Pete the Pirate

[Pete the Pirate @fairjac]

@Hitchslap1 The average person is emotional about both. I certainly don't wanna feel like I'm on the low end of some scale, and that's just how it is

... but athleticism is undeniable, whereas you can kinda obfuscate around IQ - "maybe I'm not book smart, but I'm xyz"

[Pete the Pirate @fairjac]

@ShaunUnsworth @JacobHorsfall_ right... so your problem isn't with xG then

your problem is that these tables tell you something about how clinical a team's finishing is (xG vs. reality)

but you expect them to tell you where a team will finish (xG "predicts" reality)

aka skill issue

[Pete the Pirate @fairjac]

@ShaunUnsworth @JacobHorsfall_ I could make a table based on possession%, assuming the team with more of the ball wins the game.

Would you call possession "shite" in that case?

[Pete the Pirate @fairjac]

@feito_chan @expatanon Yep. These are all an entirely different thing.

[Pete the Pirate @fairjac]

@Nymdok @Alphiloscorp Pure maths undergrad at a good university. Studied Galois theory & algebraic topology, then taught high school maths.

Light-years away from the expertise that verified this proof! But educated enough to know what these guys mean by "elegant" or "ingenious".

[Pete the Pirate @fairjac]

@feito_chan @expatanon Entirely different thing.

[Pete the Pirate @fairjac]

@Nymdok @Alphiloscorp (2/2)

"(the proof) applies... tools from algebraic number theory in an elegant and clever way."

"the model has... good intuition... they are capable of having original ingenious ideas"

"It is... an intimidating construction to see through even if you know what is going on"

[Pete the Pirate @fairjac]

@Nymdok @Alphiloscorp (1/2) If you're curious, the mathematicians' comments at the bottom of this page are good enough:
https://t.co/FNGqv5GAUw

... or open the companion remarks, for more depth.

Some choice quotes:

[Pete the Pirate @fairjac]

@Nymdok @Alphiloscorp Few would disagree with you, I think.

What's significant about the recent Erdos proof is: GPT made *exactly* that sort of inductive leap.

[Pete the Pirate @fairjac]

@AlexanderKalian Consider:

a. The conjecture stood for decades
b. The vast majority believed it to be true
c. The approach taken has been described as novel by dozens of mathematicians

Is this not evidence for 2?

Put differently: what sort of scaffolding is even conceivable, given a-c?

[Pete the Pirate @fairjac]

@jdegoes So tl;dr: AGI requires a high success rate, x%, on attempting to solve problems.

What is x exactly?

How many incorrect proofs can a human mathematician publish, before you declare them non-generally-intelligent?

[Pete the Pirate @fairjac]

@GaryMarcus 1. The (obvious) alternative is that a step change occurred earlier than 5.5Pro.

2. "Maybe... easier than some" is in poor faith. You've read the blog post; you know who those experts are, and that they have a great deal more expertise than yourself. It's not low-hanging fruit.

[Pete the Pirate @fairjac]

@GaryMarcus (3/3)
... and your point on "formally verifiable domains" feels a bit lose-lose. If you view maths as inherently "computery", then ok... but what would success look like in a non-formally verifiable domain? How would one identify a step change in, say, AI-generated poetry or art?

[Pete the Pirate @fairjac]

@GaryMarcus (2/3)

The mathematicians' view is clear: the search space is so vast, that no version of monkeys-with-typewriters produces this result. So imo, any success rate in the range 0.000001% - 100% is a significant milestone.

[Pete the Pirate @fairjac]

@GaryMarcus (2/2)

I think this is playing poorly, and underlies a lot of the "goalpost-shifting" accusations.

Is there really nothing to celebrate?

Can we not kick the can down the road on costs/success rate, and simply recognise that the Erdos result is a step change in capability?

[Pete the Pirate @fairjac]

@GaryMarcus (1/2) Maybe I'm reading between the lines too much...

But what I'm getting at is: you're a well-known skeptic, and your focus seems to have shifted recently from "AI can't do XYZ" to "AI can't do XYZ cheaply/reliably".

[Pete the Pirate @fairjac]

@KenTFoss @wool1111 (2/2) ... an entirely unrelated field of mathematics. More to the point, it identified the link - something no human has done in 80yrs of effort.

It's worth you reading at least the commentary of professional mathematicians here, before "assuming fraud because AI".

[Pete the Pirate @fairjac]

@KenTFoss @wool1111 (1/2) This absolutely does not describe the Erdos result.

The search space for counterexamples is unimaginably vast; "brute force + time" would never work in the lifetime of the Universe, and is not what happened here.

GPT found a beautiful and inspired application of...