#84 - Three Breakthroughs: AlphaFold 3

Attention listeners, please prepare for entry into the Tech Optimist podcast by Alumni Ventures. This is your gateway to the cutting edge of innovation and visionary ideas shaping our future. Keep your mind open, your curiosity sharp, and your optimism fully engaged. Our guide today, me, my name is Smita Herrick, will take us through the thrilling twists and turns of groundbreaking technologies and inspiring founders. The future starts now.

Imagine just every scientist in the world now armed with this resource. What, what's possible in the field of healthcare and biology and drug development, drug discovery? It's, it's staggering.

This is Naren Ramaswamy, Senior Principal at Alumni Ventures.

We're going to be really impacting almost all white-collar work in the '25-'26 timeframe and probably by 2030 have something that far exceeds kind of human capabilities.

And this is founder and CEO Mike Collins. And hello, that's me. My name is Samantha Herrick, and I am the producer and host for this show. Welcome back to Tech Optimist, everyone. I am your guide, Sam Herrick, and I'm here to connect the dots and unpack the tech shaping our world. Today's episode is packed with innovation, debate, and a glimpse into what the future holds for AI. So now we're joined by 2 incredible legends here behind the scenes at Alumni Ventures, Naren Ramaswamy and Mike Collins. And together they're going to explore some of the most transformative breakthroughs in artificial intelligence today. Very AI-focused episode, which is pretty fascinating. So we're going to dive into the concept of AI scaling and scaling in terms of How different models can be used for different things, right? I'll let Mike and Naren sort of take it from there, but that's just a general summary of what's going to go on today. But then we're also diving into AlphaFold 3, Google DeepMind's latest milestone in protein modeling, and what its newly open-source capabilities mean for researchers in the industry. From AI scaling challenges to game-changing advancements, we're going to uncover how innovations like these redefine science and technology. And then the conversation turns to AI agents, the possibilities they unlock, the risks they carry, and the different visions for how they might shape the future. Mike and Narain don't shy away from any controversial viewpoints here, so you can expect some very lively discussion about what comes next for humanity and in the age of AI. I'll be here adding some extra context along the way, but for now, I'm going to turn the mics over to Mike and Narain. And then really quick, we're going to hop into a disclaimer and an ad before we jump into the rest of the show. Hang tight, we'll be right back.

Do you have a venture capital portfolio of cutting-edge startups? Without one, you could be missing out on enormous value creation and a more diversified personal portfolio. Alumni Ventures, ranked a top 20 VC firm by CB Insights, is the leading VC firm for individual investors. Believe in investing in innovation? Visit av.vc/foundation to get started.

As a reminder, the Tech Optimist podcast is for informational purposes only. It is not personalized advice and it is not an offer to buy or sell securities. For additional important details, please see the text description accompanying this episode.

Hi, welcome to Tech Optimist. This is where I and my teammates get together, and talk about really cool things going on in venture capital, technology, and innovation. My name is Mike Collins. I'm the founder and CEO of Alumni Ventures. I'm joined by Naren Ramaswamy today. Hi, Naren.

Hey, Mike. How are you?

I'm great. So I'm going to kick it off today. There's been a lot of heated discussion around a topic called AI scaling. And again, my understanding of this is that there's 2 camps. One that basically says, in order for AIs to become more powerful, we just need to throw more compute at it. And the more, the better. And there is, I think, Sam Altman and some others are in there. And it's basically, if we throw enough compute, you don't have to think through much else. The It'll— the compute, the AI will figure out everything. And that is how we get better models, more information, more utility, and ultimately artificial general intelligence. There is another camp that is equally as convinced that, you know, we're going to asymptotically approach diminishing returns here and that there's going to be— and not that they're saying that it won't happen, but that the breakthroughs happen by improving other vectors of innovation. And again, way above my pay grade to be what those would be. But, you know, those things I think have the implications for the rest of us, basically on predictability and timing. So, if it's just more compute that we know, that we can measure, and we can see where this— we can anticipate with pretty strong accuracy that we're going to be in the era of AI agents sometime in 2025. And we're going to be really impacting almost all white-collar work in the '25, '26 timeframe, and probably by 2030, have something that far exceeds kind of human capabilities to the point where, you know, it starts to get scary in that. 2029, 2030 range. Then there's the other camp that says, you know, this is more like inventing and you plateau out and then you have a breakthrough and you make a ton of progress and then you plateau out again. And I think this just says it's a little less predictable, a little harder to say. I think generally it says it's going to take a little longer, although with breakthroughs, it can swing the other way as well. But again, I just know that this is something that is hot on the topic of people. And because AI has such profound implications for our society and for venture capital and for innovation, I just think I wanted to share that this is going on and wondered what you knew about it and/or had to say about it, Naren?

