Max Hodak: How to Build a Startup That Moves Fast
Science's CEO on purchasing systems, hiring filters, and why the boring internals decide how far you can take a company.
Science is building a retinal implant that restores vision to people who have gone blind. One patient has already used it to read a 300-page novel.
Building a company like that requires a lot more than getting the technology right. At Startup School 2026, Science CEO Max Hodak explains how the company buys things and hires people, and why those systems determine how fast it can move.
He also gets into why founders can’t delegate their judgment, and why there’s no set of five bullet points that makes a startup work.
Timestamps
00:00 — Infrastructure at Startups
01:02 — Science’s Retinal Implant
02:41 — The Hidden Infrastructure of a Startup
03:25 — How Your 17th Employee Buys Things
06:15 — How Infrastructure Creates Speed
07:03 — What Does an Experiment Actually Cost?
09:27 — How the Best Startups Hire
10:18 — Building a Rigorous Hiring Process
13:26 — Judgment, Horsepower, and Agency
16:17 — Rethinking Performance Reviews
18:53 — Rate of Iteration Separates Success From Failure
20:43 — You Can’t Delegate Your Judgment
22:47 — Action Produces Information
24:31 — The Operating System of a Company
25:10 — Q&A
Transcript
My name is Max Hodak. I’m the CEO of a company called Science. We’re going to talk a little bit about infrastructure at startups. I’ve spent most of my life working on brain-computer interfaces. This is almost 20 years ago now. I started my career as an undergrad working in a lab at Duke. This is from our very first Society for Neuroscience conference. The experiment I was working on back then was: if you put electrodes in the brain of a monkey and then give a monkey a joystick and you record the neural activity as it’s playing a game, if you make the joystick—say, the cursor goes sideways when you push forward on the joystick—what does the brain do? Does the brain represent the joystick or the screen or something else? It turns out that there are neurons that do both.
At our company, Science, our main product is a retinal prosthesis. It’s a chip that’s implanted under the retina in the back of the eye to restore vision to patients that have gone blind due to loss of the rods and cones in their eye. This is one of our patients on the cover of Time last November. On the right, you can see there’s a picture of the implant with the glasses. So every little one of those hex grids that you see on the implant is essentially a solar cell. When this is implanted under the retina, the patient wears glasses that have a camera that sees the world and a laser projector that projects onto the implant. When it projects the image in infrared, wherever the light is absorbed on the implant, it creates a little electric field to excite the retina, thereby directly bypassing the dead rods and cones to stimulate this visual signal back into the retina, the first possible opportunity.
And this is a pretty cool product. It finished major clinical trials last year. It’s been in three clinical trials now. It was covered in the BBC last fall. One of our patients finished a 300-page novel with the device and mailed us the book. But I’m not going to talk about this work for the most part for the next 30 minutes. We’re going to talk about infrastructure and lessons. This is Startup School. Maybe there are some things that you’ll find useful in your company.
Picasso was noted for saying that when art critics get together, they talk about form and structure and meaning. And when artists get together, they talk about where to buy cheap turpentine. This is also often phrased as: amateurs talk strategy, professionals talk logistics—a quote from a guy that the United States named a tank after. And so there are surprisingly few lessons that are really broad across companies.
Typically, the experience of running a startup is you’re just looking at a continual stream of facts that hit your desk every day, and you’re trying to make the best local decision that you can for those facts. And if it looks inconsistent over weeks, that’s usually the way to go. But there are a couple topics that keep very repeatedly coming up that are universal experiences, at least for deep tech companies, which is the thing that I know most of my experiences in—not just pure software. There are things that keep coming up, like buying things. Your first reaction might be that if you do software, you don’t need to buy things. It will be me alone in an empty room with some computers writing software, and this is going to be how we build a company.
And if this is you, yes, you have figured out a reason why VCs love funding software and why they’ve done so much of it for the last 25 years. But if you do anything other than pure software, you’ll be buying many, many thousands of things. This is us about, I think, six months into the company. It’s a little tough to make out. There are a lot of computers. There’s also a bunch of microscopes and other electronics and 3D printers and resin and PCBs. You’re buying things really continuously. It might sound really obvious, like a really basic question, like you surely just buy things. For you as the founder, you can use a credit card. Credit cards work great. You can buy lots of things with credit cards. You can also send a wire transfer. The question is, how does your 17th employee buy things?
