François Chollet: Measures of Intelligence | Lex Fridman Podcast  

I really enjoyed this conversation with François. Here's the outline: 0:00 - Introduction 5:04 - Early influence 6:23 - Language 12:50 - Thinking with mind maps 23:42 - Definition of intelligence 42:24 - GPT-3 53:07 - Semantic web 57:22 - Autonomous driving 1:09:30 - Tests of intelligence 1:13:59 - Tests of human intelligence 1:27:18 - IQ tests 1:35:59 - ARC Challenge 1:59:11 - Generalization 2:09:50 - Turing Test 2:20:44 - Hutter prize 2:27:44 - Meaning of life

We need Joscha Bach to come back ASAP!

Listening Francois gives so much insights on everything. He is just simply amazing

“Everything we do creates ripples into the future and that’s the path to immortality”

“I’m not interested in psychometrics as a way to characterize 1 individual person. I think psychometrics is most useful as a statistical tool. It’s most useful at scale.” This man is a measure of intelligence himself.

Hey, Lex, I love how you dig up the guests that we didn't know we needed to hear and provide high stimulating intellectual discussion, keep it up

Lex, Thank you for providing us layman with this extremely beautiful and captivating podcast. We all appreciate the time and effort you put into this project of yours. Life would truly be quite different without you!

When Francois talks, he gives a new insight every time. What a genius!

“Intelligence is how efficiently you learn new things” “Measure of intelligence is ability to change” Being able to learn comes from ‘understanding structured operation logic” model between events. And being able to change comes from the ability to use causal inference on a perceived event against the model.

I really like the way this man talks for some reason, like he has a fairly thick accent yet is more intelligible than a lot of native English speakers.

I love the way these two people think completely differently. I've had conversations about the way I think compared to my significant other, and its surprising to see how both similar and different ours ways of thinking can be!

'So you disagree with Naom Chomsky?!' '𝓨𝓮𝓼'

Fransios is such a guy that warms your heart with knowledge!

I listened to this after consuming thc. I had several mental breakthroughs by applying the albeit hard for my intelligence level to decipher. Highly valuable information contained within. Lex you have an amazing mission to question the questions, That has helped sculpt my mind into one that is slightly more efficient than it was. More efficient at detecting bias. I still have many biases. But I am getting better at detecting them. Thankyou youtube for providing such valuable content. The world is both better and worse for the torrents of information literally at our fingertips. From one random guy on the internet. Choose the content you consume wisely. Turn the play speed up incrementally over time as you adjust. This adds profound gains in efficiency of absorption. The faster play speed becomes a second language. You will become philosophically bi-lingual. Other people will likely not understand 2.0x play speed yet it sounds normal when you become fluent. 3.0x is only 50% faster than 2.0x. My average level of intelligence. Or maybe the frequency perceptiveness of my ears, will reach a performance ceiling before others with more capacity. Neurolink has the ability to increase the speed of uptake. In a realm where consciousness consists of stored information combined with interpretation and adaptation. Learning the dataset is just as important as improvisational ability. Any efficiency increase should produce measurable results in all round performance. /consciousness.exe

Legend says that when Francois applied to Google, the interviewers began to question the level of their own Googliness

What a great discussion! I loved the metaphor of having a subway map in your head as you write a paper. I do that all the time.

Love your work, mate. You're an amazing speaker and an even more amazing interviewer (or vice versa 😃)! Currently chugging through your extensive backlog of interesting content! Love it and keep up the good work!

This channel is a hidden gem on youtube. Carry on lex.

This is an amazing interview. One of the most rewarding ones I have ever heard. Fantastic!!

Thanks for these podcasts, lex. I've been looking for a podcast with things that I am interested in. Keep these things coming, I find a lot of value from it.

Thank you both lex and Francois. This was a great pod cast. Thank you for this genuine and honest podcast. I as a viewer appreciate the knowledge you two are able to explain in understanding ways. God bless u both for ur good knowledge on things

This was really beautiful. Especially the very end starting from the meaning of life. Hoping to be a patron soon. Best wishes

Loved the Piaget reference. One of the experiments described in Genetic Epistemology has had a profound impact on me.

This one got me really excited

This show breaks my brain everytime I watch it. Love it. Thx Lex and guests. Hella outlet.

Hell yes, I was waiting for exactly this episode ever since I read the paper.

François is such an amazing person, loved this!

I request a part 3. I’ve watched a number of these AI interviews, but in my opinion Francois has the deepest and clearest insights on this topic.

What a great interview, makes me realize how crucial the ability to simulate reality is to intelligence.

