No Priors 第140期｜与 Benchling 联合创始人兼 CEO Sajith Wickramasekara 对谈

Hi listeners, welcome back to No Priors. Today I'm here with Sajith, the co-founder and CEO of Benchling, the system of record for biotech R&D. Today we talk about the state of AI and bio, Benchling's bet on AI agents to help scientists make better decisions, experiment faster, and deliver drugs more effectively, why drug programs are so expensive and fail so often, and how to build a culture of science and software together.

Sajith, thanks so much for being here.

Thanks for having me, Sarah. Excited to be here.

For our general listener base, can you give us an overview of what Benchling is and the scale of the business today?

Sure. I'm one of the co-founders of Benchling. We make modern software for scientific progress. We started the company about 13 years ago. It's been a long time.

Oh my God.

I know. I'm a software engineer by background, but I worked in a biology lab. I was really interested in medicine, and coming from the world of software, I knew that software developers have amazing tools for working on code and collaborating. When I got to the biology lab, I found that scientists had paper notebooks and spreadsheets that would sit on their desktops, and it was terrible. It was really hard to work together, and I think that was really frustrating for me personally. I thought, a little naively at the time, how hard would it be to build good tools for scientists? So I started working on Benchling, which helps scientists design molecules, plan their experiments, run those experiments in the lab, get the data, organize it, analyze it, and then share it with their colleagues.

Today, we work with about 1,300 biotech and pharma companies and scientists at over 7,000 academic institutions and universities all around the world. Our software powers household names like Moderna, Sanofi, Eli Lilly, and Regeneron, but also cutting-edge biotech startups—the future AI biotechs like Isomorphic Labs and Xaira, and companies like that. We get to see the innovation happening across the entire biotech sector and build software that helps power it.

I'm excited to use that vantage point and ask you a bunch of questions about biotech and the macro. But just so people who don't come from the domain can picture it a little bit better, I think I can picture gene sequences.

Sure.

And the assay, like, said yes or no. What other types of data are actually in it?

I think what's really interesting for everyone to understand is that making a drug involves 9,999 steps after you come up with a molecule. To make a medicine, you have to find a biologically meaningful target in the body—something you want to drug. You have to design a molecule, optimize that molecule, and test it in Petri dishes and cell lines and in animals—various kinds of animals. Then eventually you get to the point where you can take it to a clinical trial, and you're testing it in subsequently larger groups of humans.

All the while, you're figuring out how to manufacture this thing and develop a process to make it scale economically and safely, with high quality, all while navigating regulatory bodies, so that eventually, in 7 to 10 years, you can have a drug that you give to people commercially. And even then, there's still more work there. It's just an incredibly long and complex process.

Where Benchling focuses is on all of the scientific data that comes out of the lab. Everything from all the different types of molecules that are being created, to how they're related to the work that went into creating them, to the different types of tests that you're running on them, to the data coming back from the animals, to the scale-up data coming out of the fermenters when you're figuring out the process to manufacture it. All of that incredibly rich and heterogeneous scientific data has to be brought together in one place, organized, and made searchable so that scientists can make decisions based on it.

If we zoom out, for people just looking at biotech from the outside, it seems like a very macro-sensitive industry, right? And we are perhaps coming out of an ugly period. Can you characterize where we are in the biotech macro cycle?

Yeah. I'm definitely not a macro specialist; otherwise, I'd probably be an investor or something like that.

But it's all your customers.

Yeah, it is. I would say biotech has definitely gone through cycles. The last couple of years have probably been the equivalent of the dot-com bust happening for biotech. It's been a tough time.

COVID was sort of the peak, when mRNA was this thing that kind of reopened the world, and there was a lot of generalist money that came in and a lot of exuberance and excitement. The dot-com bust equivalent wasn't just because of that. There were changes in interest rates, tariffs, regulatory uncertainty, China, and a bunch of different factors, including scientific technologies that we got really, really excited about, that are still very important and promising but maybe haven't become commercially successful as fast as people wanted. So, a whole confluence of factors there.

What are you referring to in terms of scientific technologies that got people hyped?

Yeah, I would say there's a lot of generalist excitement for gene editing, cell and gene therapies, and RNA.

So, new delivery methods. I would call them categories or form factors of medicines—modalities is the word. The last decade, maybe even longer, of biotech has really been this story of new categories of medicines being invented and taken to patients. Some of these have approved gene-editing medicines. There are approved cell therapies where you're reprogramming the patient's immune system, approved gene therapies, and approved mRNA medicines. These are real categories, but I think investors and companies got very excited and put a lot of money into them, and we're kind of in the trough of disillusionment for some of them now.

And my understanding is that they have taken longer and been more expensive.

Yes.

Than people expected—than investors expected.

