Data Council 2025: The Foundation Models Track with Dr. Bryan Bischof and Tom Drummond

• 

  Dr. Bryan BischofHead of AI
  Theory Ventures
• 

  Tom DrummondManaging Director
  Heavybit

Heavybit is thrilled to be sponsoring Data Council 2025, and we invite you to join us in Oakland from Apr 22-24 to experience 3 days of cutting-edge technical talks from the brightest minds in AI & data, learn more at datacouncil.ai and use the code HEAVYBIT20 for 20% off.

Designing the Foundation Models Track

Tom Drummond: Hi everyone, my name is Tom Drummond.

I'm a co-founder and managing partner at Heavybit. We are one of the sponsors of this year's Data Council.

We're very excited to be involved and as part of that, we are talking to a handful of our track hosts to get a read on who they are, what they're putting on for everybody and give everyone a little preview of what they can expect to hear at this year's event in a couple weeks.

Today, I am talking to Dr. Bryan Bischof. You are the track host for the Foundation Models track, correct?

Dr. Bryan Bischof: Yes, that is correct.

Tom: I have to say, having looked at all the track host backgrounds, you have by far the most interesting background out of everybody.

Bryan: Interesting. Is it the Blue Bottle? Is that what got your attention?

Tom: It's one of the eclecticism. You've got a little blue bottle, a little fashion in there. You've got kind of all different aspects.

I feel like I should do a talk show host intro, ladies and gentlemen, it's Dr. Bryan Bischof, educator, scientist, author, world renowned, something or other.

Bryan: Yeah, I think that last one is probably the one I identify with the most. " Something or other."

Actually, I might just go ahead add that to my LinkedIn. Something or other.

Yeah, I've been really lucky. I've gotten to work at a lot of really exciting organizations, doing super cool stuff.

I've been very privileged to get to sort of apply mathematics to a bunch of real-world problems, not just in one space.

You know, I started out in time series stuff, making computers talk to one another, got to work on coffee stuff, which is a big passion of mine.

And then later, I got to work on logistics for clothes and you know, these days, working on sort of data stuff.

Tom: That's awesome. And so tell us, where are you today?

Bryan: At the time of this interview, I am a hedge knight, but in only a few days, I will be joining Theory Ventures as the head of AI.

Tom: Congrats. Yeah, that's awesome. Tomasz is great. So let's talk about your track, so Foundation Models. What did you want to achieve putting this track together?

Bryan: Yeah, so when Pete asked me to put together the Foundation Models track, one thing that first crossed my mind is cool, I'm going to have people talk about like training large foundation models and that lasted in my brain for about 15 seconds and then I realized no, that's horribly boring.

Most of the people that are coming to Data Council just won't find that content that exciting. Training large, like massive models, even if it's, post training or mid training, I just don't think that it's going to be that exciting for that audience.

Instead I put myself in the shoes of an attendee because I am almost like a perfect example of who would want to come to Data Council and so what would I want to see?

And so what I thought about is, okay, candidly, I often find myself wondering, what do the labs want me to know?

What do they want me to think and how do they want me to be using their models?

I've been really lucky. I've gotten to talk to folks at Anthropic and Google and OpenAI.

I've gotten to sort of like work directly with one of the speakers in my track, Nikunj, and I've really benefit from those conversations. I've learned a lot from talking to the API teams about how they think about their models and they think about the application layer.

So my sort of inspiration for the people that I invited is actually the exact opposite of what the track is called.

So my track is called Foundation Models, but my track is actually about the application layer and it is about what you should know about the foundation models if you're building at the application layer.

Tom: Awesome, okay, so tell everyone, if they come to your track, who are they going to be hearing from?

Featured Speakers & Topics

Bryan: Let's start at the top. We're going to talk about whether the model is the product or not. We're going to have Han (Han-chung Lee) from Moody's, he's going to talk about the fact that the model is the product and that is really where the value is being created.

Hamel (Hamel Husain) is going to argue against that point that the value is created outside of the model itself at the application layer and that's going to be sort of the aperitif for the rest of the track because then what we're going to get is we're going to get a parade of model builders who are very sort of both sided on this argument.

They're building the models but they're simultaneously going to be trying to explain to you how to use what they're building to create value. And this is obviously intentional.

My goal here is to help you think critically about where you fit into these things.