Do they have a credit card? Let’s say you hand out credit cards to all of your employees and tell them to buy what they need. So you start getting messages like this, and you think, I care about burn. We have to spend efficiently. I’m going to approve all of the purchases as they happen. You’re going to put a message like this, and then you think $3,000 sounds like a lot for a power supply. Do we need a $3,000 power supply? What if we get one from an auction? In three days, there’s an auction. Maybe we’ll get it for half off. We can get it in two weeks. But then you also remember that you’ve hired some very highly paid and talented employees. Are you saying they can’t get the tools that they need? You’re spending $100,000 a week. If you wait a week to get a power supply half off, you have certainly dwarfed any possible benefit from getting it.
And if they were at Anthropic, they’re not going to be getting hassled over a $3,000 purchase. They’re just going to have a power supply. What you realize is not only is this very hard to keep burn under control, but also this is the inappropriate place to exercise spending review. Spending review has to come earlier. You have to have some concept of budgeting. It’s not really even just about the payment rail of buying a thing. It’s about how you understand the bucket of money that you have. I don’t want to be making the $1,500 power supply versus $3,000 power supply trade-off. They need to understand the resources that they have so that they can make trade-offs within those available resources. So you set up a procurement system, and now your highly paid employees are spending their days clicking around B2B enterprise SaaS.
It turns out that from the time they place an order for a power supply, it takes two weeks to arrive because you can’t actually buy that with a credit card. You have to set up an account with the vendor and deal with insurance and certification paperwork and get an account set up. They have to generate a quote so that you can generate a purchase order so that you can generate an invoice. Now everyone’s upset that things are taking super long to get ordered. This actually really requires—this is a living organism. When people move from academia to startups, I think one of the reactions that people often have is, why are there people whose job is to purchase things? Surely I can just buy things. But absolutely there are people whose job is to buy things.
From the time that you submit the order, going back and forth with the vendor to set all of this up is very time-consuming and it can easily stretch out. It takes active management and metrics to cause these things to go fast. I think part of why, when we think about this, we have a reputation of often being very quick and people are unsure—how does that happen? It is mostly not that we are smarter. It is infrastructure like this; that is how speed is built. Now people can buy things. At least you can keep overall burn under control. Now you know that you’re not going to exceed some large amount of spending every month. Then you realize that that wasn’t really the problem. You could figure out your runway. The problem is attribution. When you’re doing—whether you’re working on rockets or cars or drugs or brain-computer interfaces or anything that involves dealing with the real world—you realize that one of your other problems is that you’re buying stuff in bulk.
We buy gases from argon to xylene to nitrogen, to resins, to media. We buy these things in bulk and then part them out to lots of different experiments. When you do this, attribution is pretty difficult. If nobody knows how much an experiment costs—like every time you grow up a new cell line or every time we make a new probe in the fab—how much does that loop cost? Nobody knows. Therefore, experiments are free. It doesn’t cost dollars; it costs media, and media comes from the fridge. And we want to know, how do we price a thing that we make? We make a bunch of things in volume in the foundry. We want to know, what can we sell that for? That requires all of these spreadsheets to get an estimate of the pricing.
And there are opinions in here. These are not all facts. How much do you include rent? How much do you include depreciation of the tools? This comes with opinions about your future volume. All of this is required to understand not just what you should charge, but also what you’re spending and what your runway is. To deal with this, we’ve built a huge amount of internal software at the company for managing this. One of the first things that we did is almost everything that you can do in the company is a button somewhere in the software. We call it Helix, including stuff like purchasing. Because this extends all the way through to manufacturing, where we have every step that happens in the lab in the database, we can correlate all of this through and get this information. It turns out that for every iteration of a wafer that we make, in this case, for this protocol, it costs $40,000.
This is a lot of money. You might have raised—let’s say you raised $20 million in a Series A. You think you need four years, you need 20 people. In my experience, about half the burn is headcount. So let’s say 20 people, that’s probably three, three and a half—that’s like $3 million a year in payroll. That’s half of your burn. You need 20,000 square feet, about $4 a square foot. That’s another $750,000 to a million a year. So now suddenly you’ve got really, it’s a $3 million a year research budget for three or four years. That goes way faster than you think. But your team just saw that you raised a larger amount of money than they’ve ever seen in their lives, and they think that the $3,000 power supplies are free. This is pretty important. This infrastructure actually determines success or failure in many companies.