Absolutely. In 2021, every biotech was getting told by investors, “You need to build a platform company that's going to cure a bunch of different diseases,” and, “Here's hundreds of millions of dollars, and capital is free.”

Investors can change their strategies a lot faster than companies can. A lot of those companies, because they were taking such a big risk on a new form of technology—they were going to be the next RNA company or the next cell therapy company—picked diseases that might have been simpler problems to solve, with smaller patient populations. Then all of a sudden, investors changed their minds: platforms were no longer valuable. You're working on a thing that's just supposed to be a proof of concept, and all of a sudden it's defining you. So it's a really tough spot for those companies to be in.

What's the relevance of China in all of this?

If the last decade was about biologics and these new modalities, I think the next decade is going to be about speed and cost. People want more drugs, and they want them cheaper. China is very good at things related to speed and cost.

All of a sudden, in the last couple of years, you've seen this rise of Chinese biotech companies that are able to create molecules and bring them to patients in clinical trials in China—just early phases of clinical development—really fast and really cheaply, even in some of these new modalities. We've seen this huge uptick in pharma going to China and buying molecules that they typically would have bought from American biotechs.

And this is top-30 pharma?

Yeah. These are the biggest companies in the world—the Mercks, Pfizers, Lillys, and so forth. Many of them have gone to China and bought molecules that historically they would have bought from American biotechs.

Are there medicines that people would recognize on the market today?

One of the most notable medicines that people might recognize is called Carvykti, and it's a Johnson & Johnson medicine. Johnson & Johnson partnered with a Chinese biotech called Legend Biotech. They saw the data that Legend presented, and it was a time when people were pretty suspicious, so they were like, “The data is probably not going to replicate. It might not be real.” But J&J, I think, saw it and realized how promising it was, and they've taken it. It's actually a cancer immunotherapy. I think it's for multiple myeloma. That medicine is very commercially successful and widely distributed in the U.S. to those cancer patients.

What's been the reaction of Western biotechs to this?

It's a mixed bag. I think there are some folks who are inspired that American biotech needs to be faster, cheaper, and more competitive. There are some more nationalistic reactions, I think, of, “Hey, there are different regulatory or ethical standards over there. Are the data going to replicate?” So there's some skepticism as well.

But by and large, I think it's very much here to stay that China is going to be a major, major biotech player.

Elad Gil

Yeah, we can spend this whole time talking about macro.

Yeah, I want to get to the meat of our discussion, which I also think is that there’s some premise that the answer to faster, cheaper, better might in part be AI in biotech. I think I believe that now, and it’s really interesting to see the general public, big tech startups, the model labs—everyone is saying AI is going to cure a disease. So it’s very good that everyone’s excited by that.

Elad Gil

I don’t think of you—I think you’re an amazing CEO, but not really a content-marketing guy to date. You wrote an essay very recently that I thought was amazing about how we can possibly change the scientific field in biotech with AI. Can you give us the CliffsNotes on it? Then we’ll link it in the show.

Absolutely. Yeah, I think, maybe to step back, one thing I just wish people would appreciate more is that medicines are magic. I think we take for granted how awesome medicines are.

I think prescription drug sales are 9% of healthcare spending in the U.S. We obviously have this healthcare cost problem, but drugs have an amazing ROI. The best part about drugs is that they go generic, so a drug today is only going to get cheaper over time, and it works just as effectively. I take a statin today that probably costs almost nothing, and 20 years ago it was some expensive medicine.

Elad Gil

It’s not obvious that any other part of the healthcare system gets cheaper over time.

It’s not. Yeah, the rest of healthcare is very labor-dependent, and labor generally gets more expensive over time. I’m very optimistic for AI to help there too. But drugs are this amazing thing. We should want more of them, and then we get to stockpile more and more of these amazing medicines.

It takes over $2 billion and generally about 10 years to bring a medicine to market. Most of those medicines will fail very late in this process. You get 7 to 10 years in, you’ve spent hundreds of millions of dollars, and the clinical trial fails. The medicine is not safe or not effective.

So it’s an unbelievably difficult pursuit. It is probably easier at this point to send things to space or to put people on the moon than it is to get a new medicine approved. I know $2 billion probably isn’t that much—I feel like AI has desensitized us all. Everything is like 100-billion-dollar data centers and whatever. Two billion dollars—what’s that? But when there’s that high of a failure rate, it’s very difficult for investors to underwrite that.

While we had all these new categories of medicines being invented over the last decade, I think that’s important and it’s here to stay. But the industry has to change. The pressure on biotech to be faster and cheaper is just higher than it’s ever been before.

A lot of that cost comes from how artisanal the industry is. Biotech is this place where, if you look—I sort of take the digital and physical realms for a second—they’ve actually done a good job of systematizing the physical realm. You brought up sequencing earlier: Illumina has put sequencers on every single bench in every single lab, and now sequencing is this accessible tool to all of science. You could say the same thing has happened with different reagents, lab consumables, and things like that.

