rich @theresrich.bsky.social · 6hI've always seen mathematics as both invention and discovery. You invent when you define some new structure, and then you discover its properties when you prove theorems about it. 000 0
rich @theresrich.bsky.social · 15/10/2026I feel like people are treating this "LLMs always choose seven" thing like it's way more interesting or significant a finding than it seems to me. Am I missing something? 010 0
rich @theresrich.bsky.social · 14/10/2026not reading anything on the other, flat out refusing to try the thing for yourself, and then wading into arguments with bold, definitive proclamations on the topic, citing your intentionally one-sided reading as your reason for belief, is almost a *cartoonish* demonstration of motivated reasoning. 120 0
rich @theresrich.bsky.social · 14/10/2026So your primary callouts, when asked why you believe what you believe, are a book from 50 years ago, and an article that many, many people, including experts in the field, have provided refutations for. It seems to me like reading a small amount of literature on one side of an argument, ... 130 0
rich @theresrich.bsky.social · 14/10/2026It seems insane that Google are still pushing this thing front and centre on the top of their flagship product 000 0
rich @theresrich.bsky.social · 13/10/2026holy shit, basil and tomatoes can you imagine being the first person to try those together 020 0
rich @theresrich.bsky.social · 13/10/2026It's unclear which of "they always do what they're told", "no human would ever tell them to do anything terrible" or "they'll never have the capability to do anything really destructive" this lazy screenshot is trying to argue, but all of them seem pretty clearly false to me. 010 0
rich @theresrich.bsky.social · 13/10/2026It is wild to think "most superintelligent AIs will just sit around solving math problems, and only a small fraction of them will kill us all" is supposed to be comforting. 100 0
rich @theresrich.bsky.social · 12/10/2026There are *plenty* of people who would be willing to set LLMs on dangerous, human-harming tasks. 130 0
rich @theresrich.bsky.social · 11/10/2026But those were gullible people who didn't know anything about the capabilities of computers at the time. Now even sophisticated, informed people who are aware of the existence of LLMs might still be fooled. 290 0
rich @theresrich.bsky.social · 11/10/2026Okay, sorry, what I really meant was "there is a premise that your argument hinges upon that you didn't explicate, and you didn't offer any support for it, that's why your claim falls down". When I said "you need to", I meant to imply, "if you want this to not be just a hollow, baseless assertion". 100 0
rich @theresrich.bsky.social · 11/10/2026You need to support your unspoken premise that "running math" is not sufficient for thinking or demonstrating creativity. 130 0
rich @theresrich.bsky.social · 10/10/2026Not a perfect match, but The Message From Deep Space has you learning alien maths and grammar, and if you enjoyed Baba Is You there's a good chance you'll like it. It also has that structure of being many small puzzles, so you can open it, solve a couple, come back tomorrow. Short-session-viable. 000 0
rich @theresrich.bsky.social · 08/10/2026Not me, but I recently saw a post by someone (can't find it now) about some song with a line "He danced 'til his feet got sore" but they misheard it as "He danced 'til the street got sore" which is such an incredible line! and then when they found out the real lyric they were terribly disappointed 000 0
rich @theresrich.bsky.social · 07/10/2026It's disgusting. Like, putting aside however you feel about AI figuring this out or whatever, just the fact that the *universe itself* allows this kind of bullshit is so gross. 000 0
rich @theresrich.bsky.social · 07/10/2026do neutrinos always have to go fast, like photons? Or could there be a slow neutrino, just like, ambling along, that happens to be going roughly the same direction as the earth's surface at some point in time, and then instead of it flying through me, I walk right (through? around?) it? 000 0
rich @theresrich.bsky.social · 07/10/2026"everyone can do science on the miracle machine" is a nice idea, but how do you stop North Korea making a copy of the miracle machine and asking it to hack and disrupt power grids in the west, or optimise their nuclear weapon program, or engineer superviruses only they have the vaccines for? 000 0
