Stanford CS25: V4 I Jason Wei & Hyung Won Chung of OpenAI

April 11, 2024
Speakers: Jason Wei & Hyung Won Chung, OpenAI
Intuitions on Language Models (Jason)
Jason will talk about some basic intuitions on language models, inspired by manual examination of data. First, he will discuss how one can view next word prediction as massive multi-task learning. Then, he will discuss how this framing reconciles scaling laws with emergent individual tasks. Finally, he will talk about the more general implications of these learnings. Slides here: docs.google.com/presentation/...
Shaping the Future of AI from the History of Transformer (Hyung Won)
Hyung Won: AI is developing at such an overwhelming pace that it is hard to keep up. Instead of spending all our energy catching up with the latest development, I argue that we should study the change itself. First step is to identify and understand the driving force behind the change. For AI, it is the exponentially cheaper compute and associated scaling. I will provide a highly-opinionated view on the early history of Transformer architectures, focusing on what motivated each development and how each became less relevant with more compute. This analysis will help us connect the past and present in a unified perspective, which in turn makes it more manageable to project where the field is heading. Slides here: docs.google.com/presentation/...
About the speakers:
Jason Wei is an AI researcher based in San Francisco. He is currently working at OpenAI. He was previously a research scientist at Google Brain, where he popularized key ideas in large language models such as chain-of-thought prompting, instruction tuning, and emergent phenomena.
Hyung Won Chung is a research scientist at OpenAI ChatGPT team. He has worked on various aspects of Large Language Models: pre-training, instruction fine-tuning, reinforcement learning with human feedback, reasoning, multilinguality, parallelism strategies, etc. Some of the notable work includes scaling Flan paper (Flan-T5, Flan-PaLM) and T5X, the training framework used to train the PaLM language model. Before OpenAI, he was at Google Brain and before that he received a PhD from MIT.
More about the course can be found here: web.stanford.edu/class/cs25/
View the entire CS25 Transformers United playlist: • Stanford CS25 - Transf...

[@yoesemiat]

The fact that giving more freedom to the model and having less inductive biases affected by human subjectivity actually improves performance is really iluminating. Thanks.

[@jean-pierrecoffe6666]

Nothing new under the sun, this is just the Bitter Lesson

[@michaelbernaski7337]

Excellent. First talk is practical. Second is profound. Thank you.

[@sanesanyo]

One of my favourite talks in recent times..learnt so much from this.

[@ariG23498]

He has his slides in his head! Loved the content.

[@zyxbody]

I dont understand anything but I like how these people teach.May all get to understand the concepts thats my only prayer.

[@sady01]

What an amazing lecture. It was simple, yet groundbreaking

[@ricopags]

Really grateful for this being uploaded! Thank you to both speakers and to Stanford for the generosity. Highlight of the video for me is the Hyung's sheepish refusal to get into predictions on the staying power/relevance of MoE or any specific architecture. It felt like a wasted question since the premise of his talk is "tl;dr Sutton's Bitter Lesson"

[@inforoundup9826]

Great talks by both speakers

[@TrishanPanch]

Outstanding. I teach an AI class and there are loads of great pedagogical nuggets here that I am going to borrow.

[@ankitthawal1313]

Can you explain, what are those?

[@lugia8888]

Nice, a fake class.

[@irshviralvideo]

@@anshuraj4277 why bother going to college to learn ?

[@calm694]

@@anshuraj4277 learn english first before making going to AI CS

[@packsw9243]

@@calm694 "before making going" yeah you're a real genius

[@atdt01410x]

This lecture is super useful. really appreciate.

[@izumskee]

Very great talk. Thank you

[@adamlin120]

Great and inspiring talks

[@Aditya-ri7em]

he came and started teaching like a teacher .

[@laalbujhakkar]

Thanks for all the extra popping into the mic during the intro brrrruh!

[@indylawi5021]

Great lecture and insights on LLM.

[@jasonmeyer495]

Amazing content. His use of simple examples to explain deep concepts is extraordinary. So lucky to be living in a world where content like this is so easily discoverable and accessible.

[@JasonKendra]

Don't let this setback define your trading journey. Keep working hard and striving for success.

[@Faustordz]

Very intriguing!

[@MatijaGrcic]

Amazing!

[@CrazyFoxMovies]

Great lecture!

[@lugia8888]

All of this is BS 😂

[@itsaugbog]

Hilariously Jensen Huang from NVIDIA just spoke in an fireside chat recently about how they're already dependent on AI and models for designing chips so that last comment is already happening. Great talk.

[@simonfranco644]

Can you support this with a doc or link. I am very keen in exploring this. Also, it was hilarious to me that the attendees laughed at the doctor for explaining that and I giggled when he mentioned that it might just be official in two years or so.

[@jooholee_]

Greak Talk. Very Inspiring.

[@zacharykosove9048]

The students were asking some great questions, no wonder I don't go to Stanford

[@roro5179]

im the dude at the end (dont go to Stanford xd)

[@mprone]

Questions looked pretty naive to me. What's "great" about them to you?

[@Lalala_1701]

Andrew ng also took same kind of example to explain LM.

