Andrej Karpathy and Software 2.0 | Chris Lattner and Lex Fridman

Lex Fridman Podcast full episode: • Chris Lattner: The Fut...
Please support this podcast by checking out our sponsors:
- Blinkist: blinkist.com/lex and use code LEX to get a free week of premium
- Neuro: www.getneuro.com and use code LEX to get 15% off
- MasterClass: masterclass.com/lex to get 15% off annual sub
- Cash App: cash.app/ and use code LexPodcast to get $10
PODCAST INFO:
Podcast website: lexfridman.com/podcast
Apple Podcasts: apple.co/2lwqZIr
Spotify: spoti.fi/2nEwCF8
RSS: lexfridman.com/feed/podcast/
Full episodes playlist: • Lex Fridman Podcast
Clips playlist: • Lex Fridman Podcast Clips
CONNECT:
- Subscribe to this KZbin channel
- Twitter: / lexfridman
- LinkedIn: / lexfridman
- Facebook: / lexfridmanpage
- Instagram: / lexfridman
- Medium: / lexfridman
- Support on Patreon: / lexfridman

[@swarajshinde3950]

Next Guest Andrej Karpathy :)

[@sebastianfischer3380]

The idea with the feedback loop of the assertions into the model made my day!

[@MintyMinnie]

Fantastic indepth conversation! Thank you

[@shazzadhasan4067]

when you are going to sit with Geoffrey Hinton , Fei-Fei Li , Andrej Karpathy and Christopher Manning

[@al3030]

I wish I could have friends as smart as these guys and chat about questions like these. Oh well, at least I have the privelege of listening to them.

[@arjunsinghyadav4273]

you got it bro - I am sure you are already doing that with chatgpt now

[@masoodkhanpatel]

very interesting!!!!

[@RobertLoPinto]

This guy Chris is extremely intelligent. I absolutely love the way he cuts to the core of a challenge and articulates it so succinctly. I could listen to him speak all day. The NLP demos today using GPT3 to generate code are definitely fill in the blanks type solutions and don't impress me. They use predefined scaffolding using pattern matching against a corpus of code snippets. When the scaffolding itself is machine generated that will be a quantum leap. Assertions are the requirements and having them expressed in natural language is great, but language is full of ambiguity. The ability for AI to identify and resolve any ambiguities by asking follow up questions would be an amazing accomplishment.

[@arjunsinghyadav4273]

still think the same about chatGPT?

[@RobertLoPinto]

@@arjunsinghyadav4273 What a difference two years makes! Even more mind blowing is Midjourney. Transformers have enabled a real step change in AI, be it text to code, text to image, audio to voice, and vision to vector space (Tesla's FSD). Most recently is Vall-e, Microsoft's oneshot learning with just 3 seconds of audio input that can then replicate a person's voice and speech pattern is incredible and also uses transformers.

[@davidbell304]

I really enjoyed the snippets out of this interview. I've always thought of programming as expressing ideas very precisely, generally once you can describe a solution without ambiguity you are half way finished programming it. And programming languages help you do this by constraining the things you can and cannot say. With GPT3 the question is as what point of complexity does a problem get more difficult to express unambiguously and concisely in natural language than in a programming language? For example, instead of the red button, think of a simple arithmetic calculator app. I don't use html very much but I would think you could do that in ~30 lines without any ui or math libraries. Even with GPT100 and perfect natural language processing, could you describe that layout and functionality, again without special functions/libraries being available, as concisely?

[@Grand-MAGHREB]

I kinda wanna hear ur opinion now after GPT4

[@davidbell304]

@@Grand-MAGHREB Fair question. I have been using just the ChatGPT that comes with MS Edge browser now, which I think is 3.5, and it's very impressive and useful. I would still say the above is at least a little bit true. My experience so far is that if you ask it to write a function with just a basic description it will often do a good but not ideal job. If you ask with a more detailed prompt - specifying the name of the function, the name and type of the arguments, and the type of the return value - then it's far more likely to do a really good job. But one thing I did entirely miss was the interactive nature of these things. You can start with an OK prompt giving OK result, and then say 'fix this, fix that' to build the result you want, and that is really powerful.

[@DustyBilbo]

8:14 When I watched this the first time I never thought we could just chat with it.

[@user-wr4yl7tx3w]

The caption is totally off.

[@anon4153]

From “go tos” -> “if then else”

[@etorawa9367]

Code reuse....lol, yeah welcome to my world. Much easier than reinventing the wheel!

[@hectorarcelus6602]

Could you have a discussion about Spacex and Tesla AI. How do they compare in solving landing rockets and driving cars? The complication seem different, but both are solving 4D problems or are.they?

[@junaidali1853]

Hey Lex, why don't you bring Andrej Karpathy to your pod? That would be a worthy-to-watch interview.

[@gobdovan]

He just did! Check it out

[@lars3743]

Haha. For a long time I’ve also thought that humans pattern match and the logical reasoning side is simply to reverse justify the decision to ourselves. It’s quite a cool trick that we think our decisions were based on logic though.

[@matthewfinch7275]

wonder how much they knew about co-pilot during this conversation

[@quosswimblik4489]

One stream of multiple queries takes you to algorithmic objects on streams and can run the function inside itself and all can be self edited also it can finish with a complex array output. I call it the dynaveo function Multi tied bitwise multi bit piecewise functionality as instructions on a chip could help the neural net AI test upto polynomial time compute space efficiently. Only some military AI has verified what I'm saying so far no people as far as I know have. It is technically an abstraction function for Software 2.0 to get it's MIT's into. I'm hoping a lot of people are very wrong and with the right functionality AI can take us to a mathematician's Christmas/solstice every day. One task I would try Software 2.0 on would be in trying to find fast algorithms to accurately generate faster any of the hard to generate integer sequences on the OEIS.

[@omarnomad]

Next Guest: Edward Snowden 👈

[@ConorFenlon]

It seems like this guy hasn't seen the GPT-3 demos at all.

[@demultiplexdfunc177]

Perhaps Chris knows too much of existing programming paradigms, that he find impossible to step out of what he and many of his colleagues have developed in the last half century. GPT could very well be the beginning of a revolution.

[@hminhph]

Is nobody interested in the demo that Lex has mentioned? Can you share any of those demos where you type in this kind of commands what you want in html and the program gives you source code? For me this idea is mind-boggling and I can not really believe that technology like this was presented anywhere

[@leonardbaumgardt]

Just google "gpt3 code generator", or search on KZbin for videos. I think most of these early tools based on the GPT-3 were taken down because Open AI introduced a pricing plan and a review process for apps based on the API. I would guess these coding tools were not valuable enough to be economically viable (i.e. not useful enough that somebody would pay for it).

[@julianriise5618]

It seems the argument against 2.0 comes from a place of fear, rather than rational, which I can understand. Sure, the simple stuff could be programmed in 1.0, however, the 2.0 and GTP-x will make 1.0 obsolete at some point, just a matter of time im(uneducated)o.

[@pretzelboi64]

Did you even watch the video? "Software 2.0" is literally just an invocation of a neural network in a regular piece of software. It's not a replacement for simple, efficient code that runs directly on the CPU.

[@fredt7518]

So you interview all the "Trevor Milton"s guys who think they are smarter than Elon hahaha !

[@marcusklaas4088]

I usually can't stand Lex Fridman's interviews, but my love for Chris outweighs my dislike for Lex. Excellent communicator and very knowledgeable.