Yann LeCun on How to Fill the Gaps in Large Language Models
In this episode, Yann LeCun, a renowned computer scientist and AI researcher, shares his insights on the limitations of large language models and how his new joint embedding predictive architecture could help bridge the gap.
While large language models have made remarkable strides in natural language processing and understanding, they are still far from perfect. Yann LeCun points out that these models often cannot capture the nuances and complexities of language, leading to inaccuracies and errors.
To address this gap, Yann LeCun introduces his new joint embedding predictive architecture - a novel approach to language modelling that combines techniques from computer vision and natural language processing. This approach involves jointly embedding text and images, allowing for more accurate predictions and a better understanding of the relationships between original concepts and objects.
Craig Smith Twitter: / craigss
Eye on A.I. Twitter: / eyeon_ai
why would you answer a phone call in the middle of an interview 29:00
Shows the age of the interviewer. Kind of cute.
Please turn off your phone the next time you do an interview. otherwise, great conversation with Yann LeCun.
I'm very confused by this channel. Some of the most brilliant minds in the industry are willing to come on the show. The interviewer knows what he's talking about and asks good questions. But then the whole thing gets cut up in a way that makes it extremely strenuous to watch and listen. And just to top it off, Yann LeCun, who takes a chunk out of his day to do this, gets told to hold on while the interviewer answers his phone?! Leaving the sound on it a rookie error but answering the damn thing while he's talking is just absurd. So you assume the whole thing gets pulled through (pourly performing) auto-cut software. But then it looks like somebody actually took the time to mark chapters in the video but the embarrassing moment of answering a phone is left in the video while there are about 20 cuts per second...🤯 End of rant. 😅 Amazing opportunity for quality content here but please leave the interview uncut and let inpatient people play it at double speed. And for the love of God turn off your phone during an interview 😄
Fantastic interview!
Can you use subtitles with action props to describe the next step in a movie that way it would be Language that’s directing the next event with something like that be possible?
Which paper is he mentioning on 51:59?
In my opinion visual input is just another form of modality, which is not necessarily more truthful than other forms. Good to have, but grounding solely based on visuals also have the same issue just as generative models
You have no idea what you're saying, do you?
Why is there so many short-cuts
Its super annoying. Otherwise great interview
Makes it sound like a deepfake
You don't need an emotion like fear if you can think quickly enough. Fear is a heuristic.
10:30 most of what we learn has nothing to do with language but rather common sense
And how do you learn vining sense without language? You can’t even define it without language. Our brains are pre-programmed to learn language, and it’s the same pre-programming that allows us to recognise and respond to music from as early as the womb. There’s a tower of language that suggests it came from music, certainly it shares much in terms of rhythm and cadence and lights up the same parts of the brain.
What about taking Ai with you through a day with a human explaining everything that’s happening as it’s happening so that the AI can tag along with being a person with something like that work?
AI will be embedded into all our phones and computer systems soon, including ERP systems, so they will learning all that stuff almost by default
A few points on this 1. You see things differently and a Machine see things differently this is true for both Ideas, data, inputs many things 2. We ourself don't know why we are doing things half of the time, why the decision that we made has been made, why I took left on a road, why I am breathing fully etc 3. Quantization of Data, input, feelings, emotions even if we substitute every nerochemical to value it is still hard to quantize them 4. Processing of the input too is very different in a human brain then a robot(Computer ) So yeah, the approach you are saying might help but it will be incredibly taxing on you ( the explainor) and might not give the results you desire. I welcome all improvements on this answer
Current AI systems require massive amounts of data. You'd need to take them through millions of days to learn much.
What did Beauregard want?
To be famous.
What's the architecture that he keeps referring to? Jon tiang architecture? Jon ting architecture? I couldn't understand, he was speaking really fast 😂
joint embedding architecture, or JEPA(Joint Embedding Predictive Architecture) without "P"(Predictive)
I think setting them goals is not a good idea. Sounds problematic.
Ali Taj Ghahramani.
Moral: researchers still have zero idea of how to build intelligence into machines. What a shock!
But not impossible!
Reversible computing is the future.
mmhhmm mhhmm mhhmm mhhmmm mmhhmm mhhmm mhhhhhhmm mhmm mhmm mhhmmm mhmm mhmm
30 years living in the states and his accent still sounds like it's his first day lol (I'm French, don't hate me)
Only french criticize how people talk
Mad libs
Quite a bit of what Yann says here seems obsolete, according to GPT-4 release
how is that? The way gpt4 works is by converting the visual representation to text and then work on it. It still isn't a proof of conscience
@@Bargains20xx really? OpenAI didnt disclose how they integrated vision
@@Bargains20xx certainly seems to have some theory of mind. There seems to have been a collapse in the definition of how we’ll know and what tests are valid for consciousness, and a split into practical sentience as opposed to philosophical sentience
AI is moving so fast anything you said last week is likely to be out of date this week
Yann is right on quite a lot of things here. It depends on what you define as AGI. If you want something that is as true as possible to the human mind , his approach of grounding it in reality makes the most sense
Lame academic takes from the paperwork guy with zero practical tech industry experience.
he is the chief ai scientist at meta
@@sinenomine3710 I know that.
The comment appears to be critical of Yann LeCun's expertise and experience. Firstly, the commenter refers to Yann LeCun as a "paperwork guy", which could suggest that they view him as someone who is primarily focused on theoretical research rather than practical application. Secondly, the comment implies that Yann LeCun has "zero practical tech industry experience", which suggests that the commenter believes that Yann LeCun's work has no relevance to real-world applications or that he lacks the necessary experience to understand practical issues in the tech industry. However, these statements do not accurately reflect Yann LeCun's background and contributions to the field of artificial intelligence. Yann LeCun is a renowned computer scientist who is recognized as one of the pioneers of deep learning, which has revolutionized the field of AI. He has a long list of practical contributions, including the development of convolutional neural networks (CNNs), which are widely used in computer vision applications, such as image recognition and self-driving cars. Therefore, while everyone is entitled to their opinion, this comment appears to be unfair and unsupported by facts when it comes to Yann LeCun's actual expertise and experience in the field of AI. [ChatGPT, March 14 version]
@@dovel9149 Yann's contributions to Instagram censorship are noted.
Yann has been training Neural Nets while you were in your mothers womb.