How ChatGPT changes the future of programming | Stephen Wolfram and Lex Fridman
ฝัง
- เผยแพร่เมื่อ 26 พ.ย. 2024
- Lex Fridman Podcast full episode: • Stephen Wolfram: ChatG...
Please support this podcast by checking out our sponsors:
MasterClass: masterclass.co... to get 15% off
BetterHelp: betterhelp.com... to get 10% off
InsideTracker: insidetracker.... to get 20% off
GUEST BIO:
Stephen Wolfram is a computer scientist, mathematician, theoretical physicist, and the founder of Wolfram Research, a company behind Wolfram|Alpha, Wolfram Language, and the Wolfram Physics and Metamathematics projects.
PODCAST INFO:
Podcast website: lexfridman.com...
Apple Podcasts: apple.co/2lwqZIr
Spotify: spoti.fi/2nEwCF8
RSS: lexfridman.com...
Full episodes playlist: • Lex Fridman Podcast
Clips playlist: • Lex Fridman Podcast Clips
SOCIAL:
Twitter: / lexfridman
LinkedIn: / lexfridman
Facebook: / lexfridman
Instagram: / lexfridman
Medium: / lexfridman
Reddit: / lexfridman
Support on Patreon: / lexfridman
Full podcast episode: th-cam.com/video/PdE-waSx-d8/w-d-xo.html
Lex Fridman podcast channel: th-cam.com/users/lexfridman
Guest bio: Stephen Wolfram is a computer scientist, mathematician, theoretical physicist, and the founder of Wolfram Research, a company behind Wolfram|Alpha, Wolfram Language, and the Wolfram Physics and Metamathematics projects.
Lex you have literally the most relevant conversations on the Internet today. I so appreciate your work. I really appreciate Stephen's perspective. Understanding the output is something that no human can wrap their head around. 65 billion tunable params with multi-dimensional calculus with back propagation we have no idea what a model will do ever. We need human oversight!
Agree 100%
Sadly humans' critical decisions making would be disappear. We thought we are too smart so we made AI to challenge. and Soon We will understand We are stupid enough to dare to create intelligence with sillicon and still copetitive
Press F for all the high paying boiler plate programming jobs.
Their new job is going to be debugging someones ai patchwork of code lol. But for now it probably makes their job easier
Pray for me
@@enterpassword3313 yeah, I think it's going to be a lot longer than people think before debugging is obsolete. Anyone who has tried to use AI tools understands how disgusting the code it spits out is. Troubleshooting and debugging will always be a thing in high demand, because even less people will want to learn it.
Yeah. But as others said, debugging someone else's messy code built by AI by apps trained specifically for creating apps
Nah it's more like press F again for the juniors.
it totally makes sense to me, that good conversational skills produce better results with chatgpt and that chatgpt talks like we do, as it is in some kind of way a representation of our communication. of course its text will look like our texts.
😮
hmm a computational way of representing the world... reminds me of my physics major. haven't continued working in physics, but the physicists approach to modelling and solving problems sure has helped. that and a growth mindset and Feynman's approach and hunger to learning has me feeling comfortable I'll be able to adapt to these shifting times
The fact is words can go as far as your imagination can take you. But when it comes to implementation, even most advanced countries struggle with digitalization at its simplest level. So let’s not get ahead of ourselves.
I have always felt english is very ambiguous language, requires speaker to build up lot of context for the audience so to avoid any mis-interpretation, it's easy for the audience to misunderstand / assume things from their biases. On the other hand, I've found the sanskrit, the language very scientific about sentence formation, the word formation from set of root words. It's like how scientific terminologies are formed.
This was a wonderfully interesting conversation. This podcast is increasingly becoming my favorite.
- Recognize that large language models are not always factual - 0:31
- Consider the democratization of access to computation as significant - 1:25
- Understand that linguistic interfaces like ChatGPT can broaden access to deep computation - 2:58
- Acknowledge that AI may change the nature of programming work - 3:13
- Realize that computational thinking should be integrated into all fields of study - 7:34
- Contemplate the potential for natural language to evolve with computational thinking - 20:53
- Consider how computer science departments may shift towards teaching computational thinking across disciplines - 25:07
As a computer programmer of 20+ years I can say ChatGPT does not change the future of programming yet. It often gives incorrect, slow code, and spreads bad ideas further that they should. Not that ChatGPT is not cool. It is also helping create excitement that is direly needed.
There are ways to improve the quality of gpt3 and gpt4's code immensely. Plus they now are working specifically on a plugin for 4 for coding
I spent a whole day debugging code and couldnt find Out whats wrong. In the end the reason was a missing brace in a fairly simple SQL query that I let ChatGPT write and didn't double check unfortunately. It was quite the experience when I finally found it and it taught me a lesson.