But if you look at the digital realm—how people collaborate, how data is structured and shared, and the workflows that are used in science, which is all about collecting data—all of that is basically bespoke and invented one-off by every company. It’s because those companies are playing a one-time game. The process is so long that you’re just trying to survive until you get 6 or 7 years in, show some clinical success, and a pharma company comes and buys you. So you’re not really building for scale and building for durability.

That seems like it also comes from some of the structure of where the innovation happens, right? Because if you were doing it across a whole portfolio, actually starting at zero, and you owned the innovation, then you would invest in the systems.

Totally. Yeah. If you were setting out to build a company that was going to—you wanted to build the next great pharma company and have a whole portfolio of medicines—you’d probably care a lot about that. But that’s such a high-capital, long-term, high-risk thing to do. It’s very hard.

After 7 or 8 years, when you have some good clinical data, do I roll the dice again and keep going for another 10, or do I sell? So I think, because it’s so artisanal, there’s this huge opportunity now with AI to get more shots on goal faster and cheaper, make better molecules, and then bring them to the clinic safely and faster. I think that’s the big opportunity.

Elad Gil

People get very focused on clinical trials because they’re the biggest line item, and they’re important, don’t get me wrong. But I think it’s actually a bit of a red herring. Yes, there are operational problems—some studies are designed badly, it’s hard to recruit patients, and the sticker price is really big—but at the end of the day, a lot of molecules are just not good. So we need better molecules, and we need to move them to people faster.

One other criticism that you imply in your essay as well about why the industry isn’t more efficient is that even the large pharma companies are not as good at buying innovation and finding it as they could be, right? Examples of GLP-1s and Keytruda—some of the amazing breakout successes—were not super obvious to the buyers.

Yeah, I think those two stories are really interesting.

Elad Gil

There’s a great quote from Dario, the Anthropic CEO, in his essay about the returns to intelligence in scientific progress being very high.

Talking about “Machines of Loving Grace.” Yeah.

Yeah. So the returns to intelligence are very high, and I think the stories of GLP-1s and Keytruda are great examples of that. GLP-1s obviously have just transformed obesity as a treatable disease when, by the way, it was a totally unfundable category of things 5 years ago.

Why do you think it was unfundable?

I think, again, because we know so little about biology and there are so many failures in that space. Running a clinical trial for obesity, where you need huge populations of people that you monitor for very long periods of time, is super, super expensive, and everything has failed before. Pharma companies generally aren’t willing to underwrite that stuff.

Neurodegenerative diseases are the same way. Alzheimer’s is this graveyard of billion-dollar failures, and it’s getting back in now, but there was a period of time where everyone left the space.

The core science for GLP-1s was kind of sitting on the shelf in some sense. It’s been known since the 1990s, and so it took some insights and conviction. Then, all of a sudden, we have this category-defining medicine that’s going to go on to probably be the best-selling drug of all time. That’s happening.

Keytruda is a similar story, where there’s a molecule that’s gone through a couple of different acquisitions, and it’s almost like it’s at the bottom of some list to be out-licensed and sold off. Then a competitive threat pops up, and someone sees that this is kind of like Keytruda. Credit to Merck: they had the courage to go all in after they realized what it could be.

It’s just another example of there’s a lot—it’s a pretty inefficient system, and people are pretty rational actors. It’s just that we don’t know a lot about biology, and our ability to predict what’s going to happen in the clinic is so poor. The cost to get there and to make those decisions is so high.

If you can get to the clinic faster or cheaper, failure in the software world is that you work on a product for a year or 2, spend a couple of million bucks, and it doesn’t work. But in biotech, you’re underwriting 4, 5, or 6 years, a big team, and hundreds of millions of dollars. So how do you compress that so you get feedback faster?

Elad Gil

So Benchling is a system-of-record company. It’s a data platform. What is Benchling AI?

Benchling AI has 2 major components to it. The first is tools for simulation. This is taking open-source models, proprietary models, and a company’s internal models and making them accessible to scientists directly in their workflow. The right model at the right moment in the scientific workflow is already set up so that a wet-lab scientist without computational skills can use it effectively, and then the results are linked to all of their other information in Benchling.

We also see that leading up to being able to help scientists recommend the next best experiment to run, based on all the work they’ve done in the past plus all the public literature available. We think it’s an exciting way to approach the co-scientist problem.

The other facet of Benchling AI is agents that automate work for you. We’ve released this deep research agent. It works similarly to the deep research agents from Anthropic and other foundation-model labs.

What it does is work over Benchling data with the context of the Benchling data model. It enables scientists to ask these very difficult questions, and science is fundamentally about asking and answering questions.

