The Mastermind Behind GPT-4 and the Future of AI | Ilya Sutskever

When he says we will find out very soon , it really does send chills to my spine!

Love when Ilya starts teaching everyone. Nice home page, too.

You can see there is so much going on in the amazing mind/brain of lya Sutskever. A historical interview.

takeaways: • [00:04] Introduction of the speaker, Craig Smith, and his guest, Ilya Sutskever, co-founder and chief scientist of OpenAI and primary mind behind GPT-3 and ChatGPT. • [01:00] Sutskever's background and interest in AI and consciousness. • [02:30] Sutskever's early start in machine learning and working with Jeff Hinton at the University of Toronto. • [03:45] Sutskever's realization about training large neural networks on big enough data sets to solve complicated tasks. • [06:33] The breakthroughs in convolutional neural networks and how they led to the imagenet competition. • [08:36] OpenAI's exploration of the idea that predicting the next thing is all you need for unsupervised learning. • [10:24] The development of GPT-3 and the importance of scaling in deep learning. • [11:42] The importance of scaling something specific in deep learning and the potential for discovering new twists on scaling. • At 13:01, the speaker discusses how scaling matters and that even small changes can have a big impact. • At 13:46, the speaker talks about the limitations of large language models, explaining that their knowledge is contained in the language they are trained on, and that they lack an underlying understanding of reality. • At 14:32, the speaker comments on the difficulty of talking about the limits of language models and how they change over time. • At 15:13, the speaker argues that learning statistical regularities is a big deal and can lead to a better understanding of the world. • At 16:33, the speaker talks about the limitations of language models and their propensity to hallucinate, but expresses hope that this issue can be addressed through reinforcement learning from human feedback. • At 17:52, the speaker discusses how teaching neural nets through interaction with humans can help improve their outputs and reduce hallucinations. • At 21:44, the speaker comments on Jana Kun's work on joint embedding predictive architectures, and expresses the belief that multimodal understanding is desirable, but not necessary for language models to learn about the world. • High dimensional vectors with uncertainty are a challenge for prediction, but Auto-regressive Transformers can handle them (26:02) • Auto-regressive Transformers work well on images (26:02) • Large language models learn compressed representations of the real world processes that produce data (29:40) • The goal is to make language models more reliable, controllable, and faster to learn from less data (33:44) • Learning more from less data is possible with creative ideas (35:51) • The cost of faster processors for training language models may be justified if the benefits outweigh the cost (37:48) • [25:28] The paper makes a claim that predicting high-dimensional distributions is a major challenge and requires a particular approach, but the current autoregressive transformers can already deal with this. • [26:02] Autoregressive transformers work perfectly on images and can generate images in a complicated and subtle way, with the help of supervised representation learning. • [27:09] The vector used to represent pixels is like a string of text, and turning everything into language is essentially what is happening. • [29:40] Large generative models learn compressed representations of the real-world processes that produce the data they are trained on, including knowledge about people, their thoughts, feelings, conditions, and interactions. • [31:31] Human teachers are needed to guide the reinforcement learning process of a pre-trained model to achieve a high level of reliability and desired behavior, but they also use AI assistance to increase their efficiency. • [35:10] It is possible to learn more from less data, and there is an opportunity to teach AI models skills that are missing and convey to them our desires and preferences more easily. • [39:57] In the future, it could be desirable to have some kind of democratic process where citizens provide information to neural nets about how they want things to be. • [41:15] It is probably impossible to understand everything in a complicated situation, even for AI systems, and there will always be a choice to focus on the most important variables.

Wonderful interview! Thank you Craig and Ilya!

Ilya talks so smoothly that i couldn't turn off the video literally

Thank you for uploading. I learned so much detailed nuances about LLM from this interview. I really like Ilya's way of communicating subtle but important points.

Thank you for all the hard work, everyone who do their best for these new systems to be implemented with the least possible disruption to human societies. We are still humans, and we must go from the perspective of love - to the future and beyond. Much gratitude. :hand-orange-covering-eyes::trophy-yellow-smiling::cat-orange-whistling:

That was really good - as someone with a general interest it’s one of the best video podcasts I’ve seen on this subject, and with a very central individual to the progress being made on AI. I liked the historical reflections at the beginning, it helped put things in context. I’ll be downloading the transcript to go through and will listen again. 10/10 👌

Thank you for sharing your insights and explanations Ilya.

- His intention was specific: to make a very small but real contribution to ai. ( in the time that people were certain computers can't learn, 2003!) - Auto regressive transformer is a very powerful tool that researchers underestimate. - "humans can be summerize in sequence", do you remember Devs miniserie!? - "To predict well, to summarize data well, you meed to understand more and more how the world that produced the data." - "maybe we are reaching a point where the language of psychology can be appropriate to understand these artificial neural networks!" - he doesn't believe these models don't have any real understanding of the nature of our world! - "human teachers are using ai assistance, and they are so efficient." By human teachers, he means people working on reinforcement learning from human feedback. - "make models more reliable, more controlable, make them learn faster, with less data and less instructions. Make them halucinate less. How far are they in the future? These are topics he intrested in and work on them right now!" The interesting thing is in OpenAI, he can't talk specifically about what he is working on, the open in opanAI annoy me a little! - "The costs are high, but the question is, does paying this cost actually generate something useful? Does what we get after paying the costs outweigh the costs?

Thank you for this interview and asking good questions and directing the conversation. Some good passages here to pause and really think about.

The subtle production of zooming and the downtime used in the intro is a good touch. Always good to show consideration for the audience instead of a ramshackle Facetime.

Awesome interview! Questions were great. Please more.

On the subject of hallucinations, I think they are more clearly explained by the problem space that the engine is trying to navigate. When having no relevant information on the subject, but it is still asked (one might say compelled) to say something, whatever it says must be either off-topic or false. And I believe Ilya is very insightful when he says the language of psychology is starting to describe these systems, because we have hallucinations, too. Whatever compels us to output something when indeed lacking skill or knowledge about a subject also affects GPT systems as well. When do people hallucinate or ramble? When they have no imposed limits/feedback, like a dictator or celebrity that is never told they are wrong or some guy living all alone in the wild or a child that has not been educated yet. Or a Twitter user. With social creatures it is the meaningful interaction with other social creatures (and the physical world) that generates these limits. Which I find promising and fascinating, because it means that the supervised learning step Ilya is talking about can also be performed by other AIs, not particularly humans. The brain is also composed of two hemispheres that keep each other in balance. Very interesting indeed.

Interesting and important discussion Craig and Ilya. Thank you Mike

Thank you for the great interview. One followup question I have for Llya is whether hallucinations stem from the compression or the output process. I suspect they are inherently encoded in the embeddings thus it is much harder to totally get rid of by just aligning the outputs.

Hi, thank you for this interview.  I have a tangential question for you: Would you happen to have any good leads on papers/researchers on the anticipated economic impacts of AI? I'm finding old stuff, but nothing new. Qualitative as well as quantitative forecasts would be really helpful. Thanks for any guidance you can provide.

Great upload! The future laws put in place as guard rails will be a huge player in the speed of AGI and possible adoption in some countries.

Ilya mentioned LLM learn color from text... I was wondering could it be that it learned from the color code in many HTML and CSS files? The RGB, hex code definitely have some structure that a text model can learn their relationships...