@mjd I guess the pasted sqlite output with the vertical bars confused something. Anyway, doc/fresh. Its explanation is "file versioning slang: doc=an old saved document version; fresh=a new unsaved draft; test='he opened the [doc / fresh] on his desktop'".

(Claude found ant/dec but not the very apt ant/syn. But I don't think of Ant and Dec as opposites anyway, although I guess I've never seen their show, so maybe?)

There's the sort of typical LLM losing-the-thread stuff:
(loo/dry) Indian meteorology: loo=hot dry wind of the subcontinent; dry=... actually reconsidering — domain: card gameloo: loo=lose a hand and pay a forfeit; win=succeed in taking tricks without penalty; test='she managed to [loo / win] on the final hand

I think it might be worth reconsidering the antonym-of-a-saying model, because that's got a small enough surface that a human could review the individual decisions. But if you want to play with the cycles approach, (no need to let my tokens go to waste](https://novalis.org/antonym.tgz). This one was 100% vibecoded, so probably garbage.

@mjd It's definitely partly Claude -- it didn't find dry/Dutch, or pass/bogart or leg/car or leg/stac (admittedly, that should be "leg./stac.", but I still think it's valid). OED even offers leg/lose (that is, leg as a variant of "lag" meaning to catch or apprehend). And it has some funny ideas about slang -- or at least, I've never heard:

I can't really blame Claude here -- it's hard for humans to think of these. Claude's big advantage is that it can consider every word in the dictionary without getting bored.

re: unspecified horror

@mjd So far this producing only boring results, but I'll let Claude continue to churn for another mtok or two and we'll see. From Matt Levine today, I learned "fade", which is apparently the antonym of "chase" in the sense of following a market trend.

@karen When I play Patchwork with my friend with aphantasia, I let him test-fit pieces on his board before choosing. I guess nobody does as much lookahead in Patchwork as good chess players do; that could end up pretty confusing.

I also think good chess players consider so many moves so deep that just moving the pieces would get unwieldy -- and you would still need the memory, because you would need to remember what move started a given line. This paper says that visual working memory is different from visualization, which seems weird, but on reflection maybe isn't. There's that famous study about how good chess players do better at remembering chess positions that could occur in actual games but not better at remembering truly random positions. (Or at least that's the popular description of the study; the original seems to be in Dutch so I can't easily check). Which implies that what's going on isn't just visualization, but chunking.

@mjd @byorgey @shapr That one sentence is definitely enough to get me check the book out from the library with no further research.

re: unspecified horror

@mjd I agree, but I think that there are some useful exceptions. I would not want to use a LLM to solve a crossword puzzle or to write a letter to a lover or a subordinate, or a boss that I wanted to keep.

I have often thought about having an anonymous blog just for unpopular opinions. If anyone tried to associate me with it, I would deny it strenuously. But I am too lazy/cautious for that.

re: unspecified horror

@mjd I read all your articles (except the math ones that are above my level). But I am also really conscious that folks are (rightly) nervous about LLMs sucking the joy out of everything, so I want to be extra careful.

I think "square" vs "cat" is weak but plausible -- "square" is attested as a noun meaning a square person (aside: in Hungarian, the word "kocka" meaning "cube" is used to mean "nerd").

"Innocence is logarithm" is one Claude and I came up with together.

I think it might require a fair number of tokens to produce the antonym graph, but I'll see if it's doable.

@mjd "Square as a peach"? (Emoji anatomical connotations)

@danluu I would be curious to see your benchmarks on the the "Dan Luu" prompt vs others :)

re: The opposite of “full” is…

@mjd I had Claude come up with eight examples. The best, by far was "certainty bombed the square". (Claude agreed with this). The second-best is "poverty penalizes the italic", and the rest were, IMO, not clean (e.g. "patch" as the opposite of "worm", because code patches prevent malware worms).

(Let me know if you think using a LLM sucks the joy out of the game. I asked Claude in part because I wanted to see how well it would do, which is, better than I expected but requiring significant human curation. But I spend a lot of time thinking about that cartoon with the robots doing art while humans bike around delivering food.)

