New "50%" ARC result and current winners interviewed

14 snips

Jun 18, 2024

Guest

In this engaging discussion, Jack Cole, a clinical psychologist and ARC Challenge winner, along with AI researcher Mohammed Osman and expert Michael Hodel, delve into the nuances of the ARC Challenge, which assesses AI reasoning. They present their winning approach of fine-tuning language models, emphasizing active inference and innovative data representation. The trio debates the philosophical implications of their methods on intelligence measurement while highlighting the addictive nature of ARC tasks and raising questions about the future of AI and generalization.

Ask episode

AI Snips

Chapters

Books

Transcript

Episode notes

INSIGHT

ARC Challenge and Intelligence

François Chollet's ARC Challenge tests AI generalization abilities.
It aims to be minimally interpolative, focusing on reasoning and knowledge acquisition.

INSIGHT

LLMs and Concepts

LLMs might learn concepts and guide program combinations based on them.
This suggests they could generalize beyond database lookup, challenging the current understanding.

INSIGHT

Generalization and System 2

Models generalize within their learned distribution, performing additional processing on novel examples.
This inner processing, involving search and disambiguation, resembles System 2 thinking at a smaller scale.

Get the Snipd Podcast app to discover more snips from this episode

Get the app

The ARC Challenge, created by Francois Chollet, tests how well AI systems can generalize from a few examples in a grid-based intelligence test. We interview the current winners of the ARC Challenge—Jack Cole, Mohammed Osman and their collaborator Michael Hodel. They discuss how they tackled ARC (Abstraction and Reasoning Corpus) using language models. We also discuss the new "50%" public set approach announced today from Redwood Research (Ryan Greenblatt). Jack and Mohammed explain their winning approach, which involves fine-tuning a language model on a large, specifically-generated dataset and then doing additional fine-tuning at test-time, a technique known in this context as "active inference". They use various strategies to represent the data for the language model and believe that with further improvements, the accuracy could reach above 50%. Michael talks about his work on generating new ARC-like tasks to help train the models. They also debate whether their methods stay true to the "spirit" of Chollet's measure of intelligence. Despite some concerns, they agree that their solutions are promising and adaptable for other similar problems. Note: Jack's team is still the current official winner at 33% on the private set. Ryan's entry is not on the private leaderboard or eligible. Chollet invented ARC in 2019 (not 2017 as stated) "Ryan's entry is not a new state of the art. We don't know exactly how well it does since it was only evaluated on 100 tasks from the evaluation set and does 50% on those, reportedly. Meanwhile Jacks team i.e. MindsAI's solution does 54% on the entire eval set and it is seemingly possible to do 60-70% with an ensemble" Jack Cole: https://x.com/Jcole75Cole https://lab42.global/community-interview-jack-cole/ Mohamed Osman: Mohamed is looking to do a PhD in AI/ML, can you help him? Email: mothman198@outlook.com https://www.linkedin.com/in/mohamedosman1905/ Michael Hodel: https://arxiv.org/pdf/2404.07353v1 https://www.linkedin.com/in/michael-hodel/ https://x.com/bayesilicon https://github.com/michaelhodel Getting 50% (SoTA) on ARC-AGI with GPT-4o - Ryan Greenblatt https://redwoodresearch.substack.com/p/getting-50-sota-on-arc-agi-with-gpt Neural networks for abstraction and reasoning: Towards broad generalization in machines [Mikel Bober-Irizar, Soumya Banerjee] https://arxiv.org/pdf/2402.03507 Measure of intelligence: https://arxiv.org/abs/1911.01547 YT version: https://youtu.be/jSAT_RuJ_Cg