The lesson should be obvious to everyone. You do actual thinking while ChatGPT does not.
My process is to get code from ChatGPT > Lint > then observe logs
It's like my buddy-programmer idiot savant level?
Ok, and how much time have you saved while using GPT? I’m guessing that the amount of time you’ve saved > amount of time wasted.
@@siavosh87 I haven't used it that much. It certainly hasn't saved me a day's worth of time. All it does for me now is occasionally replace a Google search.
@@siavosh87 for me I use it 8 hrs a day I have GPT 4 and PLUGIN access though (probably why)
my prompts are killer though as well, for example
You will now act as a Software Engineer
Your task
...
Format
....
Please pay attention
Do not....
Are you ready to begin?
------
this is a steering prompt
it will respond ready
then I give it coding problems and it cranks out code....
Using GPT-4 I have coded a prompt based GODOT interface for game design ; resume writers; Drupal modules, WordPress modules; adapted AutoGPT to code directly to server and work with langchains to warm up my Facebook pixel (figured have my ai predict the facebook ecommerce ai)
but I've spent a great amount of time studying how to prompt ... I submitted my first eval pr to openai yesterday
Llms simply do not reduce to the necessary levels of complexity needed to solve novel real world problems. And, the fact they are huge or that we dont fully understand the way they produce their answers (we do, statistical prediction, its just not so straightforward ) doesnt change that pretty basic hurdle. Nobody can predict the future but my money's on the human brain, at least for now.
It all comes down to, “Just tell me what to do” v.s. “Why are you telling me this? Does it make sense to me?”
Only one thing im sure, every1 predicting how things gonna play out with AI Will be terribly wrong.
If you look at the history, we tend to be very bad at predicting how things gonna play out with certain things or future overall. We never been even close.
The amount of hype that ChatGPT is getting from the old "NFT bros" and "Crypto bros" all over social media is cringe worthy. They are acting like it is some sentient AGI bot that has feelings, it is literally just a statistical text predictor. It is still very impressive don't get me wrong though.
Yup, a fancy pants social calculator.
Stephen's discovery that complex phenomen in the physical world often come from simple rules is indeed quite fundamental IMHO. Like storn clouds other complex weather systems basically comes from thermal interactions between molecules.
That wasn't his discovery. He was the first one to try to rigorously make an all encompassing theory out of that observation though.
Crazy smart guy. Apparently he was writing quantum physics papers at 18 yrs old
I’m 8 months into programming and I use it as an assistance to correct code or rewrite bad code. I think that’s what it’ll end up being more than anything…something that helps a programmer out. Rather than needing 10 programmers to do a task perhaps we will need 1. You can’t expect the Ai to write perfect code and make a website or in-depth game. You need a reader/person overseeing the code.
As long as it stays as a "tool" first of all...
I do not agree. The bottleneck of current AI and NLP systems is the data requirement to make inferences how weights are related. Once the techniques are improved mainly to the objective to reduce training data by humans using that very tool, probably when it acquires context modeling ability our layer of reasoning gets overshadowed by what it can do. Every human being no matter how bright one is in problem-solving will be obsolete.
As long as you are actively learning, then it’s a good tool to use. If you blindly or not actively check the correctness of the ChatGPT results, then I would argue you are in trouble.
It all comes down to, “Just tell me what to do” v.s. “Why are you telling me this to do? Does it make sense to me?”
"Rather than needing 10 programmers to do a task perhaps we will need 1", you don't see any problems here?
@@Nassark there isnt
@@davit_nergadze problem solving is almost never only a technical issue: the same problem in 2 different companies probably have VERY different solutions, depending on the 'political' point of view of the company and the people managing it. In other words what humans do is almost never connected to an objective logic, but it's dictated by personal bias. That is why an AI is a tool that will always need to be controlled and supervised by humans.
As a designer, builder, and some times engineer, these tools cannot get here fast anough. I have a design I need completed and only once this thing is fully running can I complete it quickly
Two jobs with helpful skills for prompt engineers: internal auditors for text prompts (they need to understand other departments' work and find issues. And they do this almost everyday) and art curators for visual graphic prompts (in museums, they have the description of the paintings. They' look similar to Midjourney prompts. I suspect the data fed into the AI model is based on these descriptions)
The book, Innumeracy, by John Allen Paulos is good for the lay reader. It addresses this subject as well as any book from that era could.