For our customers, it helps them do a type of question that previously would have taken weeks or months in just a couple of hours. A great example is a customer that was getting ready to run some mouse studies. They were looking at 20 different mouse models and used our Deep Research capability to look at all the historical mouse studies they had run. It turned out that a bunch of the mouse models they were about to investigate—which would have taken 8 months, at huge cost, for a big experiment—had already been studied before, but the information was trapped in a lab notebook from many years ago, from a company that had been acquired, and all the people were long gone.

There’s so much science that lives in folklore and institutional knowledge, and that’s just lost over time. We view this as unlocking memory for these organizations and helping make scientific data reusable over time.

And they could just accelerate because they didn’t have to do that piece of experimentation anymore.

Exactly. We’re working toward a world where AI agents can do all sorts of different tasks in the scientific process, whether it’s generating reports and asking questions, or even composing experiments while you’re in the lab with voice and vision and things like that.

If you project out a few years, everybody loves to talk about this idea of the AI scientist, a lot of autonomy, AI co-scientists. What do you think is the role of scientists a couple of years out?

Oh, wow, that’s so interesting. When I hear about AI scientists, I think it definitely evokes this image of a fully AI-driven design-make-test-analyze loop, where we’ll sit back and let the robots give us drugs.

While I would love for that to happen, and I’m maybe more optimistic that, on a longer time scale, we will get there, I think in the next 1–2 years—which already feels like an eternity in AI time—I’m a little bit more bullish on the augmentation model. I think of it as a Waymo-versus-Tesla approach. You can do the Waymo approach to autonomy; you just need a lot of money and a lot of patience, and it’s going to take some time. I think the Tesla approach has been a little bit more about taking steps. I don’t want to call it incremental, because it’s not.

If you can get those ingredients and take the Waymo approach, which some companies have, that’s awesome. But for the rest of science, there’s a huge opportunity to make things better one experiment at a time, pick off a lot of low-hanging fruit, and see if we can get 7–10 years down to 2–3 years, with a lot fewer specialized roles and at a much lower cost to bring a drug to market. I think radiology is an interesting parallel. ML people have been saying radiologists are going to go away for 10 years.

Elad Gil

I think, like, 40. Yeah.

Probably. I think the model that’s worked there, though, is the copilot model. Truthfully, at the end of the day, with a radiologist, you probably need a human to be accountable for those decisions. It’s not just about the technology; someone’s got to be there to—I don’t know—get sued if something goes wrong. Yeah, I mean, that makes sense to me in clinical practice. I’m more hopeful that some of the experimental decisions can be more automated. But one question that I think biology faces—

Elad Gil

That other fields in AI face as well is the question of how you make these agents useful and transparent to specialists outside of the domain. If you think about engineers generating a ton of code, there’s a lot of, “Looks good to me. I didn’t really read it. I don’t know if that’s a good architectural decision. What’s happening?” How do you think about that for, for example, wet-lab scientists and computational analysis? They don’t necessarily deeply grok it.

Yeah. I think right now, when I look at biotech, we are in—so that’s absolutely the right point: Are scientists going to trust this, and how do we know if it’s accurate? Right now, I would say there have been amazing advances in capabilities that scientists could use in the life sciences, from the foundation-model labs, from BioAI companies, from everyone. It’s really awesome, but I think we’ve got GPT, but there’s no chat. That’s kind of how I think about it.

The chat—and I mean chat metaphorically—was the interface that made things really take off in software, and I don’t think we’ve figured out what that is in bio yet. We have some ideas, but by and large—and I just got back from a month on the road; I was in Boston, London, and a bunch of other places that are scientific capitals outside of San Francisco—most people aren’t really using that much AI in R&D yet.

Elad Gil

They all want to. They’re primed to, but there are a lot of concerns about accuracy, IP, security, and legal issues. I think the farther you go from San Francisco, the larger those concerns get. So you’re optimistic that you can make a lot of the context—whatever is important for scientists in different domains to understand about an output—legible through the product itself?

Legible enough to be useful. Yeah.

I think in a vertical, 90% of the work is actually translation. It’s taking something and making sure scientists trust it, that it’s the right point in their workflow, that it’s easy to use, and that it’s accurate. I think the AI that wins is going to be the one that people actually use.

Give us the temperature check of what large pharma and your customer base think about AI right now. They’ve got these AI officers.

Oh, yeah. There’s excitement for sure. There’s optimism and belief.

Elad Gil

Yeah.

I think they’re pretty pragmatic, though, and they’re all looking to transform, but they’re being methodical. Most of the large pharma companies that I’ve worked with have copilot tools and things like that, and they’re doing a lot of pilots of different technologies, but I haven’t seen their R&D organizations transformed yet.

