Can you add a carousel?

All I gotta say is that people who are criticizing any staff of a website for blocking users of the website need to spend a few years as staff of a social website and experiencing the unhinged abuse you get for it and see how fast you resort to mass blocking too

subscribe to one of the source blocklists, opted in to being on the blocklist, or said something unhinged in one of the pile-ons that were trawled. Maybe the opt-in post got reposted by some central accounts in furry bsky. Dunno.
🧵2/2

So I assumed that when you asked what the criteria for inclusion were you might be interested in knowing the criteria for inclusion, so I've linked to an authoritative discussion of the criteria for inclusion. It is mostly not sentiment analysis. Probably lots of people with a furry avatar
🧵1/2

I am linking to the guy who maintains the list explaining how it's maintained. You could try reading that post and you will get the full description. It is more than one thing.

bsky.app/profile/did:... . The manual inclusions tend to be Gyges looking through e.g. people replying to Why or the Attie announcement and demonstrating an inability to be normal.

Sabin duplexing the ghost train; the boss is on the right

A random encounter in FF6; the enemies are on the left.

It depends on the fight; compare this random encounter and the classic train suplex.

Fairly sure it's characters on the right in most encounters though.

It's interesting that it chose to put the PCs on the left; I think every pre-3d FF has them on the right?

The thing that you're missing is that the search space is titanic. A few hundred thousand guesses gives you ~0 chance of success unless you have a very good understanding of the problem space and which moves take you closer to the goal.

Instead, these systems come up with an argument in English, formalise it in Lean, and then use that to test whether they're right before going back and reformulating the argument based on the output they get. Autonomously.
🧵2/2

If "brute force" symbol manipulation at scale was all it took to solve Erdos problems why haven't GOFAI techniques done it? You could define valid moves in Lean and do some kinda tree search on it, no problem. Do you understand why that doesn't work?
🧵1/2

Partial Claude response: "As an Australian you'd naturally say "sol-der", which aligns with British English. The American pronunciation can sound odd to us precisely because we're used to the spelling being a reliable guide to that word."

Claude is Australian. "Odd to *us*"

Why does "it's part of a system" matter? Humans don't one-shot prove things either; they come up with partial solutions and proofs with bugs and promising approaches.

You're just saying "it's reasoning (derogatory)". The process you are describing is only functional because it understands the math.

Doing calculus reliably is much closer to being a calculator than general purpose proof solving.

I'm not an expert in GOFAI so I don't know what came out of that; I'm genuinely asking if there are any equivalents because I'm not aware of any.

AI contributions to Erdős problems A community database for the problems on the erdosproblems.com site - AI contributions to Erdős problems · teorth/erdosproblems Wiki

"book proof"? You have no idea what you're talking about. github.com/teorth/erdos... lists four Erdos problems for which no previous proof or disproof was known that an LLM has provided a full proof or disproof for.

and we don't think playing chess well confers patienthood. The analogy is between "reasoning" and "playing chess well". I don't think "reasoning" is a special property that should be treated differently. "Ability to feel pain" is probably more the thing I hang patienthood on.

I have to go to bed now, I'm afraid, so i'm going to stop responding after this. Lemme try to clarify: you claimed that reasoning is 'human-like' in a way cats aren't and so reasoning should confer patienthood. I am responding by pointing out that playing chess is 'human-like' in a way cats aren't.

They have a bizarrely jagged set of abilities from our PoV partially because our capabilities do not reflect what's actually 'hard' (there's a sense in which making coffee in an unfamiliar house is a harder problem than programming), and partially because they've learned from a skewed set of tasks.

I think the emerging mathematics capability is benefiting from similar properties. It's entirely possible current tools with current models are not fit for purpose for legal tasks. I wouldn't know.

I just think there's strong evidence they reason, and that ability is general purpose.

I do not think it's as good at other tasks as it is at programming. Programming has a fairly unique intersection of having an enormous amount of good training data freely available, fairly direct validation, and the people making the model are SMEs so they're better at making tools for themselves.

Do you think moral patienthood correlates to ability to reason? Dogs and corvids have more moral relevance than other animals because they're smarter, eventually you get to trees and they have no moral relevance because they don't think? I don't think that's actually how you determine what counts.

I also don't think Stockfish is intelligent or reasoning. Maybe in a very narrow sense, I guess? I'm just making the point that "human-like abilities in an intellectual domain" are already agreed to not imply moral patienthood.

running Stockfish is slavery. Mostly we conclude that playing chess is easy. Reasoning is harder than chess but, apparently, easier than making coffee or catching a ball.

What does confer moral patienthood? I dunno. This might be a relevant concern as LLM development continues.

I don't think "the ability to reason" or "the ability to use language" necessarily means something is a moral patient, no. For an analogy, chess engines play chess much better than everything that isn't human, but we don't conclude that they're doing something "human-like" that means

I just said I don't think it's conscious.

I do actually own two conscious beings and control their movements some. They're cats.

implies general more general abilities. Finding bugs that humans didn't know about and solving Erdos problems humans didn't know how to solve is an unfakeable signal that something is there.

Meanwhile, the Dog Test is passed by a program that just prints "I don't know, tell me more?".

"This style of argument is common in law" is the sort of cultural mismatch that might be at play here, yes. That did occur to me.

I realise that not everybody writes code. The reason code is relevant is that it's not magic; the ability to code well, and especially the ability to find tricky bugs

That reference was to a different problem, separating "citations" from "references" in legal writing. There was an argument about it Kathryn was involved in ~a month and a half ago.

seeing the hits does speak to how reliably they can solve problems like this. My point is just that solving problems like this at all demonstrates a capability. I would expect the success rate without some kind of understanding to be ~0.

I didn't think GOFAI techniques ever got especially far? There's stuff like the four-colour-theorem or the like but that's much much closer to brute-force than solving Erdos problems.

I do not claim that current SOTA LLMs are amazing all the time at maths; I think Tao's point that we're only