David Friedberg on AI-First Startups & the Future of Biology, Business & Creativity | AI Basics

Yeah, I mean, we had a couple projects that we shut down. We have a few that are still going. And then Ohalo is just a very special company that had these very important breakthroughs where we changed how plants can be used for plant breeding using CRISPR-type tools. So that's created a new system for plant breeding, unlocks tremendous yield potential, which is the key driver for agriculture, and unlocks the ability to make seed in crops where you can't make seed. And so the general kind of storyline is this becomes a huge game-changer for the entirety of agriculture. And so it was so important where we had these breakthroughs. I said, we can't mess this up. I'm really committed to making sure this company succeeds. I have to make sure it reaches its potential. So I stepped in in November of '23 as the full-time CEO.

Well, it's changing every week. As anyone who's building a business today can tell you, it is every job, every function, every workflow can be improved, levered, scaled up using AI. And so, like, I'll give you an example. My chief of staff, Laura, who you know, she's great. We posted up a job req for like a partnership manager for supply chain role. So I put it on LinkedIn. I have a lot of followers on LinkedIn now. Get a bunch of resumes come in. So hundreds of resumes come in. It's like, how do we filter through all these resumes? So Laura asked AI to write a Python script for her to download all the resumes from Greenhouse, scan them, score them, and give a result. She did this all in a couple of hours. So she's not an engineer. She used AI as a tool to create a tool to give herself an incredible amount of leverage and very quickly screening through these resumes, prioritizing them. That's a feature that should exist in one of these apps. For some reason, it doesn't. And she was able to cobble it together in literally hours and get leverage out of it. It was a great example.

We're creating a new system of plant breeding, and we wanted to do some experiments using controlled environments and LED lighting. And so we had AI, rather than have a project manager put together the whole project plan, which would take weeks or months, we had AI put it together by feeding the AI all these research papers, giving it some guidance. And I did it all myself within a few hours. I got this output, and here's the project plan. And now the team is executing on it. There are so many examples of that where what would otherwise or historically have been large kind of project or human-intensive workflows, you can simply create the script or create a plan using AI in minutes. And so just the people that know how to use the tools are just getting incredible leverage, and that gives you, you know, faster turnaround. You can do more with more nimble teams. It's just an incredible set of tools.

That's right.

Yeah. Yeah. Yeah, I mean, when we sold Climate, we were 160 employees, and then we grew to... I had hundreds of people. I think I ended up with, I don't remember, over a thousand people in my org after we were acquired. But it's very different. So a lot of people assume you just have small team sizes, and I think that that's under the assumption that the companies are doing the same things and look the same as they did a decade or two ago, which is like making an app, a Web 2.0 app was a company back then, and you needed engineers and front end and back end and DevOps and QA and all sorts of stuff. But today there's the opportunity to build deep tech companies, to build technically difficult companies, to build businesses that historically may have been the realm of thousand-plus-person companies with 100 to 200 people. And I think that that's really the key thing that I focus on. Everyone always talks about, oh, you can have a whole startup for two and raise $2 million and get to profitability. It's so different. VCs are dead. I don't think that's true. I actually think the addressable economic opportunity for startups has actually grown. So startups that otherwise might not have been able with 50 or 100 people to go after a multi-billion-dollar seed market, which is kind of the market that we're in, can now do so because of the leverage they get with AI. And that includes software and automation, automation in the lab for us. We're doing more automation in the greenhouse, better decision-making because of the AI. So there's a tremendous amount of leverage and efficiency that allows smaller, nimbler teams relative to big projects to address them. Venture capital money is not going to like suddenly dry up and shrink down because everyone's thinking, oh, everything's a web app. It's not true. I actually think we can do a lot more with startups out of Silicon Valley or out of any small company that knows how to use AI and that knows how to leverage these tools.

I'll give you another set of examples. Like if I were to say design an automated drilling machine that can get me into the earth to go find rare earth minerals. And you know, the fact is we have a need for rare earth minerals, right? So that's a problem that society faces, that civilization faces. We got to get them out of the ground. So we know of a few mines, but it turns out we know one millionth of 1% of what's in the Earth's crust, where it is. We got to go get it, right? And by the way, everything exists everywhere. Meaning I can go get all these rare earth minerals. They're not just in like Inner Mongolia. The reason we mine them in Inner Mongolia today is because they're at the surface level at some particular location. We only look at the surface in mining. If we went into the crust, all of the stuff we need, all of these rare earth minerals exist in North America. We just got to go deep enough, and we don't have the technology to do that. The energy cost is a square of the depth you go when you do mining. So the deeper you go, the cost is squared.

