François Chollet: Keras, Deep Learning, and the Progress of AI | Lex Fridman Podcast #38

[@lexfridman]

I really enjoyed this conversation with François. Here's the outline: 0:00 - Introduction 1:14 - Self-improving AGI 7:51 - What is intelligence? 15:23 - Science progress 26:57 - Fear of existential threats of AI 28:11 - Surprised by deep learning 30:38 - Keras and TensorFlow 2.0 42:28 - Software engineering on a large team 46:23 - Future of TensorFlow and Keras 47:53 - Current limits of deep learning 58:05 - Program synthesis 1:00:36 - Data and hand-crafting of architectures 1:08:37 - Concerns about short-term threats in AI 1:24:21 - Concerns about long-term existential threats from AI 1:29:11 - Feeling about creating AGI 1:33:49 - Does human-level intelligence need a body? 1:34:19 - Good test for intelligence 1:50:30 - AI winter

[@megaconus9174]

Great content as usual, Lex. I would like to know if there is any possible way of contacting you in private for a possible interview/conversation.

[@kwood1112]

Thank you for this - very helpful!

[@JoeWong81]

Lex Fridman you are absolutely killing it with these interviews lately!!!!. Elon Musk, Yann Lecun and now Chollet (3 of my fav twitter accounts). Keep it up bro!!

[@strongsyedaa7378]

I want his number I am becoming a Data Scientist

[@sapienspace8814]

@ 34:04 That is interesting, "hiding of details as magical", wish it could be expanded.

[@klimentserafimov9924]

This guy is the most grounded researcher I've seen.

[@vineetgundecha7872]

This is like Francois tweeting for 2 hours

[@kevthedestroyer1044]

well he didnt go into politics so it's better than that :D

[@Wardoon]

10:50 "Science itself as a system, as an institution, is a kind of artificial intelligence, problem-solving algorithm that is super human"

[@thelonespeaker]

@@kevthedestroyer1044 Does he often dive into politics?

[@kevthedestroyer1044]

@@thelonespeaker that's what he does on Twitter

[@thelonespeaker]

@@kevthedestroyer1044 Yeah I've just checked I don't like it though... politics are such a complicated issue, taking sides like he does kind of discredits his points about intelligence and the necessity of listening to data

[@grungecrunge]

My favorite podcast right now, I can’t get enough info on this topic.

[@beta5770]

Most human like definition of intelligence for me

[@metafuel]

I fully agree.

[@RalphDratman]

Chollet makes an important point when he says that human intelligence is not applied to substantial tasks by single individuals but rather by groups of humans collaborating. This becomes clear when you consider how error-prone the work of a single individual may be. Everything has to be checked. The situation of "knowing" something is a group process too, because individuals have to agree on what we know before it becomes fact. Even for solitary work, such as that of an artist or a mathematician, the product of the work must be confirmed by others as good or correct before it becomes useful or respected.

[@dr1ce]

The quality of your interviews are outstanding! The calabre of your guests are amazing. Keep up the great work!

[@andreipitkevich1088]

Amazing interview! I especially liked how François exposed the idea of ​​an exponential increase in technological progress.

[@chantelleboutin]

Hi Lex, I appreciate how you challenge those you interview. This is real journalism that you are doing. Thank you.

[@peterpetrov6522]

He makes great points on why the singularity might be a pipe dream. Reminds me of a joke: A tourist asks 2 cops for directions but they couldn't understand his language. Then he asks them in English, French, Spanish and they still couldn't understand. He leaves and the first cop says 'We need to learn new languages so that we can communicate with others.' And the other cop says 'That guy spoke several languages. Was he able to communicate?!'

[@PartyRockAdviser]

A fantastic sci-fi story I read a long time ago ended with the moral: "To ask a question you need to know half the answer." I never forgot that story.

[@Hexanitrobenzene]

@@PartyRockAdviser That's insightful. What is the title of that story ? Researchers often say that half of solution is formulating the right questions.