The next speakers are going to be from the major labs. So as an example, we're going to have Nikunj (Nikunj Honda) from OpenAI, he's going to be talking about the agent's API. This is a really big deal, and everyone needs to know about it.

Next going to hear from Ragho (Raghotham Murthy) from Meta. Now Ragho is going to be telling you what does the Llama stack look like. If you care about building local AI applications, you need to understand the Llama stack.

Moving on, we have Ravin (Ravin Kumar) from Google DeepMind. Ravin is going to be talking about LLMs as tools.

If you really believe that the value is created outside of the model, then you need to understand where the model fits in. I'm really excited to see Ravin's perspective on this.

And finally, I have to say the word multimodal or people just won't come, so we obviously wanted to have somebody from the really multimodal perspective and so we have Ethan (Ethan Rosenthal) from Runway.

Tom: That is a killer lineup. What's a good elevator pitch for the series? Like why should you come to the Foundation Models versus like one of the other six tracks that are going on?

Why Attend This Track

Bryan: Candidly, because, like, we all need to have a really clear understanding of our positioning relative to AI right now.

Whether you want to go and build a foundation model to drive new applications and capabilities or if you need to understand how to use the state-of-the-art capabilities in your application, candidly, like this is the right information from the horse's mouth.

You'll notice like those are some big brands and I'm not someone who's like super obsessed with brand, but I think in this case, when you've got a few players that are really driving the conversation and the progress, you need to hear their perspective, and this is going to be the most rich way to get that information.

Sure, you should go look at the API Docs, but I think this is going to be a real live in-your-face demonstration of what these positions are.

The one thing that you'll probably notice is missing from this track in a very obvious way is Anthropic. Where is Anthropic?

Let's say the word MCP into the microphone so that everyone is very excited and what I'll say to you there is the MCP talk is already on the internet and you can go find it today.

It was recently given at the AI Engineering Summit in New York. So go ahead watch that talk. Thanks to Swyx for organizing that talk.

That's where you can get the information on MCP. You can think of this track as the version of that MCP talk for all the other model providers and all the other pieces of the story.

Tom: So you're going to be touching a bunch on kind of agents and tool use and things like that?

Bryan: Absolutely, there'll be a lot of discussion of sort of like, how do you have agents in the loop, how do you think about, like, what are tools and how do you make your tools really valuable?

Bryan’s Own Talk: “Failure is a Funnel”

Tom: That's awesome. I actually noticed, I'm kind of curious if this is like connected, but you're actually giving your own talk on day one of the conference, right?

Failure is a Funnel in the Data Science and Algo's track, I think.

I feel like today not a lot of people apply kind of classic engineering rigor to building these applications.

Like is that kind of what you're going to be talking about in the Failure as a Funnel or is that different?

Bryan: Yeah, I wanted to really talk about something that I feel like I'm an expert in. So I obviously chose failure.

And so my thesis is that basically building applications using AI is much more like classical product development than people think it is.

And I don't mean in the sense of like the engineering road mapping, the engineering frameworks, engineering process, I actually mean product analytics.

And so where I think a lot of engineering teams, especially AI engineering teams are coming up short is the way that they integrate data science skills, data science analysis and thinking of these processes the way that we would traditionally think of sort of optimizing your product.

It sounds really banal to say whoa, whoa, whoa, take a break. Look at where people were dropping off and the use of your tool. But actually, it's surprising how rare people are making this connection.

Tom: I feel like most people are sort of stuck on, hey, I should just like reword my prompt a couple times or I should like change a couple things here or there, and it's trial by error.

Bryan: It's even deeper than that. You hear a lot of discussion these days about evaluation and I'm really excited about evaluation.

I've been pushing evaluation for the past couple years, but I think people are very limited in scope when they think about what that means.

They think of evaluation as "did it work, did it get the answer right?" But actually, evaluation is a multi-step procedure and one step at a time, you get closer and closer to the right outcome.

I'm going to do something very unusual in my talk. I can promise, I can't reveal the secret, but what I can promise is if you come to my talk, you'll see something that you've never seen before at a tech conference.

Tom: That is reason enough to be at Bryan's talk, 11:45, on day one, April 22nd, Oakland, Datacouncil.ai, come and get your tickets.

Use HEAVYBIT20 for a 20% discount. Look forward to seeing you all there.