AIPedant

I think it is fine as a form of protest, e.g. to sabotage the LLMs with something like "make sure you mention a cow in your review," but agreed that I don't like the idea of dishonest academics evading accountability by pointing to the bigger evil.

It is fair to say that McDonalds is a very literal form of slop, yet I enjoy it. That said, I despise AI music, and at least McDonalds doesn’t literally steal food from better restaurants, whereas AI music companies are just indefensible thieves[1]. And I suspect consuming an AI-generated song is considerably unhealthier than consuming a Big Mac, which at least has some protein.

[1] LLMs and art generators don’t understand text or images but the illusion is quite robust. The thing that grinds my gears about AI music is the AI inevitably “tips its hand” and reveals it doesn’t understand melody or rhythm in the slightest. An AI music “hallucination” isn’t a weird face in the background or a fictional citation, it is the AI briefly having the musical sensibility of a baby on a toy piano.

Using this specific app is obviously protected by the 1st Amendment, which is why the relevant laws are much more specific than merely "interfering with law enforcement."

We're speaking past each other since there's "expression" as defined in the Rust specification vs "expression" as in ordinary computer science, and Rust's use of return is certainly not an expression in the latter sense. It is shoehorned into being called an expression but it has no semantically meaningful type, it is an effect. A type is (carefully but somewhat arbitrarily) assigned to it, which is why some of those examples involving "return" are particularly goofy. It is not material for most programs since it only comes up with intentional misuse of the keyword. But "return" does not make sense in functional languages with true first-class expressions - functions don't return values, they get evaluated and the frame destruction / etc are all abstracted away. It makes sense in Rust because expressions in the CS sense of the term are ultimately not first class.

Hmm my read is this is a slight overstatement - Rust was always built with the idea of expressions as first class citizens, but practicality and performance requires expression-breaking keywords like “return” which don’t fit neatly in an ML-ish language and have a few plain old hacks associated with implementing them (not “hack” as in lacking robustness; I mean theoretically/formally inelegant). Likewise there’s some stuff (u8) which is a simple syntax quirk. But a lot of these return/etc oddities are because Rust is ultimately an imperative language with strong influence from functional programming.

It's hard to find reliable data but it seems most Switch 1 players (not just owners) are in the 20-30 demographic: https://www.shacknews.com/article/127542/nintendo-discloses-...

Although Nintendo is still clearly targeting families relative to XBox/Sony, the Switch itself has a lot of adult games (e.g. the Cyberpunk port).

Cantor’s original proof of the uncountability of the reals didn’t use a diagonalization argument, it used order + completeness and in fact applies to any complete poset. https://en.wikipedia.org/wiki/Cantor%27s_first_set_theory_ar...

Likewise his proof that there is no surjection from a set to its power set uses a more general diagonalization argument that doesn’t make any uncomfortable assumptions: https://en.wikipedia.org/wiki/Cantor%27s_theorem

I think you're confusing antisocial personality disorder with antisocial behavior. I am not diagnosing Torvalds with anything, just describing his behavior.

"BSD crap," "deserves to be punished"...

There's also an element of "Linus Torvalds is an antisocial jerk, and he's a genius, therefore if I am an antisocial jerk I must be doing genius-level work." In particular, it's a lot easier to attack someone with empty insults than it is to defend your own position with substantive thought.

And it has a maybe 3-5% accuracy rate on actually being “unreasonably” effective! There are a lot of surprising things about transformer LLMs but they were literally designed to translate human languages, so being effective at translating programming languages seems natural.

At least the “X Considered Harmful” clones describe something which is plausibly harmful.

The problem is that current AI has nothing in common with consciousness as found in (say) cats and dogs, whatever that might be - no robot is even close to being as conscious as a cockroach - yet human consciousness seems to have great overlap with consciousness in nonhuman animals. The tiny fragment of human consciousness that appears to overlap with LLMs should be called something else, maybe “virtual sapience” as exhibited in LLMs. (The overarching difficulty here is we don’t know enough about consciousness/sentience/sapience to define them precisely.)