@bkuhn @wouter @mjw @collinfunk @pinche_compinche This is why man strstr gives the arguments as haystack and needle instead of trying to use the word target.

re: The opposite of “full” is…

@mjd If I get a card with "full" on one side and have to pick between "empty", "self", and "new", it's going to just be guess. But maybe if I had a sentence -- "the drop is self today", and had to be the first reverse it into "the moon is full tonight", maybe?

@mogul First of May, first of May, outdoor fucking starts today!

@mjd There's also Liberia, in Costa Rica. I was once very confused when I saw an airline offering flights to bunch of Central American locations and one African one, before I looked it up. Also, I think a Saturday Times Crossword once had a clue about where one would find Angola and Brazil.

re: The opposite of “full” is…

@mjd Or "new", if you're looking at the moon.

@mjd @christianp I heard a suggestion that the stops on the 6 train be used to remember degrees Celcius -- 33rd St is ~0 C, 42nd is ~5C, 50th is ~10C, etc.

@johns Make sure you photograph all of the identifying numbers on the packaging as well as your receipt before you notify the FTC.

@crystalvisits @jeridansky Orange is the new banana.

@lumi I frequently say that my wife and I are in a mixed marriage: I use emacs and she uses vi.

@bkuhn @zacchiro @cwebber @ossguy @richardfontana Actually, I just thought of one proprietary software company that would be much happier not to have LLMs around: Salesforce. Nobody's going to buy their overpriced shit when the alternative is to vibecode something that works exactly with your business process and that you can change, yourself, any time you want at the cost of a couple hundred bucks of Claude and a few hours of work.

@bkuhn @zacchiro @cwebber @ossguy @richardfontana I don't even know if I agree with my supporting arguments. But I don't even think that it has to be someone in the proprietary world that brings a lawsuit -- it could be anyone whose code or text is trained on.

@zacchiro @cwebber @bkuhn @ossguy @richardfontana I would say it's dramatically less safe. First, there's very little incentive to go after some OSS project over an unauthorized inbound=outbound contribution. Second, if someone did, the damage would likely be a small part of a single project. Third, only a small number of parties (the employer, or maybe some other single party whose code was copied) have the ability to sue.

With LLMs, it's different. When the authors sued Anthropic, they all sued. Is a shell script that Claude generated a derivative work of, say, the romantasy novel A Court of Thorns and Roses (to pick a random thing included in Anthropic's training set)? Well, it's hard to show that it's not, in the sense that that novel is one of the zillion things that went into generating the weights that generated the shell script.

Now it happens that the authors sued Anthropic (and settled). But I don't know if their settlement covers users of Claude (and even if it did, there are two other big models). And that's only the book authors -- there's still all of the code authors in the world.

So yes, I think the risk is high. I mean, in some sense -- in another sense, it seems unlikely that Congress would say, "sorry, LLMs as code generators are toast because of some century-old laws". At most, they would set up a statutory licensing scheme for LLM providers which covers LLM outputs. Of course, Europe might go a different way, but I think they would probably do the same. Under this hypothetical scheme, if your code were used to train Claude, you would get a buck or two in the mail every year. Authors got I think $3k per book as a one-time payment, but that was a funny case because of how Anthropic got access to the books.

Still, there's a risk that Congress wouldn't act (due to standard US government dysfunction).

It seems like most people are willing to take this risk, which I think says something interesting about most people's moral intuitions.

re: genai. ethical harms. bit rambly

@lumi @bkuhn @ossguy @mastodonmigration I have always been in favor of a narrow definition of Free Software -- that is, I think it means software that respects the four freedoms. A piece of Free Software could be bad for other reasons. Bitcoin comes to mind as being unnecessarily bad for the environment. Perhaps software useful only to send spam. Or (hypothetically) software made with enslaved labor.

re: genai. ethical harms

@lumi @ossguy @bkuhn @mastodonmigration Right, that's the car analogy: cars aren't sustainable.