[@gmccreight2]

Thanks for the talk! Really interesting stuff. I had one question. At 1:04:00 Hyung suggests that uni-directional attention is preferable to bidirectional attention in turn-taking scenarios because it allows the reuse of calculated information in the KV cache. I'm trying to understand how this fits into his broader thesis that we should be moving towards more generic approaches. On the surface the use of the KV cache doesn't feel particularly generic. Does it make sense because masked self-attention is necessary for next token generation, anyhow, so using a causal attention mask universally makes sense?

[@doinitlive3015]

Types of leadership can be used as an analogy in the area of using less structure but at the same time performance is higher. A leader who utilizes an authoritarian type of leadership increases productivity within the team but decreases the team's creativity. Whereas a team under a democratic type of leadership are able to solve problems with increased creativity leading to innovative ideas.

[@heyitsjoshd]

How do we know what is small vs large? For example, with emergent tasks, it highlights that more data could lead to more accuracy with enough compute. The small LM would have not seen accuracy improvements but the large LM did. For the tasks currently indicated as flat, couldn't we just not have enough compute now to know if these tasks would get more accurate?

[@DanBillings]

Please put the subject of the talk in the title. You can then market the OpenAI speakers

[@lc.sin.]

Besides compute, I guess the eponentially cheaper network bandwidth, data storage, sensors to capture real world input should also be part of driving forces

[@dkierans]

Yeah, this is a pretty great talk. It is quite hard to figure out at what technical level to hit the widest audience. This is nice. Not as nice as those flaxen locks though.

[@aliwaheed906]

Maybe the emergent behavior happens because for that task to be learned there are a set of pre-requisite tasks that need to be learned first. Just brainstorming here.

[@dodowoh3683]

Surprised by the amount of hair an AI scholar may have retained.

[@erebi8386]

형원게이 힘내라

[@Arcticwhir]

im more curious about the 22% of completely flat set of tasks and what the solutions are to change that. Also for larger models, showing that less structure is generally better but needs more compute, does that mean the model will need less RLHF to have a desirable model for humans...

[@elcanmhmmdli3305]

Azerbaijan❤

[@akashdeb9823]

Jason can do 18 pull ups no breaks

[@robertwilsoniii2048]

Something that always bothered me was that adding in random terms increases predicability power, holding sample size constant (scaling compute without increasing data size). The peoblem is it decreases explanatory power and ability to understand the individual contributions of each variable. It's like pop-astrology, star signs -- libra, gemini, leo... etc. -- adding extra variables improves scaling compute and predictability, but does it add anything to clarity? I suppose to make predictions clarity doesn't matter. That always annoyed me.

[@hedu5303]

Strange world. This dude is almost a kid and gives a lecture

I am happy to learn from any kid :)

[@chaidaro]

His intuition is older than me

[@vireyes1595]

nah man gotta recognize game when you see it. dude’s a future titan of the industry and we’re out here getting his guest lecture for free. pretty solid win for all parties involved in my book

[@SuperHeromindNsoul]

True we can all learn from each other and Speakers here also learn from someone

[@MrAmgadHasan]

Indeed. Many of the recent breakthroughs ML were achieved by people in their 20s, mostly during or briefly after their PhDs.

[@hh0686]

Please…why can’t the presentation be done on a projector instead of a whiteboard. The kind of visual is so horrible.

[@Umarbit]

Please remove the noise from audio

[@wenhanzhou5826]

Dude just learned how to manually classify lungcancer to better understand the neural network he is building 💀

[@rasen84]

The second half is 100% wrong on the idea that scaling is what matters and adding complexity into the model, adding inductive biases bites you in the ass later. You're not considering the considerable amount of human labor allocated to data curation and handwritten instruction tuning data. That is necessary because the model is too simple and too dumb. The model doesn't have the necessary inductive biases to intelligently take any data. You need to add more inductive biases in order to obviate the need for human labor on data curation and creation.

[@user-se3zz1pn7m]

He is not talking about the immediate moment. He is discussing what kind of model would be preferable when there is an abundance of data and computing resources. He mentioned that due to the current limitations in computing resources, it's necessary to use models with some degree of inductive bias. Although he didn't say it explicitly, he probably thinks that models with inductive bias are also needed due to limitations in data. However, in the future, as more computing and data resources become available, models with less inductive bias will be better.

[@rasen84]

@@user-se3zz1pn7m what I’m saying is that the data collection, creation and curation process should count towards model complexity and scaling hypothesis. You could be removing complexity from the model and offloading that complexity to human data curators and creators.

[@user-se3zz1pn7m]

​ @rasen84 , I believe we are on the same page. I agree with your point that "You could be removing complexity from the model and offloading that complexity to human data curators and creators." However, I think he is talking about the trends and the distant future, perhaps 10 years from now. Yes, if we remove complexity from the model and training methods, we will need more resources to compensate for the trade-off in data preparation. However, in the future, there may be a vast array of open-source data available and synthetic data generated through self-play approaches. Then, our goal will be to reduce assumptions in the model, give it more freedom and make it bigger . I believe this is what he intended.

[@hang_8169]

@@rasen84 I would argue even if you use old method which has more structure in it, you still need spend the same amount of effort on data if not more to be adhere to the structure that you impose on the model. Because your model has MORE assumptions on data that it expects not less.

[@rasen84]

@@hang_8169 then it’s time to add more inductive biases.