Francois Chollet - LLMs won’t lead to AGI - $1,000,000 Prize to find true solution

Here is my conversation with Francois Chollet and Mike Knoop on the $1 million ARC-AGI Prize they're launching today.

I did a bunch of socratic grilling throughout, but Francois’s arguments about why LLMs won’t lead to AGI are very interesting and worth thinking through.

It was really fun discussing/debating the cruxes. Enjoy!

Links:

Timestamps:
00:00:00 – The ARC benchmark
00:11:53 – Why LLMs struggle with ARC
00:19:43 – Skill vs intelligence
00:28:38 – Do we need “AGI” to automate most jobs?
00:49:11 – Future of AI progress: deep learning + program synthesis
01:01:23 – How Mike Knoop got nerd-sniped by ARC
01:09:20 – Million $ ARC Prize
01:11:16 – Resisting benchmark saturation
01:18:51 – ARC scores on frontier vs open source models
01:27:02 – Possible solutions to ARC Prize

[cmfrtblynmb]

One advice to host: You need to give your guest space. You are not a salesman. Or a missionary. Challenging them does not mean repeating the same argument over and over again. It was suffocating to listen to your challenges. If it was not for the call and patient demeanor of Chollet, it would be impossible to watch. We were not able to listen to Chollet expanding upon his ideas because host just reversed the clock to zero by repeating the same "but memorization is intelligence" argument. It should be about your host, not showing the supremacy of your ideology or beliefs. If your host is wrong, you can prove them wrong by showing their arguments and ask questions as they expand upon them and then show if they are inconsistent. Not repeating the same thing over and over and over again

[MM-uvvb]

Need more interviews with legit AI/AGI skeptics to balance out the channel

[TheMrPopper]

Dwarkesh dismissed a few completely valid answers to try and steer the answer in his preconceived idea of LLMs, I didn’t like that, dude is smart, let him finish and actually take onboard his answer before asking another question

[therobotocracy]

What I learned from this: Be the guy who speaks slower under pressure, not the guy who talks faster!

[SiphoNgwenya]

Thank goodness for Francois' infinite patience

[driziiD]

Francois doesn't even sound like a skeptic, just an informed educator

[perer]

Dawarkech does not appear to have the ability to adapt on the fly in this interview! 😂

[BrianPeiris]

Really glad to see people like Chollet are willing to say the emperor has no clothes. I'd recommend Subbarao Kambhampati's talks as well. He goes into some theories about _why_ people are being fooled into thinking that LLMs can reason.

[ayandas]

Francois Chollet is an amazing guy. The best thing is he, like all the LLM guys, also wants to work toward AGI! He just doesn't think the current LLM paradigm will get us there. I'm really excited to see where this goes because he's challenging the current option space in exactly the right way

[snarkyboojum]

So good to see someone let some air out of the LLM bubble. Dwarkesh might be a little challenged by this, but it’s great to get out of the echo chamber regularly.

[josteveadekanbi]

Train an LLM to solve math but don't include anything related to calculus in the training data.
Then ask the LLM to solve a calculus problem and you'll see it fail.
That's essentially what Francois Chollet was saying.
Isaac Newton was able to introduce calculus based on his FOUNDATION OF MATH (Memory) and actual INTELLIGENCE (Ability to adapt to change)

[GabrielMatusevich]

There was 20mins of:
"Can LLMs replace programmers?"
"No"
"But can they?
"No"
"But can they?
"No"
"But can they?
"No"
"But can they?
"No"
"But can they?
"No"
"But can they?

XD ... it simply becomes clear that LLMs can't replace programmers when you start using them everyday on your programmer job and realize how Bad they perform when you start to do just slightly complex logic

[aidanshaw]

This is by far the best interview until now. We need to hear the skeptics too, not only the super optimists. I really like the french guy.

[jonatani]

literally tortures him with "okay but isn't this AGI why is that not AGI?" questions and after not having any positive feedback asks "okay but let's suppose you lost your job to AGI"

[rossmccannell]

"But if we gave LLMs as many Adderalls as I popped before this interview, would then get AGI?"

"Ok, that may work."

"That was a trick question. I snorted the Adderall."

[diracspinors]

It is pretty glaring how different Francois's interview is from Illya's. Maybe part of it is the result of Dwarkesh polish as in interviewer where for Illya it was a series of questions Illya minimally answered and here Francois openly expanded upon the problem. But, also from the start the two seemed different. Where Francois maintains all the features of researcher, who values an open exchange of ideas, Illya values secrecy and the advantages of being a first mover. I definitely enjoyed this interview much more.

[jsn]

I think Francois really got his point across in the end there. I was leaning somewhat to the scaling hypothisis side, he made me question that more. In any case you have to give him credit him for actually coming up with interesting stuff to support his arguments, unlike many other critics.

[telotawa]

1 million for this is as ridiculous as 1 million for P vs NP, it's a multi trillion dollar problem, it's like offering $1 for someone to find your lost supercar or something

[sjkba]

I love how Francois explains his answers with great patience and a subtle smile in his eyes. What a dude.

[studywithmaike]

I must say I don’t really like the way he keeps interrupting François during the interview