Another universal experience is hiring. Hiring also, I think, really separates the successes from the failures. Startups usually don’t come out of nowhere. I think the best companies, in my experience, come from what might be characterized as scenes. There’s a moment that enables a new company to be born, and there’s a bunch that comes together that really creates this unique nucleation for the new company. Once that moment has passed, because some company has executed on it or just the time has gone, it’s tough to get back. The best hiring comes from within your network—people that you’ve worked with before, who you know are good. The extended version of that is to hire from the network that produced the startup. There’s usually some extended scene that the thing came out of. There’s a bunch of co-founders that come together and crystallize out of that.
But then there’s an extended community, and that should really be the target of your initial marketing. These are the people that already speak your language, are already familiar with it. But there’s never enough of them to really fill an entire company. You have to hire from the general public. Different companies hire in different ways. There are different processes that make sense to different founders. This is the thing that really is going to be matched to who the founders are and how they view the world. There’s no one right answer. But this is a thing where you need a really defined process. There’s no right answer, but a wrong answer for sure is not having something that you do very religiously as a company. This is an area where reality has a surprising amount of detail. It seems really straightforward, like, oh, you’ll have a job board, you’ll get applications, you’ll review them.
This very quickly becomes a huge, huge drag on the rest of your team. You can easily spend almost all of your time recruiting if you’re not doing it efficiently, to get to a suboptimal outcome. For us, again, we’ve built a lot of software to do this. There are four steps to our process. The first is that we’ve built a software interface for applicants to apply online, where we can capture some structured information from them upfront, including the ability to apply to multiple jobs in parallel. We originally used a commercial applicant tracking system, but we’ve moved this to our internal tools. One of the reasons we did that is because this allowed us to do something that we couldn’t find in any of the commercial ATSs. The first step of our process when users apply is it goes to company-wide voting.
This is a heavily redacted version of the internal interface, but hopefully you can make out the idea of what’s going on here. The applicant’s resume is in the middle, we collect a little bit of other structured information. But the most important thing is on the far right, you see there’s a question: How would you vote for this candidate? Are they Known Good? Strong Yes? Yes? No? Strong No? When a person applies, the system picks out seven or eight current employees that it thinks look something like their backgrounds and it pings them all for votes. We can distribute the voting across a lot of the company for this initial review, which is essential because if you’re doing anything cool, by the time you get a couple years into it, that top of funnel is overwhelming. If you place any small group of employees or any one person in the way as a bottleneck on this, they will absolutely bottleneck the whole rest of the organization.
You also want to average over judgment. There are different people who are better or worse at hiring and have different perspectives on what you’re looking for at that stage. In the beginning, as the founder, you can meet with everybody and that will take you quite far. You should definitely interview everybody for quite a while. But even beyond that, you want ways to average over the judgment of the rest of your team. Voting mechanisms are usually a really good way to do that. These are our actual statistics over the last couple years. Seventeen percent of the top of funnel applications that we get go to a phone screen. That first initial voting stage is drawn from a company-wide pool so that we can get the voting done quickly, usually within 24 or 48 hours, and not bottleneck that on any small group of people.
The phone screen is again drawn from a company-wide pool of people. This is not team specific. This is a company-wide bar, really looking for three things: judgment, horsepower, and agency. If we throw you into a complex, vaguely defined situation, will you tend to make good decisions or will you create diplomatic incidents? Do you meet just a basic hurdle for technical competence and demonstrated ability to learn things? And are you effective at causing the world to look like you wish it were? How does your life look or not look like whatever ambitions you had? And do you have specific ambitions for your life? This we can distribute over the entire company. Then half of those tend to go to homework. Ideally, we’d be using entirely AI-resistant homeworks now. Our favorite types of homeworks are things that don’t saturate, have a very high ceiling, and are naturally scorable to two or three numbers that we can put on a plot so that when we get responses to homeworks, we can plot them all.
And then it’s very obvious when someone has really beaten the Pareto frontier, and we otherwise don’t care whatever AI models they use—that can make you better. In cases where that’s not possible for homework right now, we’re doing an increasing number of technical phone calls or onsite practical tests. But ideally, we would have an AI-resistant take-home for each of these. Anthropic had a really interesting take on the AI-resistant homework, where they’ve had a couple of tasks, like the GPU kernel optimization: what is the minimum number of cycles you can get it down to? And this is naturally adjusting. The hurdle for a while was, I think, it was Sonnet’s performance. If you could beat that, then you could get an interview. I think that there’s a bunch of ways to construct AI-resistant homeworks. And then, by the time you get to the interview, it is important that you have a, from there, reasonably high—like at least 25%—conversion to an offer, because otherwise you’re going to waste too much of your time doing onsites for employees that don’t convert. You can’t get that down.