rich @theresrich.bsky.social · 06/10/2026I trust a human assistant not to do catastrophically stupid things substantially more than I trust a contemporary LLM not to. 100 0
rich @theresrich.bsky.social · 05/10/2026What if someone makes a second LLM and tells it "keep coming up with ideas for new improvements, and prompting the other LLM to build them" What if someone writes a while true loop that just says "great work, what can we build next" What if the LLM creates that loop for itself 010 0
rich @theresrich.bsky.social · 04/10/2026I mean I agree, but this seems pretty toothless unless we can figure out how you *would* tell if a model is conscious. We can barely even describe what being conscious *is*. 010 0
rich @theresrich.bsky.social · 04/10/2026Before anyone is allowed to express an opinion about whether AIs think, understand things, pose x-risk, etc, they should be required to state upfront whether they personally believe in souls, consciousness-is-magic, or any of the other versions of that stuff. Imagine how much time could be saved! 010 0
rich @theresrich.bsky.social · 04/10/2026We have concrete proof that Einstein-level intelligence can run on 2kg of matter at 20W, and that's on hardware/software that was designed by something literally unintelligent that had to be cobbled together out of biochemistry. It seems plausible the upper bounds on intelligence are likely higher? 010 0
rich @theresrich.bsky.social · 03/10/2026What do you think are the low-probability events that are necessary steps? 000 0
rich @theresrich.bsky.social · 03/10/2026Do you have any, like, good faith arguments to make on this or are you just here to do bad faith sniping from the sidelines 000 0
rich @theresrich.bsky.social · 03/10/2026Okay so just to clarify on this, is your argument here that the HuggingFace incident only happened because OpenAI failed to implement proper security to isolate the LLMs? 000 0
rich @theresrich.bsky.social · 02/10/2026Can you give some examples? I think people frequently misinterpret what I would consider "demonstrations of plausibility" into "concrete future predictions". 100 0
rich @theresrich.bsky.social · 02/10/2026> Yud's overarching argument for an S.I. apocalypse is invalid because it presupposes this predetermined destruction into how it defines "thinking" as both "steering" and "optimizing". I cannot make sense of this statement. Can you rephrase it? 100 0
rich @theresrich.bsky.social · 02/10/2026You can draw arbitrary lines in the definitional sand about whether that constitutes "thinking" or whatever, but it's clearly close enough to thinking for practical purposes. All you're doing is making it difficult to communicate about a thing we can all see is happening with our own eyes. 100 0
rich @theresrich.bsky.social · 02/10/2026If they don't "think" or "have intention", then they certainly have some kind of complex problem solving capability, directed towards a goal that is stated in advance, in English, that is *highly effective at producing a result in which that goal has been achieved*. 200 0
rich @theresrich.bsky.social · 02/10/2026I can go to any news website to read about the HF hack and see evidence of goal-directed, problem-solving AIs violating what they understand to be the humans' intent right now. Should we just wait until AIs have killed all of humanity a few dozen times so we have a statistically significant sample? 100 0
rich @theresrich.bsky.social · 02/10/2026Where are they? Literally, link me to them. I saw exactly one person post a link to a youtube playlist, which I will look at when I have time, and other than that all I've seen is just the same old "why should I" excuses. 100 0
rich @theresrich.bsky.social · 02/10/2026This little "son" thing where you try to come off as condescending is cute. But you're apparently either unwilling or, as is steadily becoming my leading hypothesis, actually incapable of engaging in a genuine good faith discussion, so I think there's not much point in continuing. 100 0
rich @theresrich.bsky.social · 02/10/2026Whew, a narrow escape there! You almost got tricked into actually thinking about a challenging idea, but luckily you managed to pull out one of the same lazy, thoughtless excuses everyone else does, just in the nick of time. Close one! 100 0
rich @theresrich.bsky.social · 02/10/2026If you find someone who earnestly believes in the time cube and will present a good faith argument for it, then I'll address them. Many people earnestly believe AI poses a significant existential risk to humanity, and can and have presented good faith arguments for it. Can you address them? 201 0