The one place I would say pharma has really leaned in and has an advantage is that they have incredible data-generation capabilities. Many of them can and should be training models, particularly for experimental data generation.

Like experimental data generation.

They can generate data to train their own models at a scale that most biotech startups can’t match. So while it’s early on the agentic, how-we-work side, I think you’re going to see very unique models come out of pharma where their computational scientists are building interesting predictive models, similar to what’s happening in the open-source world.

What do you—can you help characterize what useful models we already have on the discovery side, and where you think we are in the cycle of having enough to make a real change in the—

Elad Gil

Overall cycle time.

Yeah. It’s been really cool to see the whole ecosystem of these tools grow a ton. Open source wasn’t really a thing in biology or in science before, and in the last 2 years it feels like it’s a thing that’s here to stay. You’re seeing all these interesting models come out, like the Boltz models, for example, on the structural-prediction side, and I think that’s a really interesting thing that’s going to change biology.

You’re seeing new approaches to federated learning as well. Eli Lilly put out an announcement about a project called TuneLab, where they’re taking their internal models and making them available to the broader scientific ecosystem. A pharma company is saying, “You can use our models, but it’s give-to-get, so we get to train,” and it’s federated or something like that. They don’t get to see the actual scientific data, but I think those approaches are the beginning.

We’ve got some really cool stuff. There are problems that I think are fairly tractable in terms of structure prediction, antibody developability, and so forth. That’s really good. But there’s a lot of work in front of us in terms of models that are more predictive of what happens when you get into patients, for example.

Don’t get me wrong: Discovery is really important, but there are so many other steps that have to happen after you have a concept molecule. Even if you’re much faster at making molecules and have a higher success rate, you still have to come up with a process to manufacture that molecule. Now you have even less time to do so if it’s a byproduct of success. How do you optimize manufacturing processes to get more yield and speed, and lower cost, out of a molecule?

Is that where you would say the highest-value missing predictive-model opportunity is? A bunch of naysayers would be like, “Okay, yes, I’ve heard about AlphaFold, and people are working on antibody prediction and creation platforms, but we are still many years into this premise.

No drugs out the other end of the pipeline that are AI-discovered.

Yeah. I feel like the naysayers have this worldview in mind where it’s like, “I just type a disease in and get a molecule out. Amazing—AI-discovered drugs.” This is where I go back to my mental model: There are so many steps, and those steps are all cumbersome and difficult. This is a game of making each single thing better.

Some of the steps matter more than others. Having the right target or having a great molecule generated is fine, but there are still many, many years after that that we can compress and shave off. Right now, I would almost argue that we should be thinking about what share of experiments have been touched by some kind of predictive capability, some kind of simulation, or some kind of AI. I bet that share is getting higher every day.

Elad Gil

Part of what I think has been really interesting—and there’s good and bad about the investor enthusiasm around both AI’s potential impact on biotech and the potential for platform companies—is this theory that we’re going to have very different business models in biotech. Do you think that’s going to happen?

I would like it to happen. As a toolmaker, I think there should be more tools. Tools are good. I do think with some of these model companies in the biotech world, there’s going to be an interesting question of whether, in the fullness of time, they morph into their own therapeutics companies with their own pipelines.

I think it’s unlikely—but possible—that they’re just going to remain pure model companies that do deals with pharma, where pharma pays them $100 million upfront or something like that, and they have 5 customers. I feel like model-building is probably commoditizing too fast for that to be an attractive business model. But taking that expertise and becoming fundamentally better at research and early development, making molecules, and morphing into a biopharma company seems like one logical path.

I think there’s a world where models can be more effectively distributed to the larger biopharma community. Rather than doing business-development deals with 5 companies, it’s more like a traditional software sale. We’ve got a bunch of models in Benchling. They’re mostly open source, but we’ve also got Chai and AlphaFold and things like that. Is there a model where some of these are pay-per-use or fee-for-service, almost like SaaS, so the entire biotech company benefits from them? You can build models and have a scalable business model on the other end. I think that would be really interesting.

Then there are going to be more data transactions, I think, as well. Data is interesting for a field that really depends on data as its currency. Everything is about data on the molecule, yet you see very few data transactions.

Elad Gil

That’s because no one trusts anyone else’s data. You wait until there’s a clinical trial and the data is positive, and then you buy the molecule. You’d think you’d see a lot more selling of data before that, but you don’t. Data is very hard. You don’t know what format it’s in. Do you trust the way it was created?

If there’s tooling and normalization around it, you might be able to transact on it.

Elad Gil

Okay.

Will people be selling their negative data at some point into a pool that other people can learn from? I don’t know. There’s all kinds of crazy stuff I can think of.

Elad Gil

I’m sure you’ve heard, in the 13 years you’ve been building Benchling, the conventional wisdom that the only way to create value in pharma is through assets, not tools. Where were they wrong? Or maybe the tools just weren’t that important before, and they weren’t as embedded as they needed to be?