The best logical person i ever heard talking about intelligences and AGI.

Lex, thank you so much for yet another awesome podcast!

"Russian starts with a language and ends with a vodka" :) lol

Great guest and great interview! Thanks Lex.

Wonderful fridman, thank you.

Thank you so much for the insightful discussion!

Thanks a lot Lex

I really like the way this guy thinks. Glad to see him on again!

Would love to hear discussions on the differences in the mechanisms of emotional and classical intelligence.

Good interview (so far). Seems like the proper view on language.

Lex thank you everyday for inspire me

Masterclass looks good. I would also like to pre-order that Sci-Fi novel of yours Lex.

Another great guest and interview. Best response to the meaning of life question I've heard.

What the hell? 18k views, 140 comments? These are too low, people dafuq. This shit is the most interesting informative discussion ever Edit- ... discussion ever today, as every podcast is amazing

Have David Deutsch on your podcast!!

I liked the ad Lex made for Babble.

Very nice talk!

Wonderful and very subtle ideas about consciousness and learning. Thank you. Some historic non-quotes: Descartes on logic: “Remember to always put des cartes before des horse.” and “C’est bien joué Monsieur Des Cartes.” Accordingly to Bertrand Russell, Descartes always wore his sword at his side, even when he was in his apartment. Thank you. William L. Ramseyer

This guy.. is intelligent. I believe he has solved what intelligence really is, and that’s amazing. Also he has great insights about thé shortcomings of current deep learning and of neuralink

Language - the operating system of the mind. Incredibly profound.

The amount of no real words in this pod is out of the charts. Amazing talk as usual, thanks, I learned a lot.

I don't think more intelligence is necessarily a good thing. Rather what we need is more equality, kindness and self-knowledge

2:31:15 he is playing both sides do that no manner which one wins he comes on top, this man is a tactical genius as well

Really admire Francosi Chollet. Thank you for this amazing talk!

Who wants Geoff Hinton on the Podcast ?

Super helpful textbook 📚

my favorite twitter person is back!

Lex, as I observe the world and people around me, I also determined primary measure of intelligence is the ability to adapt to change. This is driven by deeper layers of character/mind that I have identified.

Без водки, не случился б не один прорыв! Спасибо тебе за лучшие подкасты, с удовольствием смотрю и узнаю для себя, что-то новое!

50:18 You can, in fact, train on 100x more data when you add in video and audio

Bidirectional search 14:45

This blew my mind

Not just new but also difficulty .. level of newness and complexity

Just bought his book; what a beautiful coincidence!

Hey Lex, have you heard about Zettelkasten? It's a method of organizing information which is somewhat similar to mind maps but more sophisticated. You might find it useful. Btw, thanks for you podcasts, they are amazing! ❤️

I like the way you reason lex

very interesting

Luv the poem...Pushkin request

Great conversation, general intelligence is a very interesting topic. The intelligent mind seems to me to be a confluence of culture, curiosity and the ability to address new environments and problems in a new and imaginative way, it’s a dynamic constantly evolving awareness of environment, threat level and incentives.

Great conversation. Thank you. A couple things: I wish you had talked about the ability to “know” which transcends our compressed past and our recognized ways of perceiving, by some considered legitimate and real ways of tapping into knowledge which is actually beyond of physical finiteness. Also, our interconnected ness to other life forms. Perhaps even non physical consciousness? How do we factor in ways of learning and perceiving info from these sources which can add to our compressed knowledge as well as, dare I say it, pave ways into new ways of knowing which transcend linear time and space itself? Taking into account recent revelations in the fields of quantum physics and the emerging areas studying consciousness itself, it seems like these considerations are important not only in the development of AI, but also when considering the meaning of life itself.

I really like the way this guy thinks, Francois pushes the boundaries of how to measure intelligence. I like the idea of not using the Turing test when people cheat and people will always cheat so it is best to utilize the scientific method. Like what Kate Darling was referring to, if we tell the judges they are talking to a child from Namibia then it is much easier to get them to agree that the computer is human because we have shifted the judges away from their base of reference. We are tricking them.

We need an update!

The 'child mind' is, at the beginning, pure observation. Language builds out of the interactions between other people. Thinking, pre language, is the child mind. In my opinion, all thinking continues to happen in this mind, and you have to translate on the fly. You get really good at it. Language only has meaning in the interactions between people. But the child mind still has to have a language to be translated. Other than pure observation, what comes before that?