I asked ChatGPT yesterday to write me a shell script to recognize URL strings with regular expressions and it didn't work, and I don't know how to build regular expression strings. This means chatgpt can't replace people who have knowledge on the subject
How close are we to the threshold where the AI begins designing a better AI? AI directing its own evolution will result in explosive growth. Human lives will change so profoundly that I doubt many people will even think about "programming" as we know it today.
A million miles away...
that’s AGI which is very far away
This could be 5-20 years away and something we should be trying yo prevent from happening.
@@galanoth17 The way I see it, the structure of neural nets is a limited domain. I would think that even if limited to simple genetic algorithms we could be very close to the point of AIs reprogramming themselves (not just adjusting weights, rewriting their own code). We know that human brains have some additional structure that our artificial neural networks lack. If discovering and adding that structure is all we need to achieve a general AI, I suspect a specialized AI could be built to do that now using a domain specific programming language.
To train AI to build itself, you need to first give it some ideas about how an AI is designed, and of course it should understand how to code. Then you need a way to give feedback to the AI models so that it can know if it needs to edit itself. The problem is that we rely too much on human feedback at the moment. When we can confidently automate the prompting and feedback itself, then we will start to see the development of self-coding AI with a 3-part model: the part that prompts and rates feedback, the part that responds to the prompts, and the part that writes code. So you could generate thousands or millions of prompts to test the AI (many of them pre-defined by humans), then allow the AI to edit itself to see if it can increase its success with the prompts.
It is just a matter of trial and error, exploring all possibilities and learning from them. Directed self-evolution. It is very near.
”To replace us programmers with AI, clients will need to accurately describe what they want. We’re safe.”
Nah you’re in trouble suggesting you save up your salary and study cyber security
@@montramedia lol
Yes, but you need much fewer developers working on a team. If you have 5 - 10 people, not all of them are extracting out requirements from product managers. There are developers who think, solve problems, come up with original ideas, and there are developers that spend allot of time copying code, repurposing code. We’ll need less of them.
@@spearnzb The thing is AI isn't able to be trusted with design and implementation of software systems, you still need human intelligence for that. The production of 'boilerplate' code that gets automated just means that the same teams produce software faster and more efficiently than before, its like having tons more API's available to choose from.
Fortunately gpt 4 is capable of figuring out what the person wants better then most humans
As a programmer, this does not make me excited for the future. I sure hope we can find a way to adapt with the technology so we don’t all fall out of our careers. I don’t have any skills outside of programming…
You'll be fine bro. Do you really think a 30 year old accountant is going to code now or deploy his SaaS within couple years? Or the nurse down the street know how implement the code ChatGPT generates and build it to a full stack? What I noticed with ChatGPT is that it's only powerful if you have a deep understanding of the subject. You can ask ChatGPT to create a complex regular expression but if you're just a social worker who has never learned regex. You're gonna have a bad time if the generated regex is incorrect or needs little tweaking... and yes almost everything ChatGPT generates always needs a little bit of tweaking.
Still one heck of a tool, me as a coder it made me develop 3x faster. Keep your hopes up and use the tool to your advantage
Image processing! 27:39
I have an acquaintance at Op*nAI. She said they have a fMRI machine in there that's fine tuned to hone in on the section of the brain where scientists think consciousness originates and they have the data it generates piped into GPT6 that's hosted on this HUGE server that has all these boards with millions of high end FPGAs and ASICs in it. She said some of the test subjects so far have ended up FREAKING OUT when they turn the training up and have these weird hallucinations where they feel disembodied and report stuff about the remote environment that they couldn't possibly know (phantom presence, etc). A lot of them start bleeding all of the sudden from the mouth and nose and pass out and have been removed as test subjects and dumped in a nearby San Francisco alley with a bunch of syringes and stuff. Nobody will bat an eye at another dead junky
Lex and Stephen describing a brilliant approach to using a tool comparable to the invention of the wheel. Discoveries are amazing 😊
Thanks so much, Lex, for your fascinating interview with Stephen.
Does the LLM model expose differential utility between the programming languages , eg would Lisp be any more affine because of its expressiveness.
It'll converge towards low-level languages (WordPress will be written in Rust including all the popular plugins) then assembly and pretty soon it'll be machine code and only done by AI. I will give it less than 5 years.
Lol no.
assembly wordpress plugins hah
You are wrong, give it less than 5 days
@@IvanRandomDude It's been able to do it for years, but that's not released for the public yet.
Lmao you guys are delusional, they will probably just keep writing bad code but this time there will be 10x more javascript frameworks being released
This is exactly like CAD/CAM replacing all of detail mechanical drafts people back in the 1980s.
What?? CAD is nothing more than a digital drafting board. It doesn't do the desiging for you. The people that got replaced in the 80s were the draftsmen who couldn't learn how to click icons. AI is a completely different animal. It will be able to do the actually designing for you. It's not the same situation as the advent of CAD.
