🏅 Top traders

@Bayesian I think this market should be unresolved

It’s clear the point of ARC AGI is to measure the underlying model since they already had harnesses that achieved 100% back when Opus 4.8 was scoring ~1%

It simply doesn’t make sense to think

ARC AGI 1= 98.6%

ARC AGI 2 = 94.6%

ARC AGI 3 = 99.9%

62.7% makes way more sense and is an apples to apples comparison

re: “i'll count it bc it shows up on the official leaderboard“

It’s only on there with a caveat due to pressure from OpenAI

GPT-6 Astra isn’t considered “AGI” as a 90-100% score implied at market creation (spirit of the market).

@ChinmayTheMathGuy you’re mixing up the public and semi private versions of the benchmark. Gpt6 got 99.9% on the semi private, harder version of arc agi 3, and opus 4.8 got 100% on the public, easier version of arc agi 3

The harness creator doesn’t think 90-100% means reachijg agi. I don’t get why you are bullshitting me

ARC AGI 1= 98.6%

ARC AGI 2 = 94.6%

ARC AGI 3 = 99.9%

This has more to do with impossible problem set than anything, and makes sense

obv

ARC Prize (@arcprize) on X

ARC Prize (@arcprize) on X

GPT-6 Astra by @OpenAI achieves SOTA on ARC-AGI: - Astra scores 63% on ARC-AGI-3, 99% via a new provider adapter harness - It surpasses human performance on 96% of ARC-AGI-3 levels - It builds the most precise symbolic model of novel environments we've seen Our analysis:

@Bayesian the leaderboard table puts GPT-6 Astra at only 67%, because they used a custom harness for the 99.9% headline result.

Does that count for this market?

@Sketchy i'll count it bc it shows up on the official leaderboard

@Bayesian lol it wouldn't be ARC-AGI-3 without one last(?) rule change. A fitting send-off to this benchmark

What will Spud score?

expected this market to exist, could not find it, pls @ me if it does