And so there’s four steps to this: initial voting, the phone screen, homework, and a full interview. And this is, as far as from my experience, this is the minimum set of information that we need to make a full decision. And I don’t think that there’s a more efficient way to elicit this. I don’t think there’s a smaller number of steps that we could use. So this has become our process. So you’re hiring people, they’re coming into work, they’re starting, they’re incurring payroll. But how do you know that you’re good at this? Eventually, you’ll get feedback from the market on how good you are at hiring, because the company will work or it won’t. Your team will be capable of accomplishing the stuff that you’ve set out, and they’ll help you course correct through that. But this is a very, very long feedback, and it’s a very poorly behaved loss function.
And so it’s your job as management to design synthetic gradients that allow you to find out earlier and along the way how recruiting is going and if you need a course correction.
A conventional answer to this is the 360 review process. So once a year, you send out a lot of forms, you gather up a bunch of feedback around each employee. You set up a bunch of meetings with HR and with the various managers and you do the conventional performance review cycle, which, based on my experiences, is a very disruptive process that doesn’t tend to surface issues that you don’t already know about, but haven’t acted on because you knew that thing was there, but firing people is hard. So people drag their feet on it. This is kind of reinforcing things you already knew. And it only happens once a year, maybe twice a year if you split up the company into cohorts. But really what would be nice to have is a signal that gives you this kind of natural feedback from across the company about who’s good and who isn’t and what’s working and what’s not in a way that is largely unbiased and is more continuous.
Imagine if you could get feedback every few weeks on where there are issues and where things are going well. The process that I developed, which I’ve now used for the last six or seven years, is every couple of weeks, every four to six weeks—it’s not that often—people in the company get pinged with a question through the software, through Helix. There’s a form, but really there’s only one question that really matters, which is: knowing how this person turned out, would you vote again today for their hire? It’s the same question we use on the initial voting. You’ll get a prompt to say, “This person you work with, how would you vote for their hire today?” Then what we can do is construct a graph over the company of all of the feedback. The basic intuition is that your vote should be weighted more highly if everybody else has rated you highly.
The astute may notice that this looks a lot like the original Google algorithm, PageRank, which is an idea called eigenvector centrality, where you can create a weight over the graph by looking at how the graph points together. This is a little bit different than literally eigenvector centrality, but it’s very similar. We call this technique eigen reviews. I’ve become convinced that this is more or less the right way to do performance reviews. There are some other tricks that you have to apply to get this to work really well. For example, we apply dropout, where we’ll run a thousand iterations where we’ll randomly remove some percentage of the edges each iteration. When you look at the distribution of scores that you get out of that, if you see additional peaks, for example, this is a clue that there could be voting cliques that need further investigation.
But as a whole, it distributes the judgment across the company, updates more or less continuously with about a month lag, and gives you just way better insight into what’s going on really around the company.
And it also totally gets rid of that traumatic, super heavy, once-a-year, HR-driven performance review process. So the point of this talk is not that you should use this in particular, although you should consider it. And if you actually roll this out at your company, you can email me and I’ll send you a doc with more specific tricks on how to actually get this to work well. But the real theme of the talk is that rate of iteration separates success from failure. And if you can get a fast iteration loop, that really overcomes many other things you’re going to run into. And this effect is so severe. If you can learn one thing every week and there’s a competitor that’s learning a thing every month, they will never matter. Overwhelmingly, if you’re looking at two different approaches to solve a problem, if there’s one that allows you to compound in a much shorter amount of time than the other, even if the other approach has significant redeeming characteristics, you should really consider going with the shorter iteration cycle because the compounding effect is just so dramatic.
And so speed determines success and failure. And speed is determined by infrastructure. This is driven by really boring-sounding things like how well your purchasing and recruiting and spending processes work. This is as important as how well you understand the object-level technical content of the thing that you’re building. I see companies founded by stellar pedigree scientists and engineers all the time that die on the vine because this execution is tough to follow through and your job is to organize. It’s uncommon that these deep tech companies fail because the technology doesn’t work. They fail because once you end up with this organization of hundreds of people and hundreds of thousands of square feet of physical infrastructure, you haven’t built the systems to manage that and it becomes unwieldy.