(If you're asking whether it genuinely helps, I would encourage you to look at what other experienced programmers you respect are saying -- in particular, I think @mjd is worth listening to, as he is one of the best programmers I personally know).

But also, unfortunately, it seems really unlikely that we will manage to outlaw either cars or LLMs.

re: genai. ethical harms

@lumi @bkuhn @mastodonmigration @ossguy GenAI has a case where it's useful: producing small software when you don't know how to write code. To my mind, this is a software freedom issue: what use is a pile of source code that you don't know how to modify? Sure, you could hire someone (if you're rich).

It also seems to (since November) be sometimes able to help experienced practitioners produce software faster than they otherwise would be able to -- especially in areas where they are unfamiliar with the ecosystem. You may or may not believe that this justifies the harm, but it is a use-case.

Finally, one weird-ass use-case which I admit is niche: I use it to remove ads from podcasts. Imagine doing that like 90s spam filtering, with a pile of regexps. Yuck. LLMs (while not perfect at the job) make it straightforward. My kid is much happier not listening to ads.

re: genai. ethical harms

@bkuhn @mastodonmigration @lumi @ossguy Maybe cars would be a better analogy: expensive, environmentally destructive, makes your body atrophy, but damn they're convenient.

@mjd @skewray

A: X is wrong, and you should feel bad about doing it.

B: I did X because of the incentives.

A: By telling you to feel bad, I am trying to change the incentives.

@intransitivelie Charles De Gaulle? Or Pierre Mendes France?

@lahosken A thing I like about this is the graphic design.

@mjd @robinhouston I think there's still a problem with long spiky caltrop shapes.

@mjd @robinhouston Those are called "polyiamonds", which must have been what was meant; I think it's false because of polyiamonds with holes.

@crystalvisits I once went to a podiatrist for a sprained ankle. Sure, there was an elevator to his office. But to get to the elevator, you had to go down a flight of stairs. It was a very wide flight of stairs -- 4 or 5 meters wide. They could have put in a lift. A podiatrist.

@attoparsec @beka_valentine I think I'll play the one from The Expanse (content notice: much gore).

re: irritated note on history of U.S. aggression

@mjd George H. Bush had to settle for Panama.

@lahosken I had a similar idea

@smellsofbikes @MLE_online That brick one is incredible.

@jefftk Is that the amount with or without a bottle scraper?

@waltman "Before we get started, I would like to acknowledge that I am recording today on the traditional, ancestral, and unceded territory of the Coast Salish peoples, including the territories of the Musqueam, Squamish, and Tsleil-Waututh nations, whose historical relationship with the land continues today."

@lahosken @adrianhon When I clean my cat's litter box, I always tell them to put their dukes up -- because I'm boxing.

@mjd @choroba Also named, indirectly, after the same guy: the Wistar rat.

@rose_alibi Houston Street (pronounced House-ton).

@lahosken I think -ler should be -er.

@smellsofbikes @MLE_online The ones from mangozz.com are usually good, if stupidly expensive. I always get two-day shipping (by waiting until there's a sale that offers it).

@smellsofbikes @MLE_online This is still what it's like with Alphonso and Kesar mangos are like -- they're only in season a couple of months each year. When they come back, the kid and I go through a dozen a week.

@petersanchez @icy @tedu @knapjack@bonk.cozysumo.space OK, finally caught up on this. I was unhappy with The Rose Fields. It ended in the middle of nowhere. Also, the protagonists were nowhere near the actual plot, which seemed to have something to do with money, except that Pullman doesn't care about money and thus doesn't actually have anything to say about it. Oh well.

re: semantle #1483 spoiler

@rhiannonstone Nice!

@jmac @jplebreton I think that's Raymond Queneau (describing Oulipo): "rats who construct the labyrinth from which they plan to escape."

I often think of that line in the context of techical debt in software engineering: "rats who construct the sinking ship from which they plan to escape".

@mjd @dpiponi The sections of map on a globe are also called gores.