[@PartyRockAdviser]

​@@Hexanitrobenzene Sorry. It was a sci-fi short story that I read several decades ago and I forgot the title. Some aliens had evolved to the point they were leaving the universe. They wanted to leave behind a machine that could help the younger races by answering any questions they might have. Several races came to visit the machine but every time they asked a question, using their own language and their conceptual view of the world (i.e. - their current level of technological advancement), the machine could only reply with a statement that indicated the terms they used in their questions were empty constructs derivative of their currently limited knowledge base, and therefore too inaccurate to be given a proper answer. At the end of the story the machine is frustrated that it can't really perform it's purpose, because it realizes that its creators failed to understand a crucial premise. That to ask the right question, you need to know half the answer. My extrapolation from that is the realization that knowledge is a never ending chain constantly building on itself. If you are too far behind in the chain, you can't ask the right question. Too far ahead, you can't explain anything (Imagine trying to explain a piston engine to a goldfish, even if someone gave you the ability to communicate with it).

[@Hexanitrobenzene]

@@PartyRockAdviser Oh, several decades ago... I simply assumed that "never forgot that story" also includes the title :) On a side note, judging by current progress in AI, a day where you could write this summary to Google and it would come up with the title is not too far away.

[@PartyRockAdviser]

@@Hexanitrobenzene You're probably thinking of OpenAI's GPT-3 system, which was recently licensed to Microsoft. It does exactly that and the inverse too.

[@javierfernandez6327]

THANK YOU, Lex and Francois. Amazing discussion. I'm walking away a little different after this.

[@oncedidactic]

The word for this interview series is "exceptional". Thanks again Lex

[@Lagruell]

This was an enlightening interview, very interesting. Thank you!

[@kozepz]

Great talk again! I'd like to hear a future conversation about the use of petrinets in a.i. for practical productional uses.

[@SudipBishwakarma]

Gotta admit your channel is a goldmine. Love your podcasts.

[@robertpoole9258]

Your channel's fantastic man, these interviews are amazing. Keep it going.

[@deeliciousplum]

Lex! Look. You have to slow down the upload pace of these exceptionally insightful interviews. I am still recovering from the mind expanding interviews that you uploaded from a few months back. François Chollet's ideas and concerns have added so much to the tools that I pull from while exploring A.I. topics. Brilliant interview. I'm picturing a whole busload of KZbin content creators actively conspiring to tackle you as you are constantly setting the integrity bar higher and higher. 🌻

[@monyflickyoloo4058]

I want to learn how to explain my speciality to the world like this guy.

[@connorshorten6311]

Wow, another great guest! I'm not sure if I would have continued studying Deep Learning if it wasn't for the Keras interface making it easier to get started! The Deep Learning with Keras book is also such a great resource to get a sense of many different applications of DL!

[@PierreH1968]

The best insight on AGI and science improvements density management I have heard. Thank you to both of you

[@alex_smallet]

This is the best speaker on that channel so far. I really liked how articulate François is in explaining his point of view. I could not agree more with almost everything he was talking about.

[@theredflagisgreen]

Thanks for playing DA in the beginning, really helped me understand his issue the situation.

[@GeorgeOu]

"Exponential Progress triggers Exponential Friction" -

[@piyh3962]

For every Moore's Law, there exists a Rock's law

[@nikopumo-ri9684]

Really interesting discussion, full of new ideas and unexpected perspectives. I did not expect that Francois thinks so deep about AI. Enjoyed this interview a lot...

[@Alp09111]

New perspective ,brilliant ! Thank you

[@TomHarrisonJr]

biology, evolution, physics, politics, history, language, business, astronomy, mathematics, sociology, philosophy and even a bit of computer science. What an incredible, expansive and informative conversation between two brilliant humans. Thanks for increasing my confidence in humanity.