Yeah. 200 ft is the square of the cost of going 100 ft. And it's very hot. Then you need new material science to do it. And then how are you going to get all that dirt? Where's that dirt going to go? And then how are you going to refine the ore? Every one of those steps of that deeply complicated, highly technical project can now be run in silicon. And I can do generative design on a machine. I can do generative design on the project plan. I can basically take what would have otherwise been a thousand-plus-person analyst, engineer, research program type effort and reduce it down to software written by software. And that's what I mean when I say like we can address tougher problems than we historically think about addressing. And I think we're kind of at the surface level today of thinking about the problems that we can address. Silicon Valley is still adjusting to this. We're still thinking like, oh, let's make a web app in a minute. Okay, cool. Good for you. Everyone's doing that. We're going to get sick and tired and bored of that. What's going to happen is everyone's going to wake up pretty soon and they're going to be like, holy, we can do so much more. I can design a project plan to get me to the moon in five years for less than a billion dollars. I can design a project plan to get to Mars and ship an entire city to Mars in less than 15 years for less than $50 billion. So that's the sort of like visioning that I think we're just beginning to do as we're realizing the potential of these tools. And that's a limitation of entrepreneurship and creativity and just like you said, the problems that maybe... imagination.

Right. Right. It's really weird to catch up that quick.

Eric is my... I don't... I think you've maybe met Eric Andreo. He's my VP of engineering, I guess we call him, or AI. He's been with me forever. Worked at Climate Corp, ran data science at Climate Corp, built all our models, led the teams that did that. So he's working with me at Ohalo, and he's done a couple of these kind of lunchtime seminars where we took... we actually did one where we then followed it with a hackathon. So every employee at the company came and we did a sit-down, taught them how to use Cursor, and then said, pick a business problem that you're dealing with at work, whether you work in the lab or whether you work in the greenhouse or whether you work at a desk, and come up with an app, a web app, and then you're going to create and deploy that web app by tonight. And then everyone kind of went through that exercise. So that was a good kind of set of experiences for people.

The other one is I really like this idea of challenging people when they ask for headcount about why they can't use AI to do the things that they are asking for the individual to do. That I think needs to be built in culturally into organizations to help managers, help leaders, and even help individuals figure out, hey, wait a second, I never thought about using AI to solve this problem, like designing a greenhouse or, you know, designing a new lab layout or things like that. And then I would say the, you know, the other thing we're kind of doing now is these AI-first offsites. So I'm doing one with my reports, my leadership team, the week after next. So I'm doing kind of a full day away, and we're going to go through actually systematizing how we are going to work with individuals in the organization to try and make the organization more AI-first. What are we going to ask of people? How are we going to demand outputs from them? You know, what are the things that we're expecting of folks? And so we're actually trying to build from the top down. I need to make sure everyone in leadership is using these tools and then figuring out what are the methods for making sure that they're then passing that through the organization. So it's like dissemination of any management process. I would say, you know, you've got to start at the top. First of all, you've got to make sure that you as a leader are the right person to lead in this effort. And then you can kind of get your team, basically take those processes and distill them throughout the organization. Make sure you have the right levels of checks and accountability and performance standards and so on. So I think that's pretty critical.

Yeah. And you just think a company doing what you're doing, what I'm doing, and any AI-first company's doing, you compare that to a company that's not doing it, and you start to think about the gains and the velocity and the shots on goal you're going to get as an entrepreneur. It's just going to be three, four, five, six, seven times more efficient, more shots on goal, and if you're not doing it, you're just going to fall behind.

The reason AI is fairly transformative in biology is because biology has been digitized and is being digitized. So it's being turned into kind of digital form represented through data. So going back to pre-DNA sequencing days, right? Generating DNA sequencing data was super, super expensive. That cost has come down by hundreds of thousands fold, right? So today we can sequence the entire human genome for probably under 100 bucks.

