The first time I publicly announced the v2 was actually… in May 2022 in Davos. Back then it was a joint effort with Lab42, now it's handled by the ARC Prize foundation. We certainly took a while to ship, but the results are worth the extra effort.
@fchollet
-
ARC Prize 2024 Technical Report Released
By
–
Link: https://
arcprize.org/media/arc-priz
e-2024-technical-report.pdf
… -
ARC Prize 2024 report and upcoming v2 benchmark for 2025
By
–
If you haven't read the ARC Prize 2024 technical report, check it out (link in next tweet). One important bit: we'll be releasing a v2 of the benchmark early next year (human testing is currently being finalized). Why? Because AGI progress in 2025 is going to need a better
-
Search and Deep Learning: The Future of AI Systems
By
–
Search is a fully general technique. But program search today is at the level of sophistication that deep learning was at in 2011. Hence why deep learning appears more general. Soon, you will see combined search + deep learning systems vastly outperform anything you can do with
-
ARC-AGI: Non-Deep Learning Achieves 40% on Private Test Set
By
–
On what, ARC-AGI? 40% with no deep learning at all right now (private test set). Best DL based solution is at 55%.
-
Comparing iterative reasoning systems versus discrete program search efficiency
By
–
Re: "optimal way to do reasoning". The idea here is that if you have an "iterated system 1" type reasoning system that takes ~$1k and O(hour) in compute to find a correct program to solve a simple task, vs. a discrete program search system that does the same in 0.1s on your
-
System 2 as Iterated System 1 With Self-Consistency Guarantee
By
–
"system 2 as iterated system 1 with a self-consistency guarantee" is something I considered a few years ago as one of two hypotheses, and it led me to an interesting potential interpretation of consciousness — consciousness is the consistency guarantee. It explains why all
-
Fuzzy Pattern Matching and Reasoning: Plausibility vs Optimality
By
–
It's highly plausible that fuzzy pattern matching, when iterated sufficiently many times, can asymptotically turn into reasoning (it's even possible that humans do it basically in this way). But that doesn't mean it's the optimal way to do reasoning.
-
Challenge as feedback signal for AI progress, not AGI measurement
By
–
It's not primarily meant as a way to measure how far/close we are to AGI, it's meant as a challenge and feedback signal that helps you get closer by solving it
-
ARC as a Benchmark for True AGI: Solving Human-Level Generality
By
–
Not exactly — it's the other way around: if you can't solve ARC (which humans can solve), you don't have AGI (human-level generality) More broadly, the purpose of ARC is to be a useful tool that directs your focus towards the most productive unsolved bottlenecks to AGI