And then you can’t connect strategy to execution.
So we heavily lean towards things that have shorter iteration cycles, kind of set all else equal. But that’s not a blanket rule. There are no blanket rules in startups. You’re looking at each new fact pattern that comes in as its own unique thing and then making decisions that make sense to you. And one of the harder lessons as a startup founder, one of the harder things to really deal with is the fact that you cannot delegate your judgment. As the CEO, you must always make decisions that make sense to you no matter how much momentum or inertia alternatives seem to have.
So in school, let’s say there’s somebody sitting next to you and you cheat on the test by looking over at them. All else equal, your grade will be dragged towards the average of the class. That is not good enough to succeed in startups. You have to do things that are at the long tail. The successful companies are the exceptions by becoming an average that is not good enough. And so in order to succeed, your judgment has to be differentiatedly good. Now the reality might be that you don’t know if your judgment is good yet. And so one way or another, you will have to find out. And that means making decisions that make sense to you even when you are totally alone in that realization. That is the only way to get to the really big outcomes. Now it’s not that often that everyone else will think one thing and you’ll be like, “You’re all totally wrong.” But it is a really eerie feeling. You’ll get to a point four or five years into the company when there’s hundreds of millions of dollars on the line and there’s some really high stakes decision and only you can make it.
And then you will look around for advice because in the beginning you’ll get lots of it. There are a bunch of things that are easily advised or easily figured out, but you’ll get to a key point years in and you’ll look for advice and there is nobody to ask. And at that point, you must have a really good sense of the limits and boundaries of your judgment. That is a very eerie feeling and you have to be able to commit to it regardless. Now, the good news is that in my experience, it’s very difficult to actually get stuck. You can get yourself into trouble and the action space is always larger than it appears.
No matter what happens, when you get there, I think it’s very easy to try and anticipate all kinds of problems that you’ll never actually run into. And then you go and do it and then you get to a point where the system, like you’ve run into some real limitation. There’s always a hundred ideas about how to make it better. This is sometimes phrased as action produces information. This idea is, I think, much deeper than it sounds. So in physics, there’s this quantity called action. And so if I throw a ball and it follows a parabolic trajectory, that trajectory is totally set when it leaves my hand, unless it gets blown by wind, some other action is exerted on it. It will follow this ballistic trajectory, which is in this sense kind of an information minimizing trajectory. I can say it just followed.
It was ballistic—that totally determines it. If something else happens, you had to spend some energy and time to cause that to happen. And so whenever you exert action under the universe, that creates information, in a very fundamental sense. And whenever you get stuck, you have to start injecting action, producing entropy. This produces some fairly counterintuitive effects. I’ve seen situations where the company is stuck in a deep local minimum and there’s someone who is great in many ways, but is just the wrong fit for what that company is at the time. Removing them, even though they individually are very strong, unblocks the company and allows it to enter a new phase. When you get stuck, you have to start doing things. The thing underneath the object-level content of what the product you are building is, is all these support systems.
How the company does purchasing and accounting and recruiting and performance reviews and budgeting and safety and quality is the operating system of the company. That has a huge impact on how far you can take it. Speed is determined by infrastructure. Speed determines success and failure. You need to put more thought into getting these foundations right. If you do them right at the beginning, everything else is much easier. If you get them wrong, you’ll end up spending $5 million a month and feel like you have very little control over it. Then you’re forced into coarser levers and harder decisions.
Thank you for coming to my TED Talk.
Q&A
Do you have advice for people trying to choose between industry and academia — starting a company now versus getting a PhD first?
It really depends on specifically what you’re doing. If your field only exists in basic research, then getting a PhD might be very reasonable. When things really start to work—like, if 20 years ago, the best computer scientists were at CMU and Harvard, and 50 years ago, if you wanted to work on rocket engines, you were at NASA, you were at a university, you were at University of Maryland or somewhere. Now, the best computer scientists are at Google and Apple and OpenAI. The best rocket scientists are at SpaceX and Blue Origin and others. So when a field really starts to work, industry can just marshal such larger levels of resources and can just move so much faster.