Yeah. I don’t know if this is Thermo Fisher and Danaher. They’re sneaky big companies, and I think people don’t always realize that. They’ve done it largely through the systematization of tools in the physical realm: instruments, reagents, services around them, and so forth. I think there’s something that rhymes with building great tools on the digital side.

Frankly, looking back, the technology probably hasn’t been there. When we started Benchling in 2012, the cloud was the norm everywhere, but most of the life sciences industry was using paper, on-premises spreadsheets. We spent the first couple of years—

Elad Gil

It’s wild for such an advanced field in other areas.

Yeah. That’s because you could argue that it’s such a high-stakes game of poker for them that the only thing that matters is whether this drug gets to patients and succeeds. Pharma has pretty healthy margins, so operational efficiency isn’t always going to improve the odds of success.

We spent the first couple of years evangelizing: Bring science online. It’s going to be better. Then we spent the next 10 years convincing people that structured data mattered, because that’s the core premise of Benchling. It’s a system of record that helps you have a data model, and every time you do experiments, that data model is populated with information. You can ask questions. There are people who got it and believed in it—

Elad Gil

It seems obvious to tech people.

It seems obvious, but it’s not free. A piece of paper is much easier, and an Excel spreadsheet is much easier. There are people who believe in it and people who maybe weren’t convinced. But now, with AI, I think the benefits are much more immediately obvious to everyone. That’s going to be an amazing tailwind to try to do better here, and I think it will convince a lot of people who might have been skeptics in the past.

Elad Gil

Yes. I don’t come at that from a holier-than-thou view, because one might claim that in venture investing, the only thing that matters is the quality of the next decision and whether or not you found the winner. There are a lot of tech people with a lot of pen and paper, actually, so I think that’s likely to change.

Two things. One is that all the foundation model companies—DeepMind, Anthropic, OpenAI—love to talk about AI for drug discovery, and I think there’s fundamentally a mission orientation there. I’m also a bit of a cynic, because it’s hard to say that’s a bad idea. It seems broadly good for humanity if we have more medicines, as you said.

It’s like 10 years ago when the crypto people were saying, “It’s all international remittances,” right? Why do you think it’s both so popular with the labs, and even more popular over the last few months? And tell us about your partnership with Anthropic.

I go back to that returns-to-intelligence idea, where I think science is a problem that has some shape to it that really benefits from the LLM architecture. You think about the corpus of scientific literature as this vast pool of unstructured text.

These are pursuits where there’s a ton of domain knowledge to hold in your head, and there’s so much specialization. The idea that you could truly be standing on the shoulders of giants is very appealing. If I’m a scientist at an early-stage biotech, now I can have access to the world’s best clinical-design expert, the world’s best toxicologist, or a research assistant who can read papers better than me to figure things out.

There’s a lot about science that, again, is so artisanal and inefficient that it seems like a problem AI is going to be much better at. I think that’s one thing. The other is that biotech has big problems to be solved, and there’s an incredible—

Elad Gil

Because the failure rate is so bad.

The failure rate is so bad, and the impact is huge. Everyone has now seen what GLP-1s can do. Everyone saw what COVID vaccines can do.

Elad Gil

Magic.

Yeah. When it works, it’s magic, and people need this stuff. If AGI starts automating away software engineers or whatnot, what’s left? We’ve got to make drugs for people.

Elad Gil

All right. More scientists. And what about the partnership?

We have a partnership with Anthropic. We work with and use the capabilities of all the foundational model labs, but we found that Anthropic has a strong commitment to science. Dario’s a scientist, so there’s been really good mission alignment with them. They’ve expressed publicly that science is the next frontier after code.

For our customers, trust is super important. I think Anthropic’s posture, plus its technology, really appeals to them. From the start, Benchling and Claude have natively interoperated very well. Scientists can generate reports, ask questions, and do things like that from a very simple AI interface that they’re used to.

And I think it’s just the start with them.

Can we talk a little bit about company building? Just 13 years of wisdom in two-minute takes.

Every mistake made at this point.

Maybe we’ll start with the most recent hard decisions, not mistakes. Your co-founder, Ashu, gave up all his direct reports at some point and went all in on it. I called a bunch of friends around this company and our mutual friends. That’s a big decision. When that happened, how did you make the decision? You guys started way before AI was working at scale.

Yeah. It’s funny: we started early, but at the same time, I still feel late.

One of the interesting things is that I feel like the power of being a co-founder is actually just moral authority. This was a pretty controversial decision in our company, and we needed someone who had the willingness to say, A, “I don’t need any of my Legos anymore. Anyone else can have them,” and, B, “If I look stupid, that’s okay.”

