STEP Panel: How We Restructured Teams When AI Agents Joined the Workforce.

SPEAKER_1 (0:00)

So, the next panel, it's going to be called Operational Ownership of Agents. As AI agents move from pilots to production critical systems, the hardest question is no longer simply what they can do, it is who owns them. And joining this panel is my good friend Grace Gong. She is the founder of Smart Venture Media, an AI angel investor, best selling author, and host of the leading venture, host of one of the leading venture podcast. She also does a lot more, she's also a very humble person, so you have to ask her on that. And joining Grace is Matt Carbonara, investor at Mayfield, Dan Klein, a UC Berkeley AI professor, serial entrepreneur, and former technical fellow for AI at Microsoft. Suchintale, who is a Cedar Engineering Manager and architect at Rivian.

So please give it up for our panel, and yeah, let's give it up for them. Thanks.

Grace Gong (1:06)

Thank you so much for joining our panel here. So I want to do a quick little exercise. So I know that you guys just had lunch, and maybe some of you are about to having a food coma. So I really want to make sure everyone's like staying very active and having fun. Okay, so everyone, please stand up.

Yes, everyone. Amazing, y'all look amazing. Okay, so if you're a founder, please sit down. Founders, sit down. Investors, please sit down.

Okay. Operators, please sit down.

Who are the rest of you? Okay, amazing. You guys can all sit down right now. We'll figure that out later. Today, I'm on a panel with some of the three amazing personalities here, or actually amazing profile people here, including investor, founder, and operator. I want to start with a round of introduction. Dan is an amazing professor as well as a founder of a really cool company. Why don't we start with you?

Dan Klein (2:25)

Hi, everybody.

I'm Dan Klein. I'm a Professor of Computer Science at UC Berkeley. I am also the CTO co-founder at Scaled Cognition, which is a model lab. We build models that are high reliability and can't lie.

Sachin (2:43)

Hey, everyone. I'm such a 15-year-old student.

Vijay R. (2:56)

Thank you for having us.

So that's a new map of our Mayfield front, where a venture company will be based in the St. Valley, and we're investing in the C to Series B, all of the Valley infrastructure and AI stash applications to the software, or to the nuts and bolts of AI as well, like STEM and synopsis, We will listen to some of the key things and areas like that as well. But we're really excited to have them.

Grace Gong (3:25)

Amazing, I want to start with, since everybody's building their own AI agent these days, I feel like building an agent at your own company is very different than deploying it in the real world. Maybe, I want to start with Dan. Maybe if you can give us an example of how are you building agents, and I know you guys are also building models.

What are the biggest challenges that you're seeing for deploying the agent in the real world?

Vijay R. (3:53)

Yeah.

Dan Klein (3:54)

We talk about artificial intelligence or intelligence for that matter. It's one thing, but it's not. I think the biggest challenge in building agents, or really any enterprise today is reliability.

This really traces back to the core technology and the core architecture that people will build on. So if you look at things, what's the core issue?

SPEAKER_1 (4:17)

The core issue is reliability.

Dan Klein (4:18)

It doesn't have to do with other aspects of intelligence. So if you look at things like LXA, you look at the breadth of their knowledge or the processes, their behavior or the fluency of their output, it's through the roof. But there's other aspects of intelligence, controllability, verifiability, auditability, and slayability.

Vijay R. (4:36)

Here the technologies are just not working.

Dan Klein (4:38)

They're consistently lacking. So when you look simultaneously at the fact that you can whip up an eye-popping demo or a prototype seemingly overnight, but then it takes months and months and months to shift the thing, and there's a success that most enterprise AI projects don't actually shift. This is why the things you need for a prototype are these strong points in the model. But when it comes to the reliability, the model should just lie around practically.

I think that's really the biggest challenge is architecting for reliability in the AI space.

Grace Gong (5:16)

Sachin, you are at Rivian. I want to talk about, since we talked about the payment component, I feel like financial things is some of the most sensitive topics in the AI space, because you really need to make sure everything goes according to the plan.