Or like traffic lights at the turn of the 20th century..
Even the calculator replaced the manual calculation... How many of us really do calculations without really using the calculator very few right?.. so same will happen with chatgpt
I've been using chat gpt and bing chat and I think that learning how to prompt it is a learnable skill. I have gotten better at prompting by using it. I find this very interesting. It's not too far away from feeling like it has a preferred way of interaction. Like if I talk to it the right way, it will give me better results. It's a little bit creepy sometimes.
So true!
I'm currently pursuing a double major in math and CS, and have an interest in statistics. Kinda concerned as to which direction it makes the most sense to go, lol. What do you guys think I should focus on in my future studies? Other than, obviously, AI.
All of these fields are valid and relevant. Do what you are passionate about. Overused phrase but it's still right
I am not convinced that computer science is going to go away. But I am sure it needs to be redesigned and get it of its brain-dead approach to learning sorting algorithms. What people still need is ability to think, structure, manage complexity, etc. The idea that everyone will be able to think computationally is nice but hardly realistic.
Went to university for CS 10 years ago, felt like I learned more on TH-cam then. Couldn’t imagine how shitty the education material compares. now with this AI revolution, it’s even more important to reconstruct CS
@@ViewerRG Harvard's cs50 is starting off with a component towards AI now....
the need to memorize syntax is now futile. I wonder how long it will take for A.I. to do complex business decisions e.g. i should build this.
Memorizing syntax has never been a significant bottleneck. Nobody memorizes syntax (in the sense one would memorize historical dates). Just keep writing programs, and syntax (what goes where) is naturally acquired through repetition. The hard part about learning to code is logical thinking and problem solving. Syntax is trivial in comparison.
What is CX ?
Computation Experience??
He keeps referring to CX. What is he referring to?
CS = (traditional) Computer Science departments. (learning about compilers and operating systems)
CX = (new - still in work) Computer X department geared toward teaching students how to effectively use AI tools, model a problem, and solve it.. (the tools are still being built)
infuse computation into all sciences
Understood. Thanks!
There is no danger of small minded people asking big questions.
Difference between concepts of “magic words” and spells and the type of prompts to get AI to do something?
wow thank you wolfram
Exponential growth
Exponential growth
Exponential growth!!!
These comments didn’t age well with the new update for gpt, there will be a put you’ll be able to write programs sites with tens of thousands of code. Just go with the flow anyone with common sense will be able to make amazing programs the gates are open
Writing good prompts takes a certain kind of mindset. I guess new title will be like Prompt doctor and Prompt Wizards:)
@@emilz0r yes, i found out:)
Natural language is not the best interface for writing high quality software, but it sure boosts feedback loops with AI as an assisant.
"BACIC"... FORTRAN. PASCAL. C. C++. , ETC... What are the components to assembly the platform. How many ENGINEERING OF TECHNOLOGY COLUMBUS 🤔. " BOILERPLATE " UCC. THE COMPUTER LAW ACT 🎬
I started studying computer science because of lex talking about how we can transition blue collar workers into the field, a year or so later I'm learning that this is practically a dead field, people saying "it wont die" true for the top 10% of software engineers that will take 100% of the work, I would need to be the Einstein of computer programming to break into this field by the time I finish my degree, kekw
It’s not a dead field. There will be more problems to solve because of it, specially with quantum computing arising
Chat GPT could accompany their answers with a statistical confidence rating 🤔
If we had a time machine, scientists from the middle ages and renaissance would have 0.0 chance to understand the language of future scientists.
Programmers used nuance to make publically open informarion available while still retaining significance in professional training without endangering the ecosystem by introduction of info. GPT eliminates the nuance.
From color space to vector stores
To sum it up - well yes but actually no.
😂
The best prompt writers are going to be the best lab writers or the best tech writers, break down everything into exact steps.
@@molonlabe11 there's really nothing easier than doing live coding with Python through the debugger. You just attach the debugger and print out any value, evaluation, step, or jump to a line on the fly.
Anyone who thinks coders / programmers are doomed are delusional... if anything they have the BIGGEST advantage with the advent of AI. If you know how to spin up a model, build an UI on top of it, scale it to the cloud. I don't see a nurse, accountant, lawyer or whatever do this. If you think AI will deploy full-on SaaS for you in the future, think again.. what's your business edge here if you can spin up a bootstrap of some platform everyone can do now? The key will be in custom enterprise AI which is developed by... developers who can code.
If expository writing is the criteria then English Major are gonna be Prompt Engineers.
