Research Forum: Panel Discussion: AI Frontiers
ฝัง
- เผยแพร่เมื่อ 30 ม.ค. 2024
- Microsoft Research Forum, January 30, 2024
Hosted by Ashley Llorens, VP and Distinguished Scientist, Microsoft AI researchers, Sébastien Bubeck, Ahmed Awadallah, and Ece Kamar discuss frontiers in small language models and where AI research and capabilities are headed next.
See more at aka.ms/ResearchForum-Jan2024 - วิทยาศาสตร์และเทคโนโลยี
Join us for a continuous exchange of ideas about research in the era of general AI.
*Register for the Microsoft Research Forum series:* aka.ms/ForumRegYT
i liked this video because of bubeck's first answer, where he calls his aspiration of AI is to illuminate human mind.
Great video and very insightful.
Congratulations AJ, Sébastien and Ahmed on your new roles in the AI Frontiers team!
We look forward to all the amazing work now and into the future, All the best on your mission :)
Just like how it isn't the most effective if one human tried to learn and do everything in the company, so are more specialized AI models more successful at their perspective tasks. I could see those general jack-of-all-trade models act as a sort of project managers / leaders / communicators between more specialized models.
Edit:
Ohh, that's kinda what she's talking about at around ~18:00
that was insightful, thanks guys.. now to bed! 🌞
Thank you for sharing how you see and how you work on AI research. If you are interested to outsource some of the work, it will help on making the machine to learn from different kinds of intelligence that it should learn better facilitation of human learning. Perhaps it can also fast-track empathy development and to be more context sensitive. We think that we must have agents that learns from the innocence of humans who specializes on what are made available to them by nature or by circumstances. Perhaps delusions can be transformed into instinctive learning.
❤️❤️❤️❤️❤️
Paraphrase: "we dont know the mysteries of the human mind" - how it works. Then the next question is how good of an analogy are transformers to the human mind - "terrible".
Is it me or does that make no sense at all: how do you know if something is a good analogy of something else if you dont what the something else is?
For all we know right now the human mind is just a next token predictor - we have no other model that comes close to reproducing the output of the human mind.
Next token prediction is indeed the similarity, in this case its a very good analogy. The real mysteries I assume is meant to be one level lower, and then the analogy isn't practical.
Prediction is mostly what vision is based on, with minimal input from the eyes and then just a bit of error correctness. So why is reasoning not just a mirror of that system?
I assume the distinction is in outcomes and performance. Allow me to illustrate.
Humans don't need the same quantity of examples to learn patterns. Also, the brain uses much less energy to do what it does than an LLM. Third, the human brain can reason when its exposed to a phenomenon its never seen before. Next token prediction is based on statistics abstracted from a large body of examples. LLMs can't next token predict if the sample size is 1.
@@darylallen2485No, that’s not correct. Modern LLMs are quite capable at reasoning about contexts or scenarios they’ve never seen before. This is called zero-shot learning. They are also capable of learning patterns from a very small number of samples; this is called few-shot learning. GPT-4, for example, does both zero-shot and few-shot learning extremely well. I’m not sure why you thought they weren’t capable of this, but you’re very wrong.
Also, while it is true that the average person’s brain is capable of learning new patterns with relatively few samples, this does not happen in a vacuum; by the time a human’s brain is capable of doing this that brain has already processed many orders of magnitude more information than modern LLMs are trained on. So in that sense the brain is not actually more efficient as you claim.
@@therainman7777 Yes I can agree with that, however it can still be useful to highlight the limitations.
For example, a LLM today does not in fact receive input that it can use to define the degree of which the predictions were successful in real time. This operation is very metabolically expensive so the brain have optimized its behavior to reduce the amounts of this function invocation, but it is critical to reinforcing the predictions of the model. This is where the constraint of the environment, makes the brain more efficient and adaptable.
Potentially how it can be solved is to have RAG serve as an actual input that corresponds to what in reasoning is the existing knowledge base, and in vision is light rays. What is missing from there is a reinforcement learning model applied on the predictions with the goal function being derived from the evaluation of the RAG invocation.
Your video description says Ece Kamar instead of AJ Kumar 😂. That’s not very good AI😀
It is a very good AI, accuracy is not everything, efficiency is needed here
@@joannot6706 Awesome! My AI broker will buy me 100 shares of Macrosoft when I ask it buy me Microsoft. But I'll remember and be grateful that it was super efficient doing this.
Hello to all at Microsoft i would like to have in my PC almost all Microsoft Research topics such as Cloud Computing( in Health Care and Simulations VR + AR) because i am a Windows Insider i would like to have too a enhanced support in my PC and related subjects i know that all Research Topics in a single PC isnt appropriate so i would like to have special support in all my previous and future comments here on TH-cam because i am going to spend almost all my time from now on here on this channel and others related with Microsoft Windows the PC.
how would someone contribute to your research ? i not only discovered this before it was published but also have done extensive research what sort of reasoning works for what walks of life. P.S. discovered next bit in my research better than textbooks
I wish the interviewer was someone with actual research knowledge… At least then they wouldn’t sound like they’re pretending to understand…
You mean woke sense reasoning?
IMO Microsoft needs to integrate AI into Active Directory and their server products like Exchange Server, SQL Server, Project Server, SharePoint, etc. We need that foundation in order to create digital humans that will perform tasks that humans do. To some degree, people have tried to perform some of these tasks with AutoGen and TaskWeaver, but we need that built into the foundation so that businesses can leverage that when they build out their AI foundation for their business.
When the corporate HR shill spouts shallow catch-phrases for twenty minutes and you're like, "Let the researchers talk already..."
She's a PhD CS. And a researcher at MS for a very long time.
@@silvesterjkennedy Anyone can pass exams and write papers. But maybe I missed something. What did she say that impressed you / interested you?
my god we get it you work at microsoft that woman keep telling it non stop
I really hope I will receive some technical insight but this is just waster of time
Sebastian yt video is billion times better than this this,
title should be AI Frontiers MSR AD
Probably im not the public for this. Zzzzz
vomit if it from MS - run