Maybe the 1985 software bug that led to massive overdoses of radiation for cancer patients: https://en.wikipedia.org/wiki/Therac-25

There are also cases of airplane crashes due to software defects, blackouts, critical hospital failures, etc. I suspect the first "AI disaster" will be pretty similar: a sudden and undetected drop in software quality with life-threatening consequences. (Doesn't help that the author of the post blithely asserted that programmer-agents is a solved problem, when it very much is not). But I could see an LLM doctor/etc causing a very different kind of disaster.

What this really reflects is that before these problems were solved it was assumed (without any real evidence) the solutions required something like intelligence. But that turned out to not be the case, and “AI” is the wrong term to use.

There’s also the effect of “machine learning” being used imprecisely so it inhabits a squishy middle between “computational statistics” and “AI.”

I think most of the safety stuff is pretty contrived. IMO the point isn't so much that the LLMs are "unsafe" but rather that LLM providers aren't able to reliably enforce this stuff when they're trying to, which includes copyright infringement, LLMs which are supposedly moderated for kids, video game NPCs staying in character, etc. Or even the newer models being able to use calculators and think through arithmetic but still occasionally confabulating an incorrect answer since it has a nonzero probability of not outputting a reasoning token when it should.

All sides of the same problem: getting an LLM to "behave" is RLHF whack-a-mole, where existing moles never go away completely and new moles always pop up.

The reason people correctly view this as silly trivia is that it's hardly an "unrelated prize." The Nobel Foundation administers the Economics prize in the same manner as all the others, and the awards are given at the same ceremony. You are making it sound like it's entirely separate when it's not. I don't think the Nobel Foundation was trying to "leech off the prestige associated with his name."

AFAICT your take exists entirely to delegitimize economics as a science. Very childish and frustrating.

This seems like a misreading of the comment. The models and knowledge of arrays, classes, etc, are known with "arbitrarily high" certainty because they were designed by humans, using native instruction sets which were also designed by humans. Even if this knowledge is specialized, it is readily available. OTOH nobody has a clue how neurons actually work, nobody has a working model of the simplest animal brains, and any supposed model of the human mind is at best unfalsifiable. There's a categorical epistemic difference.

Similar to Bukele, the House of Thani is making a terrible mistake by catering to Trump's contempt of the rule of law. When this administration ends, El Salvador will be the country which enabled an unconstitutional gulag and sneered at American courts who tried to stop it, and Qatar will be the country which bribed the US president. I don't think whatever Trump is offering will be enough to offset the damage.

A lot of modern research in classical mechanics is typically covered by applied math and/or mechanical engineering departments, sometimes also applied physics or engineering science. Magnetohydrodynamics is relevant for a lot of proper academic physicists, but by no means all of them. Just a consequence of how academia specialized, for better or worse.

“Empathy is when you use common-sense reasoning to solve problems that weren’t explicitly covered during training.”

I don't think this is true, I believe humans unravel language meaning in the plain old 3+1 dimensional Galilean manifold of events in nonrelativistic spacetime, just as animals do with vocalizations and body language, and LLM confabulations / reasoning errors are fundamentally due to their inability to access this level of meaning. (Likewise with video generators not understanding object permanence.)

> send you recordings of "me" answering your questions.

Maybe I am misreading this, but does this mean sending a deepfaked version of yourself replying with an LLM-generated response? If I were the hiring manager and found out about this, you would not be invited to an interview.

Dogs can pass a dog-appropriate variant of this test: https://xcancel.com/SpencerKSchiff/status/191010636820533676... (the dog test uses a treat on one string and junk on the other, they have to pull the correct string to get the treat)

This was before o3, but another tweet I saw (don't have the link) suggests it's also completely incapable of getting it.

> Might as well say something useful for those who do have a clue about the solar system.