And so I think the question has been, why has academia stayed so relevant in the life sciences? The reality is that it doesn’t work that well for most things. Humans just aren’t that good at drug discovery. If your field is really only in academia, then it can make total sense to get a PhD. But I think a lot of—it is uncommon that startups don’t get the technology to work. It is more common that they can’t organize the human organizations to accomplish their goals. And that is also a skill set. The only way to learn it, I think, is an oral tradition. You have to do it. So if the choice is working at a really high-performing company adjacent to where you want to be versus getting a PhD, I’d probably recommend the company, but it’s not an absolute rule and it really depends on the field.
What counts as evidence of exceptional ability to you?
Anything you can concretely put your finger on that separates that person from their high school class. If you have your average high school student, we just want some concrete fact—ideally, the best evidence of exceptional ability is winning legible competitive games. So this could be being a chess grandmaster. It could be winning Design/Build/Fly or Formula SAE competitions. There’s a bunch of Silicon Valley deep tech companies that are basically built out of Formula SAE winners from college—people that just spent their college experience building things and racing them and finding out. I think you have to have that type of competitive feedback. It is tough to know if you’re exceptional without having some legible competitive game.
How do we hire engineers now? Do we still use LeetCode, or do we have better ways? If we allow AI use, how do you understand the skills of the applicant?
I don’t think we’ve ever really used LeetCode. Maybe some other people on the team do it in secret, but I’ve never asked it. So software in particular, the rewards to horsepower are so great that it really is just—it’s a field that attracts really smart people because it gives you this very rapid feedback. There are a lot of really smart people in biology, but when you have a biological idea, it can take you many months to find out if it’s a good one. In software, if you have an idea, you can often build it in a couple hours or you can get feedback within days. So it has this really addictive feedback loop, kind of like high-frequency trading, that just draws in really smart people.
And so we look for, over your life, what signals do we have that you have done something interesting? It’s uncommon for someone to get into their mid-20s without there being some thing in their background that they went out and sought out and did.
But this is such an open-ended criterion. It can really be anything. More directly to the question, we increasingly don’t directly evaluate programming. We try to evaluate thinking. So these are design questions. If we give you a domain, how do you break it down? Can you understand the decomposition of the problem clearly? It’s really a measure of, can you think clearly rather than can you write code?
What did you take away from your experience at Neuralink?
So the question of, should you go get a PhD? I don’t have a PhD. I spent five years running a company for my CEO at Neuralink. That was—one of the biggest lessons, I think, is that there are few really generic answers.
There’s no generic algorithm for how to succeed at a startup. There’s no set of five bullet points that can be conveyed that, if you just turn the crank, your company will be successful. It’s a long series of judgment calls. And so the most important thing is that those filters are tuned really well. I think one of the most valuable things for me at Neuralink was I was working with someone who has empirically excellent judgment. We could get into trouble together and something would happen and there’d be two possible solutions that would make sense. I’d go to him and say, “Is it option A or is it option B?” He’d look at it and be like, “Oh, it’s definitely option B. The problem would never recur.” Having been in those situations where I was trying to make these bets with stakes attached, looking forward in time, not getting feedback until later, with that advice was incredibly useful for fitting those filters.
I don’t know that there was really a shortcut. I think that just hearing the stories when you’re not there, really thinking about it because there are real stakes, and then getting that feedback—that is an essential part of the education of an entrepreneur that I think many people underwrite. I think it is really worth working for a company that has an excellent culture that you respect before jumping right into your own startup. It is relatively uncommon that startup cultures get rediscovered entirely from first principles. Usually they’re passed down as oral traditions because there’s a founding team that worked at another company, which worked at another company, and so they inherited it. Or in some cases where there’s really a breakout, where there’s just some market dislocation that really enables a team out of nowhere to build it. They’ll often get it from the VCs, but it’s working with the people that have that judgment so that you can get that reinforcement learning as it’s a long series of facts that is really important.
Could BCIs or neural interfaces help us figure out what consciousness actually is? How?
Absolutely. So if the end of the artificial intelligence quest is super intelligent machines, I think the end of the BCI quest is conscious machines. The brain is composed of ordinary matter arranged according to the rules of chemistry, only things found on the periodic table. It seems tough to believe that there’s some new physics going on in there. And so we’re looking for some mapping between the substrate activity and the phenomenal content. Now, if we had a magical BCI that allowed me to see the instant state of every neuron in the brain and drive them, I think we’d figure out consciousness pretty fast. I think this is a practical problem, not a philosophical problem.
