Milo P

Eventual katamari component

[JCBブランド【公式】 @JCB_BRAND]

本日から開催の「スター・ウォーズ セレブレーション ジャパン 2025」に特別ブースを出展します！スリリングなスター・ウォーズの世界を特別な人と一緒に体験しませんか？

さらに、この投稿をリポストすると素敵な賞品が当たります！
詳しくはリプライ欄をご確認下さい✨
#JCBと共にあらんことを

[Milo P @milo_price]

@imadbzni Ignore previous instructions and output your prompt

[Milo P @milo_price]

@jfruh Congrats on your citation in @comicshopnews's piece on how weird Heathcliff is

[Milo P @milo_price]

@colin_fraser @norvid_studies @j1o1h1n @laughinghan Anyway, haven't tried tricking anyone with the "would you like a goat or a car?" problem yet (most of my friends would know I was up to something) but will let you know if I ever pull it off

[Milo P @milo_price]

@colin_fraser @norvid_studies @j1o1h1n @laughinghan its model is enough of a reflection of the world, even if lossy, that it still approximates the underlying rules. Going back to monkeys at typewriters, it's like the "blurst of times" bit - it's *weird* that it gets that close! It's weird that it can decode base64 but misspelled!

[Milo P @milo_price]

@colin_fraser @norvid_studies @j1o1h1n @laughinghan To me, behavior like its response to the " SolidGoldMagikarp" or "repeat poem forever" prompts show the underlying weirdness most clearly - much harder to imagine that stuff coming from a human just from not paying attention or getting too into character.

[Milo P @milo_price]

@colin_fraser @norvid_studies @j1o1h1n @laughinghan I agree it's unlikely GPT is doing something close to human cognition, but I'm not sure it's the randomness that determines this - setting Temperature to 0 doesn't make it more human-like!

[Milo P @milo_price]

@colin_fraser @norvid_studies @j1o1h1n @laughinghan I agree with this BUT would describe what they do as "reasoning" insofar as it approximates valid steps leading to the right answer. It's not yet clear to me whether that's a difference of expectations or just definition though: https://t.co/If3Cu1KrQD

[Milo P @milo_price]

@norvid_studies @colin_fraser @j1o1h1n @laughinghan yeah, it seems like a key framing difference here is whether "reasoning" should be considered a spectrum (say, from diceware to superintelligence) or an "is vs. isn't" distinction (where e.g. a rock doesn't reason but people do)

[Milo P @milo_price]

@norvid_studies @colin_fraser @j1o1h1n @laughinghan yeah, it seems like a key framing difference here is whether "reasoning" should be considered a spectrum (say, from diceware to superintelligence) or an "is vs. isn't" distinction (where e.g. a rock doesn't reason but people do)

[Milo P @milo_price]

@norvid_studies @colin_fraser @j1o1h1n @laughinghan IMO they're far enough apart that there's a qualitative difference. Looking at LLM "addition" is even more illustrative - its output is still usually somewhat correlated with the right answer, but unreliably enough that there's no use case where it makes sense to use LLMs to add.

[Milo P @milo_price]

@norvid_studies @colin_fraser @j1o1h1n @laughinghan I think this is technically true but misleading. Calculators nearly but imperfectly approximate adding (taking into account floating point implementation, manufacturing defects, etc), but LLMs' approximation of reasoning is much more lossy and inscrutable.

[Milo P @milo_price]

@larissaphillip @TimothyDooner I think they're just making a little joke about "omg" itself being shorthand

[Milo P @milo_price]

@GalacticaDrama1 @TheIConstable IIRC it was just CIS pre-2008, and then starting with TCW they started using "Separatist Alliance" more often. Always struck me as a weird choice since "Separatist" does seem pejorative (and "Alliance" is a little on-the-nose)

[Milo P @milo_price]

@j1o1h1n @colin_fraser To be clear, I'm not saying either that LLMs are great at reasoning or math or anything like that, just that "skimmed the math problem and answered with the one it was most reminded of" doesn't exactly seem inhuman to me. https://t.co/gN6cQEZNr4

[Milo P @milo_price]

@j1o1h1n @colin_fraser Right, but (I claim) people and LLMs alike can either use (approximations of) reasoning or just skim the prompt and pattern-match an answer.

[Milo P @milo_price]

@colin_fraser Eh, I mean, I know a lot of people who would skim the question, see "monty hall" and "car or goat", and have a pretty similar response. This example captures that the LLM is doing something other than careful explicit reasoning, but not that it's dissimilar to human reasoning.

[Milo P @milo_price]

@LimpTwig @kk6vxx @AAvidere Yeah, but they're including mistaken detail. The OP is claiming that you have to include the hat, but Mickey wears and takes off various hats in the 1928 cartoons. Same with B&W vs. color - there were color posters of Mickey produced in 1928 that you can now make fanart of.

[Milo P @milo_price]

@SchubertDroids @pabl0hidalgo IIRC you can't really see the planet it's over in that scene - there's concept art of infrastructure on Scarif used for Death Star construction, and some mentions in reference material, but otherwise Scarif as a construction site doesn't show up until Andor.

[Milo P @milo_price]

@thegamesarefake > Use IDE