the theme you allude to is not emphasized, with maybe one major exception. Most of the story and its themes are focused on Frieren's own characterization and experience and thus most of the world-building is more subtle and done in the background (exposition is rare). Frieren has only about three major formative periods of her 1000+ year life that we've seen details for. 1000 years ago we see some more rare flashbacks and people are dressed in Greek/Roman style clothing, so that has an implication there. However, we spend most of the time in the present with some semi-frequent flashbacks all within the last 70 years or so. We also see in the anime's current plot (28 22-minute episodes) mostly rural countryside, too, so it's hard to get a precise bead on tech development, though the source material not yet adapted eventually will show a more advanced nation. The visual vibe is maybe 1400s, we do see some pretty clean and well built out cities.

The one massive exception: tech advancement is not the focus because the show's 'tech' is magic, and magic advancement IS definitely a major theme, dealt with directly. Magic goes from restricted to humans (hoarded by elves and natural to demons, thus a major taboo) to humans leading major advances (even somewhat threatening the elves with their thousands of years of practice), and that comes up and will come up again.

Small-Scale Question Sunday for July 27, 2025

Yeah, on the margins technology (or at least something) is clearly having an effect. Maybe it's some chemical in food or water. Or one of those classic social science multifactorial explanations with 50 different causes.

The 'true human love will always prevail' rhetoric is just a story trope. Those bars, nightclubs and third spaces are still shutting down, are they not? We can see the statistics.

https://old.reddit.com/r/Infographics/comments/1c742ed/percentage_of_americans_married_by_birth_decade/ We can see the statistics on marriage like you say.

AI pornography is the worst it'll ever be.

Culture War Roundup for the week of July 21, 2025

ast Twitter! How often did it turn a profit? Why did these companies keep on getting funding at ridiculous valuations? Maybe it is a way of doing sentiment engineering at scale through various behavior modification tricks with Likes, upvotes, retweets.

TBH I'm kind of inclined to dout that the reddit board as an organism is "smart" enough to do that, except in the broadest sense. Like, with as much data and control as a social media site has, I'm pretty sure I could be way more effective at pushing my own ideosyncratic policies than any existing social media site actually does. Reddit at it's most ruthless just sort of vaguely boosts leftism in the exact same way that tumblr and pre-elon twitter did. Probably because if anyone in particular starts trying to press a view hard, there's too much disagreement on the specifics to get very far. Just imagine a world where, for example, the entire board of higher-ups at facebook were monarchists, including Zuck. They definitely have the power to make monarchism a credible political subcurrent in america... but I think they would sincerely fail to advance the cause of a particular monarch. Zuck would want himself, of course, but members of his board might be crypto-orleanists, or avowed bonapartists. In the process of promoting monarchism more generally, they'd have plenty of latitude to advance their own causes, in the end causing self-interference and averaging out.

Friday Fun Thread for July 25, 2025

royal

America

Wait, why do I hear musket fire in the distance?

Friday Fun Thread for July 25, 2025

Dangit I wish we had royal clubs in America.

Friday Fun Thread for July 25, 2025

Wait I'm confused what's wrong with Anya Taylor-Joy?

Culture War Roundup for the week of June 23, 2025

So a between-sibling GWAS of persons born in WEIRD society and one or both parents are mixed-races would find answer.

It would find a very limited version of the answer. Again, even if you find that the measured relative average IQ difference between groups A and B is caused by genetics, it doesn't necessarily prove that the measured relative average IQ difference between group C versus A and B is due to genetics. Even if you find a result that applies to "nigerian immigrants in america" the selection effects of immigration would invalidate extending the result to "nigerians in nigeria."

Why? Selection eliminates deleterious alleles from population. What constitutes deleterious depends on current environment. So you may find some population where selection for IQ-lowering alleles intensified but selection for bad running (or immune systems) relaxed.

If there's still a single target "golden brain" it doesn't matter how weak or strong IQ selection effects are for it-- every group will aproach it asymptotically over time, though some groups will take longer than other. For IQ to be traded off versus, say, faster running, you need to start thinking in terms of the actual tradeoffs for having big brains-- mainly metabolic, but also head size, injury likelyhood, pregnancy difficulty, etcetera. And when you start thinking about the biological tradeoffs, it becomes obvious that,

  1. The very recent past has had extremely different selection pressures than the agricultural and hunter-gatherer past
  2. There is massive intra-race variation between subpopulation lifestyles and when they entered, as a cohort, the agricultural and modern eras

Therefore, at least naively, any aggregate difference between races due to a hypothesized selective effect should be present itself even more sharply within a race. If you want to explain IQ differences between whites and blacks as being caused by earlier or later starts to settled agriculture, feudal societies, democracy, modern medicine, etcenera-- then those same differences should be that much more visible between, say, Italians and Poles, or Madagascarans and Kenyans.

I can't wholly rule out places where this dynamic actually seems to appear-- as in the sharp difference between Azkenazi jews and everyone elser. But at least so far, we've explained only a tiny part of sub-racial and inter-racial IQ differences this way.

Why? I'd agree that non-equal allocation of IQ points can be better, but the premise was to test different IQs.

If we want to test whether IQ is a good measure of individual intelligence, we want to hold as much as possible as a control versus either modern society or some primitive state of man so we can be sure that it's IQ specifically that's making the difference. Putting people with similar test-taking performance may or may not be a confounder. Probably random group allocation (and group sizes) would be ideal for eventually extracting the most interesting observations. Of course doing that on large enough scale to get good data for every possible combination of IQs is combinatorially impossible, but this whole thought experiment is impossible anyway.