@LivInTheLookingGlass If you can cheese the shapeshifting, maybe? Like, a redwood is organic; if you lean a bit, transform into a redwood (instantly extending your length, but at an angle, so you're falling), then, when you're about to hit the ground, transform into a flea so you don't get hurt?

@bkuhn @jzb @kees There's always the classic Justin B Rye taxonomy.

@mjd @luksfarris I think of multiplication as coming from a generalization of skip-counting. Like, you're trying to inventory a bunch of stuff, so instead of counting, one amphora, two amphorae, etc, you count two, four, six, eight, ten makes a boxful, one box, two boxes, etc.

You need formalized multiplication when you have enough (fungible) stuff that counting becomes hard.

You get division in a similar way -- if you have those 42 squirrels and six people, you deal them out into six piles one-at-a-time (keeping track of where you started in case of remainders). But I think most ways of formalizing it rely on already having multiplication (at least doubling).

@jessamyn @timrichards In NYC, we have a 311 app for things like this. Reporting graffiti did get it cleaned up. But I've used it to report cops illegally parking on the sidewalk over a thousand times (literally); as you can imagine, nothing changes.

@luis_in_brief Sure, but I think it's more important to fix unrepresentative images than to add new images, especially for the case I mentioned, where such an image would be pretty meaningless. What does an assistant professor look like? Well, about like anyone else -- it's not like they molt once they hit full professor.

@luis_in_brief How do you feel about cases like Assistant professor, which has no images at all? (It's listed under Associate Professor on your site, but the actual text refers to assistant rather than associate. My inclination is to mark them as has_diverse_images (the empty set contains multitudes?) but would like your take.

@mjd And his wife.

@ranjit @futzle Red pandas didn't exist when we were in school. They were invented in conjunction with Firefox browser in 2004, to be its mascot.

@ranjit @aubilenon I always confuse orecchiette with a hypothetical orchiette (which I assume they sometimes call Apennine Mountain Oysters).

@mjd Inspired by Melvil Dui?

@mjd @talexb @Shanmonster "I just dropped in to see what condition my condition was in."

@attoparsec Crazy idea, but what if instead you did something that took 8 or 9 clock cycles, considering one input at a time? So, you would use a clock divider to choose one input into two 4-input AND gates into a single adder (with its previous output as the other input; initial state set from self). I don't know if the necessary discrete logic components are reasonable, or if this is too complicated or won't work for some other reason, but it seems like it could used fewer chips.

@ranjit Sure, you would send it to the IRS (Island Revenue Serviette).

@dillyd Only if you are careful about cross-contamination. When I scoop baking powder, I use a clean spoon that I have dishwashed since I last used it for scooping flour. If you just use whatever spoon, then you've probably cross-contaminated something. Similarly, if you used a butter knife to spread butter onto your wheat bread, and then gone back for more butter with the same knife, and now you're going to use the rest of that butter for the cookies, you've got a problem. But if you use fresh butter, or if you are careful, then you'll be OK.

@attoparsec @Extelec There's a lot of climatic inertia. If we take dates 3 months apart, here in NYC, the coldest part of the year is Dec 6 / March 6 (by highs) and Dec 13 / March 14 (by lows). Data from 1990-2020, by eyeball. The warmest part is about opposite that, conveniently. So starting on the equinox is way closer than centering on it would be. Of course, climates do vary; Dominus notes that the four-season model doesn't really describe where he lives as well as a ten-season model would... or would have, anyway.

@ranjit @nash @ira here you go

@bkuhn @liw I've solved this problem by getting a heating pad, which I put on my office chair. Even though it's resistive heating (vs a heat pump), it's only heating a tiny area, so it's way more energy efficient than heating the whole house.

@NanoRaptor x86 assembly for me, but overclocking just requires setting the caffeine jumper.

@richardfontana @bkuhn @Ember @mason To this day, there's a cat breed which includes the term (as a shortening of "raccoon"). I have two, but have to say "like raccoon" every time I mention the breed.