It’s funny. He wrote this post that said, “I’m quitting my job to do this other thing.” I had a bunch of customers call me, and they were like, “Oh my God, I’m so sorry your co-founder quit. Is everything okay?” I was like, “Oh no, no—[laughter]—metaphorically.”

Metaphorically didn’t quit. Just going full-time on AI.

Yeah. It was controversial. Biotech obviously had this kind of bust, and a lot of our team was feeling like, “Hey, we’ve got to focus on the basics with our customers. The market’s tough right now. You have some companies that are laying people off or shutting down.”

How do you invest like that?

Yeah. Isn’t AI a distraction? But we were actually really fortunate. I think we had a good outside-the-building perspective. Ashu was very hands-on-keyboard himself, and that’s how he got convinced. He was playing with one of the models during Christmas or something like that, building for himself, and I think that really inspired us.

We also realized that if we don’t do this for our customers, who is going to do it? Again, it goes back to needing to translate some of these amazing things that come out of the Bay Area and Silicon Valley into useful vertical applications in very complex, regulated domains that we felt like we were the right people for.

Maybe because there are a bunch of entrepreneurs listening to this podcast as well who are looking at industries that are complex and regulated and want to try to bring the cloud and then AI to them—what has been hardest? What are some lessons from that?

My most trusted algorithm for this is: go talk to customers. I know that’s super obvious, but all the times when I feel like the company has been at its lowest or at its worst, or I’m feeling at my lowest, it’s because I’ve gotten too far from customers.

Product-market fit is a moving target, and sometimes your market changes. Our market changed: you had a bunch of biotech companies that were going to take over the world, and then they weren’t. I still spend—and maybe this is a vertical thing—probably 30%, 40%, 50% of my time talking to customers. I think of myself as the person who needs to go really deeply understand their problems and how they’re changing, and then bring that back to the company.

That’s the thing I have to role-model so the whole company does it. That’s the number-one piece of advice I’d give.

So that answers part of one question from another friend who used to work for you. He said that Sajith is amazing at understanding the macro of the company and then being deep in the detail on everything, especially with customers. Given that Benchling is a complicated company and biology is a complicated field, how do you decide what to focus on and then focus your team on?

I appreciate that’s a nice compliment, though. It’s like telling someone they have a large context window.

It is a good model, man.

Thank you. Thank you to whoever said that kind thing.

It was Malay. Hi, Malay. [laughter]

One of the most interesting things about being in a vertical is that you talk to all your customers, and they’re all underserved. In the world of life-science software, there aren’t many companies. It’s not like go-to-market tools, where there are 10,000 companies.

Much less great software companies.

Yeah, there are just not many. You have a very underserved demographic where they’re not used to someone coming and asking what they specifically need. They’re used to very general-purpose, horizontal software and trying to use productivity software to do kludgy things to make it work.

When you go talk to them, the first reaction is, “Oh my God, what we do is so unique, and we’re such a snowflake. There’s no way.” But then you talk to enough customers, and it turns out they almost all want the same thing.

I find that having 5 or 10 customers is actually a pretty representative model for what the entire industry needs. You don’t want to get too overweighted in one category of medicine or one size or something.

Everything we’ve built over time that has been successful is because we found a couple of customers, got very, very deep with them, and built and built until they were super happy. Then it took off. Maybe that’s the YC part that’s been programmed into me that I’ve never gotten out of: do things that don’t scale. I feel like in a vertical, that works 10x better.

Even for our new AI tools, they’re available to all of our customers now. Our team is on daily and weekly calls with 5 or 10 customers, to the point that a customer texted us yesterday about a meeting they had with the FDA. They ran a report inside our Deep Research tool, and instead of taking them a day, it took them 5 minutes and was perfect for them. That’s the level of closeness we get with our customers.

One other unique thing that I think would be useful for a lot of people today, including me, is that Benchling knows how to get scientists to work in a software company and work around a software company. I work with many more research scientists in different fields than I anticipated, let’s say, 5 years ago, when I was just doing good old engineering.

It is philosophically different, right? You have to run programs differently. You’re like, “Oh, it’s not like this is done by the next sprint.” It is: we do not know. So what advice do you have on recognizing that talent, getting them to be productive, and managing it?

That is a really hard question. I have serious battle scars from that. We’ve had to build a very interdisciplinary company to be successful. If I was only hiring software people who knew bio, I would have exhausted the pool 10 years ago. It doesn’t exist. We have to take the software people, take the science people, make them sit together, and learn from each other.

This is going to sound obvious, but I’ve actually found that the most conflict has been around how the mission gets solved. A lot of people coming from the world of science, especially academia, have a very different incentive structure. In the world of academia, your labor is basically free, so there is very, very cheap grad-student labor, and the currency is publishing a paper.

Yeah, and that’s how you get more funding to do more things and so forth.

