Thank you, welcome!!

The point is not in this case whether ChatGPT gives the right answer, the point is at 9:50 -- whether it gives you the right answer completely depends on if it's seen similar text in its training data. This means if you articulate the same question differently it may or may not give you the right answer, and you'll never know if the answer you are given by it is correct or not unless you know exactly the right answer too. This is very unreliable/scary from a product/application perspective. Also this is why ChatGPT (and lots other AI models too) is useful only in non-critical missions, with humans in the loop. Overestimating the capability of current AI models easily leads to financial catastrophes such as IBM's case with Watson. So LeCun's comment about it being "an extremely good demo" is just spot on.

@@Tbonesteak23 Agree

@@Tbonesteak23 yes I agree with you. And with Lecun. What catastrophe happened with Watson ? 🤔

@@Tbonesteak23 You do know you have just described what humans do, very few humans have any original thought. What makes ChatGPT more than a tech demo is that it is actually useful. Its knowledge is not perfect, but it can program and be augmented by symbolic AI, which I have seen done with wolfram alpha.

@@TheReferrer72 If you know how deep neural nets work under the hood you'll know it's not even about originality, it's about the fact that they do not reason or think. It's like remembering the answer to a question without knowing why. They remember what's in the dataset and try to mimic how ppl talks, however to learn how ppl think is beyond the current state of AI. These models don't really understand what you are saying or what they are saying, everything is just a bunch of transformations that create vectors in the latent space from the input and try to generate text that reads like human-generated text as much as possible, from those vectors. Saying "an extremely good demo" is not saying that it's not useful, but merely the fact that the science and technology behind it are a lot more superficial than many ppl think it is.

great questions btw, i think i love your interviews Alex

@@josgraha Thank you so much Joe! Thanks for watching :) More on the way!

Agreed, LeCun is really thoughtful and this interview was much better then most of the other coverage of ChatGPT.

You are essentially advocating for anyone who uses inspiration to create something new (virtually anything) to be indebted to whoever that inspiration came from. If that's what you want, then all of your art is indebted to another person or multiple. Somewhat irrational, don't you think?

@@andrewxzvxcud2 No, not irrational, here’s why. First, I notice that you’re hinting at a common fallacy which a lot of AI folk try to use. That is “Human inspiration = machine inspiration.” This is incorrect. Humans and machines do not learn in the same way. We do not process data in the same way, and a human hand, no matter how “talented”, cannot reproduce exactly another human’s work. Anyone who has ever tried to reproduce an old master’s work will tell you how difficult it is. On this subject, Karla Ortiz wrote a great article on this, which I highly suggest if you still believe this myth. Second, yes, to some degree we are all indebted to the people who came before us. In the academic world, one doesn’t copy-paste word for word someone else’s work. The proper practice is to quote, paraphrase, and cite one’s sources. Just copying is called plagiarism. In the art world, let’s say I make a study of an old master. For one thing, I don’t then turn around and sell it at my local art fair, claiming that I was the one who came up with the composition. That’s called art forgery. When it comes to references, we don’t just take a random photo we found, we pay to use said photo as a reference, or go out into the world and take the photos ourselves. Contrary to popular belief, art is difficult. One doesn’t just “get inspired” and suddenly come up with a million dollar painting out of thin air. That one painting you see, was actually hundreds of life drawing studies, hours of learning art history, living life and gaining experiences, practicing over and over. This is not something a machine does. The AI can only build an image off of what it is fed. You know what’s being fed into it? OUR hard work. The work of thousands of human artists. The machine didn’t do that work. A stable diffusion user did not do that work. The ARTIST did that work. This is why it’s wrong to not credit the artist, you’re riding off of our hard work. That's not inspiration, it's theft. Machines have no way to go out into the world and experience things. They don’t talk with their friends and family. They’ve never been on a hike. They don’t get in arguments with their spouses. All of these things in a human are subtly picked up by the brain, stored in memory, and put through a human filter. Art is an expression of this experience. It is straight up a fantasy to say AI is doing the same thing, it’s not. If we can’t agree on that, we can’t talk, because you’re not living in reality. A human, to whatever degree is needed, can acknowledge those who came before them. This isn’t irrational, it’s a common practice in many fields, and it’s basic human decency. Refusing to give credit is sociopathic in the true meaning of the word (it’s a lack of empathy, impulsive, and anti-social). That’s profoundly sad that some AI enthusiasts are so sociopathic, that they see giving due credit as “irrational”. PS - One last note: it’s perfectly reasonable to build a dataset ethically. It’s been done before. I have several friends who work with AI for a living, and yes, it is possible. Don’t try to bullshit me on this one.

