Foundation models and the next era of AI
Вставка
- Опубліковано 28 тра 2024
- AI advancements are outpacing expectations and presenting novel opportunities and research challenges. In this video, Microsoft Senior Principal Research Manager Ahmed H. Awadallah explores AI’s shift from perception tasks to text, code, image, and video generation and the class of large-scale models underpinning recent progress. He also examines how these models are creating new experiences and enhancing products people use today to support them in achieving more.
0:00 Introduction to foundation models
11:04 From GPT-3 to ChatGPT - a jump in generative capabilities
19:11 Everyday impact: Integrating foundation models and products
Check out the video recap, full transcript, and additional resources:
www.microsoft.com/en-us/resea... - Наука та технологія
This was an epic walkthrough!
Thank you for taking the time to explain important developments in AI! The emergence of these new abilities is amazing!
Proud Egyptian here. Let’s go! 🎉
That was extremely informative and insightful. I do say it could use more excitement to grasp the audience attention to keep them focused. Overall great entail.
So informative ,thanks for sharing
This is a awesome talk and sincerely hope there would be more on other key topics as well........
Thanks so much! This talk might be the best for explaining why ChatGPT performed beyond expectations!
well done!, good work, keep posting.
Outstanding. Thank you so much.
Keep the great work 🎉
Great work.
This is a great video with a lot of new ideas for those who start diving deeper into LLM. But there is one example at the end of the video about Super Bowl. I think it may be not easily understandable to all the people because it is a specific event in the U.S. about which a lot of people are not aware. It is better to choose a more general example. I hope this helps.
Very interesting video!
Excellent overview
Excellent!
Awesome !
'We have been getting to a level of quality, we have never seen before' was perhaps a bit superfluous unless we are going the wrong way ;)
An u😃nusually concise and thoughtful analysis
Awesome.
3ash!
This is alot to sink in , i need my ChatGPT to do a lot of explanation and alot of 5W/1H lol ..
great content overall , thanks :)
ok but why putting ads almost like an hour
I ran into this video as an ad too. I was attending to something else and the video played too long for an ad so I decided to search it up and here I am. Not many viewers would want to see a long ad except people who are curious like me.
7.50 the answer on Apples could be wrong or right "20 to make lunch" does NOT mean that those Apples are `gone` maybe the meals have not sold yet and thus they are still in the cafeteria, were as a human would presume that the question by its nature means they have been used and should not be counted, that is not explicitly stated nor is it the only interpretation of the words used. Pushing models to get the `right` answer in such cases could reduce abstract thought and non linier thinking, its `arithmetic ability is connected to its other abilities as they are emergent you need to be careful around `correct` as it could restrain wider interpretation, that flexibility in humans is very powerful.
The models are not generating from scratch, they are being trained with videos and pictures and end it up being a realistic interpolation of many samples. The issues is the copyrights of training data. Even if the generated does not resemble a percentage of the original, it is still being used for the training.
The output of many of these models are absolutely NOT “interpolations of many samples.” I see so many people speaking so confidently about AI these days who clearly have no idea what they’re talking about. AI is the hot topic right now, so of course everyone’s an “expert” all of a sudden. These models do not interpolate. You don’t know what you’re talking about. Please stop polluting the information space.
@@therainman7777 Generative ML model tries to learn manifolds. The learning algorithm assumes the data points are distributed on the manifolds. Please tell me how it is not an interpolation.
@@devoch2031 For one thing, the “manifold hypothesis” that you’re referring to is just that: a hypothesis. The truth is that no one actually knows yet how a model the size of, say, GPT-4 actually works. Its own creators admit that openly. And secondly, and more importantly, the fact that many of these models contain a lower-dimensional representation in some of their hidden layers (the “manifold” that you’re referring to) does not make their outputs “interpolations” in any meaningful sense of that word. The _output_ of these models, which is what we’re talking about, is of identical dimensionality to the input; it’s an N-dimensional vector where N is the number of distinct tokens. Ultimately what an LLM is doing is approximating a probability distribution over sequences of tokens. If you had a set S of samples drawn from a one-dimensional, normally distributed random variable X, and used a model to approximate the mean and sd of that normal distribution, you could then sample new data points from the new approximated distribution, X hat. However, those new data points would _not_ be interpolations of points in X. They would be new samples distributed according to X hat. “Interpolation” is a specific term with a specific meaning, and it absolutely does not apply to the output of generative models.
@@devoch2031 It is an interpolation, it is also a reductionist attempt at downplaying the subjective value or assessment of the output. Much of human education and learning is the exact same thing at the end of the day.
I understand a bit because I am a computer student.
well done, cs.
Now, how many bits can you understand when they're together?
1:38 Ehhh..... NPR.... I used to listen to them but they lost their way. Anyways, how about some feedback on putting a pause on research so that we can align the AI that is now possessing superhuman intelligence.
Idiotic to “pause AI” unless you can magically guarantee that every country’s government and every company and every person will pause, which is impossible. Pause in The West, and it means China gets to advance.
LNNs are the future. Large Neural Networks for Architectural Quantum AI.
The person that made the video editing choices needs to be fired: Microsoft has the capacity to do better than a small portion of the screen for the stuff we as viewers care about. It makes it impossible to read on anything but a large screen, which is largely wasted.
55% more productive for developers! 55% more $$$ for developers??
Call me hype driven if you want.
Any organization not exploring the implementation of AI is wasting time and money they can't even imagine yet.
There will be no real "pause".
AutoGPT and other implementations start showing how absurdly effective this soon will be.
Nobody is watching this .. spend on CSR instead of UA-cam bots 😑
Very little new information. Mostly marketing. I think a code tutorial would have been more helpful.
You are the wrong target audience for this specific video. I found it very helpful as a newbie to understand the basic concepts.
I don't understand he's talking. Je comprends pas quand qui parle.
Confused whether this was educational or the guy is selling stuff