Whereas in a company, we have to sell software. The most impactful thing has been a lot of repetition around the fact that, in order for us to achieve our mission and keep delivering great things to our customers, we have to make money.

A lot of the tension has come from the need to do that, so we have to make sure our scientific teams really understand that the better we do as a business, the more amazing innovation we can bring to our customers. By the way, if we don’t do this, who else is going to do it? Where are the next 10 companies building software for science and R&D that are going to power the next discoveries of these biotech and pharma companies?

I think we’re the only independent, scaled player doing this at this point.

Do you interview at all for this orientation?

I think we try to. I don’t know that we’ve had some amazing predictive way to find it.

I have a friend who’s a founder who asks people, including research scientists, how they feel about capitalism. I don’t find it to be a controversial question because it’s literally in my title, right? I’m a venture capitalist, but they think it’s a pretty interesting sorting function.

I will try that. [laughter] I’m not recommending it, but I do think the set of answers you get is interesting, actually, because I have tried it.

One more question for you on culture. You spend a lot of time with biotech and large pharma. You’re, at your core, a software person. What can these two worlds learn from each other?

Great question. I think the biotech and pharma world can learn something from how tech communicates and tells stories.

There’s the whole AI wave in tech right now that really resonates with me as a founder. I think biotech and pharma companies need to tell their stories. Most people, I don’t think, could name 5 scientists or 5 CEOs of pharma companies, but we could name every single tech CEO. Everyone knows who Sam is, and it’s on a first-name basis, right? For a lot of them—Sam, or Jensen, or something like that—there’s a lot of hero’s journey in tech.

A lot of hero’s journey in tech.

But I think talking about the patients and scientists would change a lot of the public perception of biotech and science. I don’t think people know how hard it is to make a medicine. I was just thinking recently: you look at COVID—even Moderna and Pfizer helped the world. Whatever you feel about vaccines, they played an important part in helping the world reopen, and yet Zoom gets more credit in COVID. Where you have Gilead, HIV used to be a death sentence in the 1980s, and there was tons of panic and fear about it in the ’90s and early 2000s, and—

Gilead has basically cured HIV at this point, and most people have no idea. Because of the way they communicate, it’s much more about these faceless companies rather than the people. I think it’s easy to hate on them and underappreciate them. So I think they need to tell their story and go direct.

That’s one thing I think tech can learn from bio. I think tech has sort of become everything, and the ambitions of tech have grown a ton. I do think some of these other industries have figured something out when it comes to rigor, validity, and accuracy. “Move fast and break things” does work for certain domains, but once you get to the point where you want to have credibility with regulators, put things in patients, or make a medicine, that stuff does matter.

Biopharma, again, for all the things that are difficult with this, has figured out how to be safe. The U.S. is still the gold standard for how to deliver medicines safely to people. I think that stuff matters more and more.

Yeah, it’s really interesting. I was talking to a friend who’s a top research scientist at one of the top labs, and I asked them a very basic question probably a year and a half ago: How can I better use these models in very quality-sensitive fields? And they said, “Just wait,” essentially.

This is an amazing research scientist, but I’m like, that seems unambitious, right? There are many benefits to intelligence in these fields that are really important to all of us, and so I do feel like more people should work with that too.

Yeah, yeah, yeah. I think in Silicon Valley sometimes it’s a little bit easy to oversimplify other people’s jobs. There’s a lot of complexity out there in the world of biotech and pharma, but I’m very optimistic that tech companies will figure out how to make it work because I think everyone is very excited about AI.

Last question for you: Something you’re excited about in AI outside of bio, or outside of Benchling’s immediate purview, I guess?

On a personal level, I haven’t written a line of code. I had not written a line of code in probably—I’m very embarrassed to say this—maybe 8 or 9 years. It’s been some time. I remember I stopped coding around the time that React started being a thing. I’m dating myself. All the kids listening are going to turn off.

I’ve tried some of the new agentic coding tools lately, and it’s just, on a personal level, fun to feel the whimsy of being able to build something very quickly again. That’s pretty cool. That’s probably the one I’m most excited about.

And then I’m excited for my parents, actually, and for them to have new technology that’s pretty easy for them to access. My mom’s a ChatGPT user, and I’m sure it’s sort of Google Search++ at this point, but that’s been pretty cool too: technology that’s so intuitive that I don’t have to go home and be IT support at Christmas.

Yeah, I think it’s actually undervalued because there are audiences that are traditionally not as lucrative as the fast-tech-adopting audience of 15- to 40-year-olds, but it is incredibly wild how the UX of natural language and voice are—

—changing who can use it, right? AI tutoring for my kids and such, and the same experience—

Learning is such a joy with AI. I hope I’m learning the right things. Thanks so much for doing this, Sajith.

Thanks, Sarah.