[@yeeboy217]

Incredible work interviewing Lex ... you seem to ask just the right questions at the right moments to keep the guest going , and pull out the highest value information and the most improtant points out of the guest .... really appreciate these interviews and the quality and care you put into the interviews and into every single question and all the follow up questions .... awesome rhythm with the guest, and highly enjoyable podcast for neophytes in these fields to advanced people in this field ... keep up the awesome work man thanks so much !! -Rafif

[@sanesanyo]

Thank you Lex for such a great podcast with Francois. Really refreshing perspectives.

[@NisseOhlsen]

Looking sharp and spiffy as ever, Lex, keep us the good work!

[@alicapwn]

Lex is the Joe Rogan of AI podcasts

[@christophersura3642]

Really quality content in this podcast! Well done.

[@Wintergatan]

Thank you Lex for all these awesome videos

[@jeff_holmes]

Another great interview! I really like the idea of allowing users to see and customize the algorithms used by companies such as Google and Facebook (around 1:17:00). "Open algorithms!"

[@unltd_j9018]

This channel is so cool. Would love an interview with Wes Mckinney on the future of pandas.

[@marekartist8441]

Thank you François. It's been a pleasure to listen.

[@SergioArroyoSailing]

I just have to say :) You have the best interview videos ever! Keep them coming ! :D

[@Wardoon]

25:37 "For many people, AI is not just a subfield of Computer Science. It is more like a belief system."

[@dridri4664]

Again a very interesting talk, thank you Lex!

[@hannahlambdaa]

Thanks for this interview, Francois is a very smart person and I appreciate the chance to listen to his thoughts...

[@blakebaird119]

best version of Lex I've seen in a podcast

[@BunDinYo]

great video and a great guest. your videos are always inspiring. Maybe I can ask a question here, how to get more involved in the deep learning topics. simple implementations and work with pyhton, keras, numpy, pandas and etc. run quite well. but how does one go from a beginner level to more advanced? how can i better grasp and truly understand the theoretical basics, let alone the different mathematical basics, many books are either too "basic" and others are just too hard for to follow. maybe someone has some hints for self study.

[@hamjii]

I really like Francois' sugar-free opinion on AI, that is the real thing from the real expert ! Wonder how I found this one so belatedly... I'll have to look up every and each interview of yours, Lex ! Another fantastic interview.

[@okko7788]

A refreshing take. Nowadays, It's getting scarier to see all these people on the internet talking about a singularity as if they were waiting for the second coming of christ or something. I've always felt uneasy looking at these takes but wasn't smart enough to put it into words. He did it perfectly.

[@Coeurebene1]

brilliant discussion, gave me a lot of food for thoughts, thanks!

[@ayanghosh8226]

This is an awesome interview. Everything that Francois says makes sense, its concrete logic. I did find that the initial line of questions being put to him a bit repetitive, same things being asked again and again. But indeed, Francois’s words are truly enlightening.

[@hv1461]

I am so glad there are such thoughtful and intelligent younger people. There is hope for the future.

[@adamgoldstein3015]

Great talk. Love the way Chollet thinks. Data and hand-crafting of architectures is a great conversation.

[@ocel12356]

Love this guy's approach to online content for oneself

[@jorenboulanger4347]

The best one yet! (The Jeff Hawkins one is second.) So many valuable stuff in here. Thanks!

[@WilianRibery]

2 hours with the python boy are exactly what I need for my saturday night

[@osemudiamedawodu3113]

Dyon Liekki sack boy! I loved that game. It’s the reason I got into programming. Remember logic?

[@denisgolovin5928]

Very interesting interview! I find François' argument where he compares science to a highly sophisticated problem solving system quite appealing, but there might be some major differences between an AGI and the 'problem solver' science. The AI Safety researcher Robert Miles presents in his video 'Why Not Just: Think of AGI Like a Corporation?' some differences which could explain the linear progress in science.

