Ep 44: Co-Founder of Together.AI Percy Liang on What’s Next in Research, Reaction to o1 and How AI will Change Simulation

"Liang reflects on the advancements in AI saying, '...this...signals sort of a change that I think we'll see going into the future,' hinting at a transformative phase in AI research and development."

"In discussing the impact of language models, Percy Liang stated, '...we think about language models is prompt comes in, response comes out...people measure...tokens per second,' highlighting the traditional view but also hinting at potential shifts in understanding AI responses."

"So it should be the case that whatever research project you pick, if GPT-4 or GPT-5 comes out, it should enhance, either enhance your research or be like kind of irrelevant in some sense."

"Again, another example, we have a benchmark on, you know, ability for a language model to solve machine learning engineering tasks. And that's going to get better."

"But I think there's a limited amount of things you can do with the model. And obviously some of those measures can be taken by, you know, the model providers themselves or net new startups."

"This is interesting because even though the math doesn't apply to the actual model, it was a source of inspiration for the actual architecture."

"I think the latter is maybe less well studied, but I think it has a whole host of new kind of potential applications that we're not really looking at right now. Even for clinical trials right now in medicine, they're trying to do this digital twin type work where it's like you give someone the treatment but then kind of simulate what would have happened to them if they'd stayed in control."

"So I think there's two types of agents... The first agent is more what we talked about earlier, agents like O1 that are able to perform really difficult tasks. And the other type of agent is more like about simulation."

"I think there's a lot of potential applications or benefits from being able to do, you know, accurate simulations. But I will say that, you know, we have to be very careful that, you know, the simulation currently is still, you know, quite far."

"I think we are closer to a kind of BERT era moment for robotics where the idea of using a vision language model, fine-tuning them and getting four foundation models for robotics is effective."

"We don't want to just have unconditional generation or even just textual generation and having it just generate the music. Felt like it wasn't giving, you know, artists enough control over like what it was happening."

"I know what I want, but my fingers won't do it. At a personal level, I'd be really excited about how can these tools really help me realize my musical vision."

"You know, there's a way to communicate that to AI that enables you to still express that creativity within you?"

"It's like we're getting to a point of model capabilities where we may just need models to write evaluations for models, exploring new forms of evaluation."

What is "Ep 44: Co-Founder of Together.AI Percy Liang on What’s Next in Research, Reaction to o1 and How AI will Change Simulation" about?

In this episode, Percy Liang, a prominent figure in AI research and co-founder of Together.AI, discusses the evolution of artificial intelligence, from simple token predictions to sophisticated autonomous agents. The conversation covers the significant challenges in AI, particularly regarding evaluation and interpretability, while also reflecting on emerging trends and the potential of new models like OpenAI's O1. Liang highlights the transformative possibilities of AI applications in various fields, including robotics and music, and emphasizes the importance of holistic perspectives on AI technology beyond just language models. The podcast concludes with speculation on the future of AI in creative fields and education, revealing a nuanced view on collaboration between humans and technology.

What are the key takeaways from "Ep 44: Co-Founder of Together.AI Percy Liang on What’s Next in Research, Reaction to o1 and How AI will Change Simulation"?

Percy Liang emphasizes the importance of generative agents in understanding complex social dynamics. Interpretability and evaluation challenges in AI remain critical topics in ongoing research. Recent AI models like OpenAI's O1 may have broad implications for developers and researchers.

Which podcast is this episode from?

"Ep 44: Co-Founder of Together.AI Percy Liang on What’s Next in Research, Reaction to o1 and How AI will Change Simulation" is an episode from Unsupervised Learning.