The popular online discourse on Large Language Models’ (LLMs’) capabilities is often polarized in a way I find annoying and tiresome. On one end of the spectrum, there is nearly complete dismissal...
The popular online discourse on Large Language Models’ (LLMs’) capabilities is often polarized in a way I find annoying and tiresome.
On one end of the spectrum, there is nearly complete dismissal of LLMs: an LLM is just a slightly fancier version of the autocomplete on your phone’s keyboard, there’s nothing to see here, move on (dot org).
This dismissive perspective overlooks some genuinely interesting novel capabilities of LLMs. For example, I can come up with a new joke and ask ChatGPT to explain why it’s funny or come up with a new reasoning problem and ask ChatGPT to solve it. My phone’s keyboard can’t do that.
On the other end of the spectrum, there are eschatological predictions: human-level or superhuman artificial general intelligence (AGI) will likely be developed within 10 years or even within 5 years, and skepticism toward such predictions is “AI denialism”, analogous to climate change denial. Just listen to the experts!
There are inconvenient facts for this narrative, such as that the majority of AI experts give much more conservative timelines for AGI when asked in surveys and disagree with the idea that scaling up LLMs could lead to AGI.
The ARC Prize is an attempt by prominent AI researcher François Chollet (with help from Mike Knoop, who apparently does AI stuff at Zapier) to introduce some scientific rigour into the conversation. There is a monetary prize for open source AI systems that can perform well on a benchmark called ARC-AGI-2, which recently superseded the ARC-AGI benchmark. (“ARC” stands for “Abstract and Reasoning Corpus”.)
ARC-AGI-2 is not a test of whether an AI is an AGI or not. It’s intended to test whether AI systems are making incremental progress toward AGI. The tasks the AI is asked to complete are colour-coded visual puzzles like you might find in a tricky puzzle game. (Example.) The intention is to design tasks that are easy for humans to solve and hard for AI to solve.
The current frontier AI models score less than 5% on ARC-AGI-2. Humans score 60% on average and 100% of tasks have been solved by at least two humans in two attempts or less.
For me, this helps the conversation about AI capabilities because it gives a rigorous test and quantitative measure to my casual, subjective observations that LLMs routinely fail at tasks that are easy for humans.
François Chollet was impressed when OpenAI’s o3 model scored 75.7% on ARC-AGI (the older version of the benchmark). He emphasizes the concept of “fluid intelligence”, which he seems to define as the ability to adapt to new situations and solve novel problems. Chollet thinks that o3 is the first AI system to demonstrate fluid intelligence, although it’s still a low level of fluid intelligence. (o3 also required thousands of dollars’ worth of computation to achieve this result.)
This is the sort of distinction that can’t be teased out by the polarized popular discourse. It’s the sort of nuanced analysis I’ve been seeking out, but which has been drowned out by extreme positions on LLMs that ignore inconvenient facts.
I would like to see more benchmarks that try to do what AGI-AGI-2 does: find problems that humans can easily solve and frontier AI models can’t solve. These sort of benchmarks can help us measure AGI progress much more usefully than the typical benchmarks, which play to LLMs’ strengths (e.g. massive-scale memorization) and don’t challenge them on their weaknesses (e.g. reasoning).
I long to see AGI within my lifetime. But the super short timeframes given by some people in the AI industry feel to me like they border on mania or psychosis. The discussion is unrigorous, with people pulling numbers out of thin air based on gut feeling.
It’s clear that there are many things humans are good at doing that AI can’t do at all (where the humans vs. AI success rate is ~100% vs. ~0%). It serves no constructive purpose to ignore this truth and it may serve AI research to develop rigorous benchmarks around it.
Such benchmarks will at least improve the quality of discussion around AI capabilities, insofar as people pay attention to them.
Update (2024-04-11 at 19:16 UTC): François Chollet has a new 20-minute talk on YouTube that I recommend. I've watched a few videos of Chollet talking about ARC-AGI or ARC-AGI-2, and this one is beautifully succinct: https://www.youtube.com/watch?v=TWHezX43I-4