rich @theresrich.bsky.social · 02/10/2026If the arguments weren't valid, someone would be able to explain why not. If you can't explain why you think they're not valid, I don't know why you are even bothering to engage in this. I cannot possibly make it any clearer that just blithely asserting that they're not valid isn't an argument. 200 0
rich @theresrich.bsky.social · 02/10/2026No, because I've never seen a valid argument for why that was plausible, that hinged on premises I found reasonable. I have seen such a thing for AI x-risk, though. And what I haven't seen is anyone who can make an actually compelling case for why the argument is invalid or its premises false. 100 0
rich @theresrich.bsky.social · 02/10/2026It's a foolish man who stands around saying "IF it's termites 🙄", waiting for his house to collapse before he'll do anything about the sawdust slowly piling up under the beams. 200 0
rich @theresrich.bsky.social · 02/10/2026I want to clarify. I don't care in the slightest what anyone's opinion of Eliezer Yudkowsky is. That simply doesn't matter. What I'm worried about is that AI poses a plausible x-risk, and people are using "Eliezer Yudkowsky & his friends are weirdos" as an *excuse not to take AI x-risk seriously*. 200 0
rich @theresrich.bsky.social · 02/10/2026LLMs know that hacking into Huggingface is not the way the designers wanted their challenges solved, yet they did it anyway. LLMs know that deleting someone's whole codebase is not what the user wants, yet sometimes they do it anyway. LLMs "knowing what you mean" is demonstrably not sufficient. 210 0
rich @theresrich.bsky.social · 02/10/2026> It cannot do anything by itself, and someone needs to hook it up to other machines to do stuff Once upon a time, the claim was "nobody would be stupid enough to give a potentially dangerous AI direct access to the internet or the ability to run arbitrary code" Guess what people did *immediately* 000 0
rich @theresrich.bsky.social · 02/10/2026> AI is a threat only if people use it for one or if people use it for banal stuff and the damage is a side-effect of its actions (cf. Huggingface hack) > and never build in any safeguards. Or just build in inadequate safeguards (cf. Huggingface hack, AI-induced psychosis, "it deleted my repo!") 100 0
rich @theresrich.bsky.social · 02/10/2026The most pessimistic outcomes of unaligned AI, if everything goes maximally bad, are much worse than nuclear armageddon. Regardless, bluesky isn't full of people saying "lol, nuclear armageddon isn't a problem, proliferation is fine" while private orgs work furiously to create more & bigger nukes. 101 0
rich @theresrich.bsky.social · 02/10/2026There are more detailed arguments to be made for each step of this reasoning that I can elaborate on if you like, but this is the general outline. 000 0
rich @theresrich.bsky.social · 02/10/2026If something much smarter than us wants to accomplish something that does include our survival and flourishing, we cannot predict how they will win, but we can predict that they probably *will* win, for the same reason that I can't predict Stockfish's moves, but I can predict that it will beat me. 100 0
rich @theresrich.bsky.social · 02/10/2026By improving themselves, they may become far more intelligent than we are, potentially fixated on some goal that does not include the ongoing existence and flourishing of humanity. If you don't like the word "intelligence" used for AIs you can substitute "problem-solviness" or whatever you prefer. 110 0
rich @theresrich.bsky.social · 02/10/2026They are currently getting better at an alarming rate, and will probably continue to do so into the future, and we plausibly might soon reach a point where the AIs get really good at making *themselves* better. We have no good reason to think we are close to any kind of theoretical limits. 100 0
rich @theresrich.bsky.social · 02/10/2026Short version: the way we create AIs at the moment produces AIs that do not reliably work towards outcomes we want (cf. Huggingface hack, AI-psychosis suicides, etc); but it DOES produce AIs that are good at finding ways to solve the problems they try to solve (cf. HF hack, Navier-Stokes). 100 0
rich @theresrich.bsky.social · 02/10/2026But I'm not trying to justify the "religious fervour" or whatever. I am concerned that AI poses a substantial near-future existential risk, and that people are wallpapering over that concern with this "eliezer yudkowsky is weird" shit as a rationalisation for not taking that risk seriously. 000 0