Yeah, I think Bill Clinton was president when he went on stage and introduced it. And so, you know, the way DNA sequencing works is you take DNA, you multiply it many, many, many, many times over. That's called amplification. And then you chop up the DNA into little short sequences. And those little sequences get picked up on a chip, the DNA sequencing chip, and they're moved one read at a time. And ultimately, you get like these little short reads. And so you've got all these little sequences in the like 140 base pair long sequences that comes out. That's the DNA read. And what you then can do is statistically reconstruct that strand of DNA. And there's multiple strands because there's multiple chromosomes. So you're basically able to put together the complete genome ultimately of the human or whatever it is, whatever the sample is that you're reading. And so all of the little pieces of that, and that's just optical scanning, Jason. So I don't know if you like realize, but it's like there's a chemical that creates a little bit of a different fluorescence for A, C, or G. And so then you're picking this up on this little optical scanner and you're running millions of these a second. So you're just taking images, images, images, converting those pixels into an A, C, T, G, and that'll give you a little readout on the DNA.

So now fast forward 25 years, and DNA sequencers are super cheap, super fast, super efficient. And now with large-scale compute, we can very quickly assemble genomes based on these little short reads, these shotgun sequencing reads, and that gives us insights into the sample that I'm picking up. Okay, why is that interesting? Well, that's interesting because then I can compare the DNA of an organism to the physical characteristics of the organism. So in the case of a plant, you know what genes and what variations of genes are causing the plant to be tall or short, have big roots, little roots. But as I gather more and more data, I can actually start to predict complicated phenotypes or physical characteristics of plants by having AI figure out combinations of genes and create a model that predicts a phenotype as a function of different genes. So sometimes it's not just one gene being on or off or one gene being one variation or another that makes a plant tall, but it's maybe 50 or 60 different genes, and there's different combinations or variations of those genes that can predict the height of the plant or predict the depth of the root or predict how much yield you'll get out of that plant and all these sorts of things.

So what's happened in plant breeding, which is an area we spend a lot of time in, is you can now say, 'Hey, I want to get plants that are tall, that have deep roots so they can survive drought, that can make a lot of food, that can grow really fast, that are really energy efficient in terms of using sunlight, disease resistant, all the things you want.' And then I can collect all this data we have of how different plants performed in different environments, and I can look at the DNA data of those different plants, and the model starts to predict what combination of genes are going to make the plant you want. So it's almost like AI is giving you the DNA sequence of the organism and saying, 'Here's the perfect organism, here's the perfect DNA sequence,' and then I can go look at all the plants I can use in my breeding system, and then there's other technology to breed plants to get them together to make that thing.

This is happening similarly in microbiology. So where they're making little E. coli cells or yeast cells that we're then programming to make a molecule. So I can program E. coli, for example, to make a biologic drug. And that biologic drug, I wanted to make the perfect protein that's going to bind to a cancer cell and then trigger my immune system to kill that cancer cell. And so there's this whole world of predictive protein structure that's being used to create new biologic drugs. And so we can look at the performance of different proteins historically. Which ones bind? I can make a prediction that this protein would bind to this specific cancer cell in this way. Based on that prediction, here's the DNA sequence I would likely need in the E. coli cell to get it to make that protein. I extract that protein and I turn that into a drug. And so there are so many applications of how we are now able to treat DNA or genomes almost like a language model. And so people are calling this concept the genome language model or the GLM.

Yeah. The GLM is effectively predicting or scoring a DNA sequence based on some objective. And that is being utilized in drug discovery, in plant breeding, in human health.

Yeah.

Yes. And they funded the Arc Institute, which has put out an open-source model called EVO 2. And EVO 2, what you can think about it as is a grammar checker for DNA. So you can put any DNA sequence into it, and it will tell you whether or not that gene will be dysfunctional or functional, or if there's an issue with that gene. And the way they built it is they put all the DNA data that they could gather from all over the world, from all these different organisms, and they basically trained this model to say that DNA sequence looks like it makes sense or doesn't make sense. No idea why, no idea how that model works, but it's actually able to predict the functionality of a gene. And we use it in production at Ohalo. So we actually use it to look at the plants in our breeding system. We feed the whole DNA sequence of every plant in, and that model will tell us, 'Hey, these genes are dysfunctional or may have issues,' and that gives us a score for that plant before we've even phenotyped that plant. And so there's a lot of ways that we can kind of start to utilize these models in different production environments to accelerate biological work like plant breeding.