I agree with Chollet that language is a layer on a cognitive base. Interestingly, the early universe, according to the standard model, was more accurately described by our language than the reality that we perceive. It wasn't until those few massless particles grew in number and created the force fields that gave the particles mass, that being as we know it came about...which is a combination of being and moving, not separated as in our language of nouns and verbs.

Do a video on vodka cocktails..

Hehe, I read his book: "Deep Learning with python" --> It's by far the best book I've read about this topic. In my opinion, Human Intelligence often uses heuristics. Computational power can extend these heuristics by the huge amount of operations. So, the proof of "Four-Color-Theorem" was just possible by the combination of human intelligence and computers.

I've always been curious, what is Francois' academic background? He's obviously a great engineer -- keras is a beautiful API and a huge success. Somehow he also has the ability and time to also work on deep questions of intelligence. Did he study physics? computer science? something else? PhD? Very curious if anyone knows.

Thanks for the pronunciation of his name. I've been struggling since I bought his book "Deep Learning With Python". Btw, he always sounds paranoid in his book. Am I the only one to feel like that?

Love his cute french accent and his calm, quiet and relaxed way of talking during the entire time!

Lex, don’t know if you’d see this - for visualizing notes you take, there are tools that allow you take notes in text files and make connections between notes very easily, so you get a “knowledge graph” for free. It’s easier to make connections between thoughts I think. For emacs and org (there’s one for md): github.com/org-roam/org-roam

We need another podcast with this based person

I am always puzzled by what exactly is "human level" driving means by different people. It seems by default it means the best human can be, which is very different from what actually is the human level driving in real world. Just because humans in theory can handle the corner case does not mean the average human would handle them perfectly.

One of the few youtube channels that increases my IQ

@47:20 This is what I was thinking. There need to be models of cognition that can interpret BASIC abstract ideas which can then be iteratively developed. AGI is crippled without these.

Astonishing interview. I agree with Chollet's definition of intelligence and I sincerely hope that we will never develop non-deep learning AI. Or at least not before the final version of Neuralink.

The algorithm yet again proved me a fool

The more knowledge you have, the more intelligent you can be.

The fact that during the entire interview (it looks like) your phone screen is on at full brightness with white background is making me anxious :)

It would be great if the guests had some way to write things down, instead of only having to rely on words. For example, when Chollet was explaining mind maps, he could draw an example really quickly instead of trying to describe it.

42:40 Lex is correct here. My guess is that the GPT-3 hype surge was just due to more people getting access to the OpenAI API.

It's been said that babies have their own built in language that is understood by other babies, which eventually gets lost as they age. Something to really think about.

By no fault of the eloquent guest, my auditory processing issues made this difficult to understand. Strangely enough, 1.25x speed improved intelligibility. Something to try for anyone in a similar boat.

I think an AGI needs to be able to „lazy“ learn from it’s environment while it is running instead of pre-learning everything through one shot learning. We humans do the same. We are motivated by something, try to move towards it and start learning when there is a delta between desired outcome and actual outcome. I think we do this by formulating problems and raising questions around those problems. Questions allow us to identify lack of knowledge or training in certain areas. We can start to reason from past experience or „pull“ new information or training required to improve/adjust our skills to be able to solve the problems and reach better outcomes.

Also, human level intelligence is far beyond what we are today: Learning to play chess is actually a solution to another complicated situation, like feed the hunger for getting recognized or get that geeky chess girl.

"You listen to Lex's show?" "Yes." Okay, then you're intelligent.

I ignored the thumbnail twice then I saw the name and clicked in my mind oh he's that deep learning by Manning publication guy 😁

9:36 I have this idea regarding language and AI... I don't know how to properly elucidate it. I don't think language is necessarily the base "language" for communication, even internally. Like.... if you raised a child in complete sensory deprivation, except an artificial set of inputs you determined, I think the brain would be capable of communicating in a purely symbolic or mathematical (or whatever) form. And those forms could reach equal complexity. I think this is significant when you're talking about AI. Because it's kind of analogous: creating these babies in pure darkness communicating along lines of predetermined telepathy. And so... if the goals is to recreate something recognizable as human intelligence, it would be worth determining if language is actually at that base level. Or has just filled that niche and then been reproduced out of material benefits. (I paused the podcast to write that. Lol. So if they address this, and I'm repeating something, I haven't heard it yet.)

14:45 bidirectional search. and it's very cool :)

for real reasoning we will need a intermediate semantic abstraction in the ai model itself - like a symbolic knowledge graph similar to web3

This is so much better than jre

The test already exists: demonstrate the ability to define, identify the unique characteristics of, ever smaller (or larger) "units" of matter.