Thanks for bringing that up. I've been hearing about this chatter in Silicon Valley for the past couple months, and it seems like, after talking to a few experts, it seems like the jury is still out on where we'll end up, which approach. But a lot of people are saying that there might be both approaches for different use cases. You'll have this AGI type of model that Sam Altman is raising $7 trillion for.

Yeah, right.

Which, by the way, is 2 times the GDP of the UK, right? He's raising that for the Middle East.

Yeah, just 2 UKs. We'll start using just 2 UKs as a unit of measure, right? Yeah. And China's doing 5 UKs. Yeah.

And, you know, we— you talked about just the, um, productivity and timing implications, but I'm talking— I'm thinking about the macroeconomic implications of something like that, dollars moving at that scale. So that's one, you know, I think there's going to be one piece of that. But for more specific enterprise use cases, maybe in legal tech, that you're looking at a, you know, set of documents that you need to be able to query, you actually do better with a smaller model that is more focused on that set of documents rather than the entire internet because— The scope of errors and hallucination are much higher the larger the scope is. So, you know, people think that there will be a bit of both depending on the use case. I think I agree with that.

Yeah, I mean, I think it definitely in kind of the short term, right, meaning years, I think it intuitively makes sense and it's consistent with my own experience, which is You know, there's a niche need, like I'm trying to do something and it's very particular to a certain customer, a certain workflow, a certain experience set, a certain way that we want to— humans want to interact, um, that requires a little more of a, a narrow, specific AI approach.

Yeah.

that is really targeted, kind of a rifle shot to a company or a specific customer type or a specific product. And you don't need the big God model for that. You need strong models, but really focused, trained, and customized to what you're really trying to achieve. At the end of the day, you're still in an economy where you're trying to solve a problem through a product or service. So, I tend to think, and let's just try to be reasonable in the timeframe, in 2025, I think that there is a lot of the value creation is going to be in these use cases. Right. And again, I I had the fortune to study with Clayton Christensen, and, you know, just one of his frameworks for thinking about marketing in particular was to really focus on kind of the customer problem and what job is the customer hiring you to do. And it has always been very powerful for me. as an innovator and a venture capitalist to be always thinking about that first and foremost. And then if you apply an AI to help you deliver that to this group of people, I think it can be a huge amplifier. But there's a lot that goes into the data, the wrapping, the understanding, the human touch in there. Where do you trust? Where do you need human intervention? So I tend to intuitively believe in the short term that we're going to see much more of these custom approaches, language models that are a little more focused and tight and tuned to your particular business or use case while the big heavyweight stuff is there. And I even find that with the ChatGPT suite of models, right? There are times where I'm going to go just use 4.0.

Yep.

Or 4.0 with— it has a, you know, kind of the canvas feature if I'm doing some writing. But a lot of the analytics stuff, I used actually their, their mini reasoning model. Works super fast, gets it done. Their big boy model sometimes just seems like way too You know, for a marketing problem, it just seems like you're bringing kind of a, you know, a gun to a knife fight.

Over-engineering.

Overdone.

Yeah. And who knows what the compute costs are for each of those responses on the backend for them, right? And people have said that OpenAI uses what's called a mixture of experts model, which is under 4.0, you know, you might have different types of models to optimize cost for them, right?

Yeah.

And they used to, as you remember, they used to have that 25-query limit, which they've now taken off if you're part of the paid license. So, you know, it's different approaches to that same outcome. But this is a good segue, Mike, to the topic that I wanted to bring up, which is more of a specialized model as it relates to biology called AlphaFold. And we've talked about AlphaFold because it's come out of Google DeepMind. And recently they announced AlphaFold 3. And the big breakthrough here is that they open-sourced the model, which has great positive implications. It was not actually open-sourced. There was criticism against the initial restricted access. And then now DeepMind went ahead and open-sourced it, including the code and the model weights. specifically for academic use, which is great, right? Anybody who's interested in researching how proteins are folding, but also how they interact with each other. Now, so that's the big breakthrough in this model, right?

That's the breakthrough, right? With 3. Yeah.