[@Torulv]

Isnt science the measurements of the physical world. And we use logic, math, deduction to theorize the possibilities of how it might work and then test. If AI is given ability to find out how to measure the world better and find ways to assemble matter better in all its various applications, isnt that enough for something close to a singularity. I feel like sometimes the sceptics gets tangled up in misleading analogies and methodology, me included. Its just something about his perspective that doesnt seem right...

[@svenoscarhansen8847]

Would love to hear you question Rupert Sheldrake. The Past exchange would be quite fascinating!

[@andypotanin]

44:32 "making design decisions is about satisfying a set of constraints. But also trying to do so in the simplest way possible" - elegantly put.

[@Marcos10PT]

Re listening to this in the context of the ARC challenge. So relevant!

[@GodsNode]

I love this guy. His twitter is the best.

[@RANJEETTHAKUR1983]

Thanks alot, This is amazing content. As a suggestion and expectation, I would like to see 2 people - Sergey Levine and Andrej Karpathy

[@cristiannapole8393]

Sergey was on the 108th episode

[@RANJEETTHAKUR1983]

@@prashantkumar9139 Can you share more details?

[@RANJEETTHAKUR1983]

@@cristiannapole8393 yeah Thanks for sharing

[@b.f.skinner4383]

François is one of the most brilliant minds I've ever had the pleasure of listening to. How he doesn't have a wikipedia page yet Kim Kardashian does makes me very upset

[@gauravdwivedi2808]

I was shocked too.

[@Hexanitrobenzene]

Make one :)

[@Likexner]

They say dont read the comment section, but thats definitely not good advice for this chap. Youre doing an awesome job, Lex. I miss the technical in-the-know things as i have no education in the field, but i love thinking about the principles discussed in your podcasts. The thousand brains theory one was my favourite. Keep going at it! Thanks for your hard work. Peace

[@EmileAI]

Giving the users the choice to customize their AI (youtube recommendations etc) is absolutely mindblowing I can't even understand why I didn't think of it as something extremely important for me and more broadly the users of those platforms

[@man.h]

I love his book! Thanks for this amazing content. Btw. hope to see you soon on the JRE.

[@alexhri]

BEST PODCAST EVER

[@RalphDratman]

Chollet is a fascinating thinker. Excellent!

[@JousefM]

I like Francois :) Thanks for the great interview again Lex!

[@abelgeressu2709]

Lex Friddman , Would you invite Ben Goertzel and le him defend Francois Chollet argument?

[@kasibert2229]

I'm at 11:05 and I stopped because I have to say this: this guy seems like a genius

[@GauravRawat14]

The point regarding if you improve part of the system the other becomes a bottleneck, is so true in so many aspects of computer science and technology in general, great talk 👍

[@En1Gm4A]

great podcast thx

[@alanrobertson3172]

Brilliant guest.

[@Crimsonwhocares]

I was exactly thinking of francois after you had lecun.

[@Consural]

Genius. I wish I was his student.

[@aahm6975]

here to give a like asap, then watching it

[@nirajabcd]

Finally Francois Chollet!

[@jinmanzhao3696]

Oh god, they talk about DL-guided program synthesis! That is my research topic! I know exactly the paper behind Excel FlashFill (58:54) and I'm working on a paper on DL-based code completion. I've been watching this series for like 20 episodes so far and this is the first time I've seen people talk about this. So excited!

[@armanmkhitaryan27]

1:41:58 "Try to picture a game where no matter how much you play this game it doesn't change your skill at that game." Love actually sounds like a legitimate candidate. Not flirting but genuine love.

[@user-gc6my9jg2c]

very interesting. I agree with a lot of what he said

[@badboat04]

Love what Francois brought up at 1:16:31 . We've been working on an interface for controlling our RL model on news recommendations. It currently looks like this: ibb.co/Ht3cjXV We're working on a slider next to control the degree to which we add "random" news instead of the RL model's. Random being a list of top 40 publishers, where we filter to have one article per source and try to avoid duplicates. The news is currently down, having issues with the API provider. Hopefully coming back soon.