EDIT: chatGPT 4 actually surfaced the correct answer. :-)

it gets it very wrong when asking in a different way: "A plane starts a level 360 degree steep turn that puts it very close to stalling, from a wind facing direction. The plane does not change the power setting during the turn and keeps the same banking all the way. Is it likely to stall as it gets into the tailwind direction?" It sounds great and smart, but it is just wrong!

Wow, can't thank you enough Figal! This really means a lot. Thank you!!!!

Thank you! Welcome aboard :)

That is not how to query ChatGPT, try asking it if it can write a program that lists all the colors that don't have the letter e. Also ask it to add some test cases.

That sounds like the older version of ChatGPT The Bing variant has reportedly reduced errors

Humans can't do what you asked either...

It can literally do both of those tasks easily and correctly.

Thanks so much, Malcolm!!

Thank you so much, Brian!!! Glad you enjoyed the interview and can't thank you enough for this. I appreciate it!!

Regarding evolutionary priors and active training, these methods can lead to computational efficiency, but they may not always produce models with a deep understanding of language or the relationships between concepts. Parsimony as a loss function could lead to oversimplification or loss of important information in some cases. Furthermore, the desire to leverage symbolic systems that have already achieved a level of reduction might not always align with the goals of creating more efficient or accurate models. It's essential to strike a balance between efficiency and effectiveness, and not every AI system will be able to perfectly achieve this.

I think his point was that the Galactica model was built to achieve a specific set of tasks for scientists and it performed well on those. But because it was asked questions it was never built to handle (nonsense questions), it returned seemingly bad answers. But this just goes back to what he was saying about OpenAI making flashy demos. ChatGPT can focus on the low hanging fruit because they need to create text that the lay person will accept with minimal scrutiny, not experts.

Imagine how small you have to be to make a comment like this. You enjoy the talk and all you can think of is to belittle someone. Do better.

yes

Nope. It can reason.

How is he wrong?

I mean for starters chatgpt has soo many parameters - it can easily explain lecunns paper example (below is the prompt I gave it). The carpentry example as well is off, like a carpenter, GPT and can put together code blocks like wood and create very visually specific websites. “I fold a piece of paper. And then I open the flaps and stretch the paper up horizontally. If I let go of 1 side what will happen?” “When you let go of one side, the paper will fall down due to gravity. The side you're holding up will stay up because of the force you're applying, but the other side isn't being held up by anything, so it will fall. Unless the paper has been folded or manipulated in a way to give it some rigidity (for example, by folding it into a shape like a cylinder or cone), it won't stand up on its own. This is because paper, by its nature, isn't self-supporting; it needs an external force or structure to give it form…” But I mean this aside its very clear to me we have reached AGI. We saw strong potential with LLMs around when Devlin et al published their work on BERT. I do agree with him in so far as we do not have sentience, but I think open ai have done great. Workflows are changing and the technology is basically a discovery like the calculator.

This "dude" is considered one of godfathers of Deep learning and artificial intelligence. He won the Turing award (Nobel prize of computer science). So I think it's safe to assume, he knows what he's talking about

@@christoforosmeziriadis9135 I'm not claiming he isn't smart. I'm saying it's just a shame that his employers greatest use of AI so far has been to sow discontent and deliver nonsense to rubes in exchange for ad dollars. If I were half as smart as this guy is, I would be choosing to put my talents to something more useful than whatever bullshit Facebook is up to next

regardless of what Zuckerberg does, Yann Lecun is saying that ChatGPT is overhyped. It is just a computer program that is good at regurgitating text. It doesn't understand a word it says.

@@christoforosmeziriadis9135 Neither do most of the people I speak to day to day

To me it is ridiculous, to believe that a computer program, which is nothing more than a bunch of equations and words, can be intelligent