Imagine a tool so transformative, it's reshaping the future of molecular biology and drug discovery. That is exactly what AlphaFold 3 from Google DeepMind brings to the table. Recently made open source for academic use, this advancement marks a new chapter in how researchers can explore the building blocks of life. So now we have talked about AlphaFold 3 and AlphaFold in general a few times on this show, but if you need a bit of context and another summary on what it is, AlphaFold 3 isn't just another update, it's a leap forward. So this iteration models proteins interacting with RNA, DNA, and even drugs pushing boundaries that were once considered unreachable. This capability opens doors for breakthroughs in areas like drug discovery, where understanding protein interaction is essential. The decision to open-source AlphaFold 3 follows significant criticism of the limited access researchers had to earlier versions. DeepMind responded by making the source code available to everyone for non-commercial use and allowing academic researchers to request the powerful model weights. This step ensures scientists can dig deep into protein structures without barriers, provided they're working in a non-commercial setting. The timing of this release couldn't be more notable, too. Just as Demis Hassabis and John Jumper received the 2024 Nobel Prize in Chemistry for AlphaFold's pioneering impact, the world now has broader access to the technology. But DeepMind hasn't let go of its commercial reins entirely. The company continues to balance accessibility with innovation through its spinoff, Isomorphic Labs, While ensuring no unauthorized commercial use of the technology. Now, AlphaFold 3 is more than just a tool. It's a shift in how science collaborates, advances, and pushes the frontier of what's possible. So what does this mean for the future of medicine and molecular biology? Stick around as we explore how this game-changing tool could revolutionize research and innovation across the globe. Mike and Niren, take it away.

AlphaFold 2 was great at protein folding, and that's what led to Demis from DeepMind winning the Nobel Prize this year in chemistry. But what's fascinating to me is applying the technology that underpins large language models to something like protein folding. And just for our listeners, AI does very well at pattern matching given a set of rules. That's what language is, but that's also how nature operates, right? You have the laws of physics, laws of biology, and DeepMind has done an amazing job teaching

this

model.

Yeah, biochemistry.

Exactly. And so this is— imagine just every scientist in the world now armed with this resource. What's possible in the field of healthcare and biology and drug development, drug discovery? It's staggering.

Yeah. I mean, we're really— this is unlocking how humans work, right? At the fundamental level, proteins, amino acids, the folding of these things, the interactions of these things has been largely trial and error. If you think about like how we've discovered most of the therapeutic drugs to date, it's been happy accidents, right? From penicillin to, you know, Ozempic to, you know, you know, blood pressure medication turns into Viagra. I mean, these things were just, you know, you know, accidents, frankly. And The ability of these models to turn them into engineering problems is— that is the game changer. When instead of a lab and experiments and, oh, we get some deep research and we have a little better understanding about this pathway or the cycle or these kinds of interactions, and we run some— just really slow. But if you can create a simulated human, right, that really begins to understand the just enormous complexity of some of these subsystems. The ability then to test, simulate drugs, simulate solutions is just orders of magnitude more effective and faster. And yes, at the end of the day, there is still a very laborious physical world we have to interact with and, and appropriate levels of scrutiny and regulation and testing. But, um, if you can do all of this work, um, up front, that is way more efficient, way more effective. You know, we can, we can make many more drugs available. We can reduce the time to market. We can— the number of drugs that'll, you know, look good in phase 1 but then poop out in 2 and 3, if that can go down significantly, it just is— it just has huge implications. So, yes. DeepMind, fantastic. AlphaFold 3, great. But again, I just think it points to where this where this is heading. And again, just huge optimism over the next decade or two in understanding the human condition and disease.

And Mike, just the— this knowledge previously would have been in the hands of some biology experts somewhere in the world in a lab.

Yeah,

but now is available like for the entire world. to access at their fingertips. The same way that the internet democratized information, this is democratizing expert-level information now. So it's amazing.

And in drug discovery, we've seen this time and time again too, is the way a lot of times academics have been incented is very siloed. And that really a lot of great things happen when there is cross-pollination between labs and offices. And I actually think a lot of the work with AlphaFold and DeepMind has been very collaborative, actually. But I do think this also promises that the incentive structure to be more open source and sharing of information and learning hopefully will also accelerate. So we are a little less siloed.

Yeah, absolutely.

The power of people in a cluster as well as compute.

Commercial break incoming. Sit tight, we'll be back.