[@ericzong1189]

wow.this young man is sharp asf.machine-like;)

[@ivavrtaric913]

GENIUS

[@hrsight]

Thanks

[@Tetsujinfr]

The occurence frequency of the term exponentially is exponetially growing after this video...

[@sherrivonch6231]

Taking info in is ok but can a person use it when they need to.

[@NadjaNiemann]

.. thanks for not mentioning him. Although his name was in the air quite a bit.

[@danellwein8679]

Lex … I finally figured out who you remind me of … the great chess player Robert J. Fischer …

[@syedabdulsalam4659]

The goal of the researcher must be to solve problems and innovate, he/she will get the bucks. However, if things are the other way round to publish and get big salaries rather than focusing on actual problems, the resource consumption will rise exponentially and productive output will remain flat. I agree with him.

[@exacognitionai]

I find myself agreeing with Francois including (mostly) about intelligence explosion. What drives human intelligence to adapt and to peak (i.e. right time, right problem) is stimuli and response mostly driven by our interaction with our world. We are at times driven to achievement & at other times not. Very intelligent people are all around but they need the right combination of stimuli, motivation & response. My belief, however, is this can be created in a self aware artificial cognition that is appropriately motivated (i.e.contextually relevant motivation such as survival). Complex, possibly out of our current reach but not impossible.

[@cheval63sg]

I would like hear Chollet in French. :-)

[@2894031]

Wow sounds like very original thinking 😳 so rare nowadays

[@Elian-]

What is it with famous French AI researchers and the non-stop expression of opinions on Twitter

[@GodsNode]

BeCahls.....whe'Ahl Fwehlch....*smokes cigarette*

[@sandeep7305]

rasko = take a note...... in Telugu language.😀

[@Elian-]

@@sandeep7305 glad to be learning my username has a meaning! And also glad to (shamefully) learn about Telugu

[@neutrinocoffee1151]

N = 2

[@michelinstarschallenger8626]

It’s our little Napoleon/Trump side.

[@CapitanFantasma1776]

Wow! Create the right Environment!

[@dbporter]

How is the universe an information processing system?

[@allurbase]

WOW this guy systemic view is beyond anything i've heard

[@rwang5688]

On what’s surprising about deep learning (paraphrase): “It is a shock that deep learning works.” On Keras: “Building scikit-learn for covnets. Hiding a lot of details but that’s the point.” (Yep, information hiding is fundamental to good software design.)

[@bearwolffish]

Impossible to define intelligence void of environment. So true.

[@nazgulizm]

He thinks his own thoughts. Very interesting conversation.

[@joyceagresta9829]

"Both the man of science and the man of action live always at the edge of mystery surrounded by it. " J Robert Oppenheimer

[@joelomondi5815]

Great expression of intelligence

[@dylanhirsch-shell9977]

One day podcasters will be asking guests about "The Chollet Test" instead of "The Turing Test".

[@nikuman8058]

I realize there's almost no change you'll read this Lex, but hello internet. So far the efforts in the field of AI only converge toward modeling instrumental rationality. Playing go =/ playing go. Lets strive towards a world where education is not just a privilege of those already showing an aptitude to learn.

[@AltumNovo]

Isn't intelligence all priors. Intelligence is a form of knowledge not an attribute that can be separated out from the other properties of the system/agent

[@neilbatten8453]

40 talks in on the subject of the possibilities of ai dangers, and not a single mention of nueromancer or Gibson.

[@matthieucneude5761]

Did I say that this podcast is brillant? Well, it is. I'm trying as much as I can to take recommendations from human beings. Not algorithms.

[@sherrivonch6231]

Someone may not wish to solve a problem and choose to not participate. Especially if the outcome is already known.