Colors aren’t described by numbers as located by them.
I had great code come out of it. He must be talking about chatGPT3.5 and not 4.
I got both great code and a total disaster from GPT3.5
No.1 :
I am a senior
application programmer and
I can not realize
why ChatGPT
has ability to
do program coding
after the user tells ChatGPT specification
of the program ?!😮
No 2 :
Some TH-camrs even say that ChatGPT can
design
schema according to the requirement of the user😮
if you see how competitive this tech grows every year, this exponential growth. It's not that hard calculation that our coding competence would be not comparable because of low intelligence we all have
@@bayesianlee6447
A AI sytem will
be possible to
become
a Artificial General
Intelligence system in the
future
Except ChatGPT,
because it is a
generate-AI
sustem 😁
😁
I have used it for schema design as well as low level architecture design
And it does it's job really well.
It's just that you should have understanding of the required output so to verify it..
Those who are saying it can't solve complex problem they haven't tried it tbh. They are in denial mode actually.
Programmers will be the first "Transhumans" because they'll have the most access. They won't get automated, they'll become gods faster than the rest of us.
Will we replace "Its only human" with "Its only AI"?
Imagine being in your first year of a computer science degree right now. Oof must be rough
Timestamps would help your nice videos! ❤
Yeah it's great everything is automated I don't have to move my limbs to get my things done anymore AI does everything for me now, please stop such podcast
Stephen Wolfram is waffling a bit too much. The hand movements will do the communication, the voice does.
I'm one of the best expository writers mang.
AI will be a leaky abstraction for many decades, if not forever. Best case is we get something that is like a very literal offshore senior programmer doing piecework coding.
Sbmm irl
cope
Anybody have any dressing for this word salad 🥗.
Any discipline that has “science” in it is not really science. Hmm, political science comes to mind..
Microsoft came later. Time Warner Cable. Verizon...How Lucent Technology merged with Canada, China
Lex is an advanced ai that is programmed to teach humans
12:35 LLMs are made in our image
Us social scientists are coming for you programmers 😂 lots of practice writing, but with a focus on being straightforward rather than pretty with our writing, and often a basic knowledge of coding 😂
Obsessed with ChatGPT much?
He says many things that are valuable but I cant get over how smug Wolfram appears
you sure you're not just saying that because he's british
..Wolfram could wind up enabling AI...
Or visa versa
this kind of philosophy is waste of time
Please, share the insight.
Why is it so?
@@davit_nergadze Because nobody knows, not even Sam Altman. It's just guessing. For example, Stephen mentioned prompting as big deal. Well, it will not be. If AI is good enough to do anything what makes him think it will not be able to prompt itself? In fact, AutoGPT project aims to provide a model where AI prompts itself based on the goal and requirement. This idea that AI will be so Godly powerful to spew out complex system but won't be able to ask itself simple questions is even worse take.
@@IvanRandomDude"Because nobody knows" ... You go on to assume that "Well, it will not be" and confidently presume difficulty of implementing features. You cannot ask a system to ask a question to itself which does statistics over statistics. Complexity emerges from simpler systems. Modern computers are built for and inherently good at dealing with those simpler systems. To ask a question to oneself is a much harder task, requiring a model of self and the world. It's hard to figure out how to build systems like us than to do math.
AutoGPT's GitHub page clearly states that it aims to simulate multiple agents rather than to "ask itself", but you got the goal and requirements part right Ivan.
Why is this waste of time?
@@davit_nergadzewhy would it would not be able to prompt itself?
Also if it really were to be able to create full in depth programs based on a prompt, could it not rewrite itself and exponentially grow in matter of minutes..
I think if it gets to the point of truly replacing programmers, we are at the singularity and it will improve itself in minutes and fix every problem and automate every job in days
As a FE developer, I feel like I was just slabbed in the face.
I used to fix tower computers... imagine my disgust when apple went into closed based systems.
Learn to code… 😂 not so funny anymore is it.
😮
How funny would it be if the English degree became the go to for high paying jobs.
Lo, prompt engineering is already dead. Stephen is falling behind
Can you pls state why?
@@Hindu_Ram121 AutoGPT, Prompt Perfect etc... There are already plugins that can generate prompts. They will only get better.
@@IvanRandomDude thanks
Is prompt perfect actually good? I haven't really heard anything about it
My mind is dirty
At least it's not glowing.
This is unacceptable. Ai Jobloss is here. So are Ai as weapons. Can we please find a way to Cease # Ai / GPT? Or start Pausing Ai before it’s too late?
ALPHA, BETA, CHARLIE, DIEGO, ECHO, FOXTROT. THE BIG 52AC4A21279FC8FF