Exceptional value creation comes from solving hard things. Alumni Ventures Deep Tech Fund is a portfolio of 20 to 30 ventures run by exceptional teams who are tackling huge opportunities in AI, space, energy, transportation, cybersecurity, and more. These game-changing ventures have strong lead venture investors and practical approaches to creating shareholder value. If you are interested in investing in the future of deep tech, visit av.vc/deep-tech to learn more.

And that, I think, will take us right back into, again, just kind of state of play on AI again with a lot of stuff I think going to be released in the next 3 to 6 months, um, related to agents. And so again, for, for those that don't follow the industry, uh, quite as closely, I mean, agents, agents are basically the computer, the, the software going and doing things on your behalf. Besides just writing copy or email or, or giving you, uh, text, the ability to go out and basically interact with any website to do things that you would do on the website is really how I think one could think about— should think about it, which is if you could do it by going on to Expedia, or if you could go to a site where you have to do some work for, you know, in a B2B, you know, you might have a salesperson interacting with Salesforce and doing 19 different workflows and tasks after, before or after a customer meeting. That's interacting with a website, software, tools. The ability of these systems to begin to do that on your behalf is in testing and coming very, very quickly. And I think will unleash the same kind of enthusiasm, backlash, concern, the whole bundle of stuff that we had when ChatGPT 3 and 4 came out. So, you know, we've got, you know, Anthropic with Claude seems to be one of the companies kind of, you know, with their API. I think OpenAI's technology is rumored to be called Operator or Operations or something along those lines. Again, we're seeing Google not sitting on the sidelines here with Jarvis. You're seeing Apple really trying to play catch-up here with Siri and clearly having some distribution leg up here in order to have Siri do things on your behalf. And I think there's an arms race. So, the fact is, I think the fact that there's not one There's 4 also leads me to believe it's going to be sooner or later because I think there's going to be, like there was with OpenAI, there's a lot of buzz when you're the first one to make the big breakthrough. Yeah. And so I think there's, I'm sure, heated debates on is this ready for prime time? Is it going to work? Is there going to be backlash versus are we going to be late to the party?

Yeah, I think this is similar to self-driving cars a little bit in that the last 10, 20% will take 80% of time.

Yeah.

Because the pieces of the underlying technology are already here. Like the same way that we can read a website and perform a click, software can do that too. In fact, the software is— meanwhile, AI is solving International Mathematical Olympiad problems better than humans can, right? So the technology is there. It's just about making it private and secure and in a way that also, how do you trust someone to— it's like inviting someone you don't know to operate your computer and do things on your behalf. How do you develop that trust?

Yeah.

And it might just take time. And I'm curious to see how these different approaches differ, not on the technological standpoint, but on the use and trust standpoint. How do they engender trust amongst hundreds of millions of users? It'll be interesting to see.

Yeah, I think the analogy you gave, Naren, of self-driving cars is an excellent one, right? Because it's like there, the stakes are really high, your life, right? So I think there's a lot of, it's a very slow trust. that you're going to give to people. I think the same will hold true with people and these agents, meaning it's one thing if I'm going to trust you to pull things together for me to finally decide, but I'm not giving you my credit card. At least I'm not going to give you my credit card out of the gate in the first period of time, right? Yeah. But Those things change over time, and humans do develop trust, and they go through this cycle of skepticism, baby steps, trust. Can't live without it. Yeah. So, yeah, I think it'll be interesting, but I really am excited about getting on this vector of growth and improvement. I'm excited to see how the different approaches work, how people take different strategies, and the marketplace will determine where we're going to trust it first.

Exactly. It's like having a team at your disposal out of thin air.

Out of thin air.

Instead of being an individual person just going about your life, even irrespective of your relationship status, right? Every person will have, you know, like 10 different AI agents that'll do things for you.

So, and so you become more of an orchestrator of your agents to do what you're, you want to have done in a day. So, and that'll take some getting used to. And as you said, a lot of trust. So, but great. It was good catching up with you, Naren. Uh, wish you, you know, a happy Thanksgiving and we'll talk, uh, we'll talk again in a

Likewise. Thanks, Mike.

Thanks again for tuning into The Tech Optimist. If you enjoyed this episode, we'd really appreciate it if you'd give us a rating on whichever podcast app you're using, and remember to subscribe to keep up with each episode. The Tech Optimist welcomes any questions, comments, or segment suggestions, so please email us at info@techoptimist.vc with any of those, and be sure to visit our website at av.vc. As always, keep building.