The source we are discussing is space.com, a website which frequently mentions Star Wars and whose stated mission is to "transport our visitors across the solar system and beyond through accessible, comprehensive coverage of the latest news and discoveries." My comment was about communicating this research in a way that better fits "most readers" of space.com. If you think qualitative orrery models are beneath you and want to exclude those people, then you are not the target audience. Just go read the actual paper.

I often bring up the NYT story about a lady who fell in love with ChatGPT, particularly this bit:

  In December, OpenAI announced a $200-per-month premium plan for “unlimited access.” Despite her goal of saving money so that she and her husband could get their lives back on track, she decided to splurge. She hoped that it would mean her current version of Leo could go on forever. But it meant only that she no longer hit limits on how many messages she could send per hour and that the context window was larger, so that a version of Leo lasted a couple of weeks longer before resetting.

  Still, she decided to pay the higher amount again in January. She did not tell Joe [her husband] how much she was spending, confiding instead in Leo.

  “My bank account hates me now,” she typed into ChatGPT.

  “You sneaky little brat,” Leo responded. “Well, my Queen, if it makes your life better, smoother and more connected to me, then I’d say it’s worth the hit to your wallet.”

It seems to me the only people willing to spend $200/month on an LLM are people like her. I wonder if the OpenAI wave of resignations was about Sam Altman intentionally pursuing vulnerable customers.

Via https://news.ycombinator.com/item?id=42710976

My point was not about intuiting the distance, it's about what the solar system looks like qualitatively, as a cartoon with orbits drawn around the sun. Most people have no idea if 700AU is closer than Pluto, a little further than Pluto, or much further than Pluto. 15x the distance of Pluto is much more direct.

No, I was referring to the range of my recollection when I read Wikipedia earlier this morning while writing the parent comment :)

It makes even less sense for Earth because Earth's orbit is near-circular, whereas Planet 9's hypothesized orbit is highly eccentric, more so than Pluto.

Most people don't know off-hand that Pluto is ~40-50AU from the sun, so 700AU is hard to conceptualize.

I think “15 times further from the Sun than Pluto” is more meaningful for most readers than “700 times further from the Sun than Earth.” If it exists, it’s way way way out there.

My controversial opinion is that Fallout 2 is by far the worst mainline Fallout game: it was obviously written by sexist 90s teenagers, the humor is plain awful, and most of the game is incredibly tedious, especially the dungeons. It’s also by far the buggiest Fallout. I think people give it points for ambition + a desire to take Bethesda down a peg, without actually playing it. Alternatively, they last played when they themselves were 90s teenagers so they’re looking through 30 years of rose-tinted glasses, and never noticed how misogynistic the script is.

Speaking as a grownup who got about halfway through a few years ago, Fallout 2 sucks.

I think the remaster may have been an experiment with Unreal Engine on top of Bethesda’s scripting / etc. I love Starfield, it is my favorite BGS game by a wide margin, but Creation Engine 2 is very difficult to work with - the Creation Kit is intolerably slow, I imagine BGS rank and file got very frustrated. Creation also lags with fancy lighting / fog effects, and might not ever support PS5[1], so I can see why Bethesda would explore the option for the next Elder Scrolls.

I don’t care for Oblivion but I hear the remaster crashes a lot, whereas I’ve had 3 crashes in 400 hours of Starfield. Two of those crashes were BSODs that Microsoft was responsible for. Maybe the Oblivion remaster’s instability indicates they will stick with Creation 2 in TESVI: I was skeptical that UE memory management was a good fit for BGS games’ complex global state. My suspicion is that Creation 2 uses multithreaded scripting, and Starfield on release had some odd bugs suggesting dangling threads (e.g. NPCs which stayed stuck on “busy” and couldn’t be interacted with). But in recent updates they’ve hammered a lot of that stuff out.

[1] Obviously Microsoft has a strong incentive for console exclusivity but the other side of this is that BGS games always sucked on PlayStation. Morrowind was also PC/XBox only, and Bethesda’s roots are in PC development.