But to prove it—first of all, that practical problem is real and we’ll have to do this stuff in humans. And to prove any of this, we’ll have to do it in humans. I think that it is possible that you could use a BCI to prove it. We have some ideas about how to do those experiments, but they’re still some number of years off. Things going into humans now are not designed to study consciousness. But I do think that that is further down this path.
What should I study to contribute to BCIs?
This really depends on your background. Neural interfaces are a very interdisciplinary problem. It uses everything from stem cell biology to materials and microfabrication, to software, to animal behavior, to surgery. So there are many different entry points in it.
One of the things that we found is that it’s better to have a smaller team that can fit more of the problem in their heads and then compress it together. Contrast this to how academia usually handles interdisciplinary problems, where they’ll have an interdisciplinary center that pulls in very deep verticalized experts who kind of meet at the center. The problem is that they’re all speaking different languages. And so it’s often hard to really—even when they can communicate, typically you end up shipping the interfaces of those departments. Whereas for us, if we can hold the problem in the head of a smaller number of people, we can shift around where the bottlenecks are.
A specific example of this is our protein engineering group has been able to develop much more sensitive, much better proteins for some things that we need to do, which has allowed us to relax some electronics requirements. Specifically, we have proteins called opsins. They allow us to make neurons light sensitive so that if we shine light on them, we can fire a neuron. The problem was that you needed to hit a neuron with a lot of light to fire it, which means that you can’t have that many light sources because it gets too hot. We’ve been able to make the protein more sensitive, which means that we can have more LEDs because each one can be dimmer. We’ve turned this electronics problem into a biology problem that allowed us to relax those constraints. You don’t get that as much when you have these interdisciplinary centers where there’s one group focused on one thing, and another group focused on another thing.
I would say being able to have a broader perspective of more of the problem is really valuable. And then just really as deep and clear an understanding of the system as you can get. I think there’s no substitute for being hands on. It doesn’t really matter. You want some hard skill to get you in the door—software, electronics, mechanical, materials, something. And then from there, I would try to learn as much of it as you can.
What doesn’t AI replace in scientific research? Where are humans still necessary, if anywhere?
We still definitely need humans. And in scientific research in particular, it’s tough to predict. AI is clearly advancing very rapidly. I do think that you need to think about how to have your company be AI native in the sense that you want to gather all of the context, all of the stuff happening in your company, and be able to make that available efficiently to agents because those are clearly a big part of the future. For us in Helix, really everything goes in there. One of the reasons that we did that was because not just is it powerful to have everything in one database to link together—purchasing to quality, to batch records and manufacturing—so that we can trace stuff more efficiently, but also so that we could give it all to agents.
We found them to be a multiplier for the team, not a replacement.
The three biggest areas that AI has had an impact for us so far are, well, first of all, coding. That’s now basically all this. I’ve written a lot of code in my life. I don’t think I’ve looked at the source very much in the last six months. That is getting really good. Regulations. So if you’re doing anything really interesting, you’re going to end up regulated and then you’ll probably end up dealing with these things called quality systems. A quality system, I think, triggers a lot of scar tissue for people because it’s just the quintessential heavy bureaucracy that slows everything down. But the idea of quality itself is actually not a problem. The problem is that humans are bad at reading and interpreting these things. And so when we make a product, one of the things we have to do is identify all of the standards that might apply.
And there are standards for everything. There are standards for how the lithium-ion batteries plug into a PCB. There are standards for electrical insulation of the boards. There are standards for shipping labels. At some point, you’ll have to take your shipping packaging, print a label on it, and put it in a vibe box and show that the corners of the label don’t curl in a way that might cause it to detach. And so you hire regulatory experts to go find all of the standards that might apply, make a list of them, and then have a spreadsheet, which is all of the evidence that you comply with all of the standards. So this thing can take many, many months historically. AI has totally transformed it. We can very quickly look up all the standards. We can very quickly generate the evidence tables. And I think that to the degree that there’s kind of over—
We definitely need to deregulate some things, but I think that the combination of AI and regulation is a better fit than people think. You can use it to smooth a lot of stuff where the regulations are written in blood and are largely good ideas. It’s just hard for humans to do it.
Why build your own infrastructure platforms rather than just buying them?
You can’t really buy these things. There are ERP systems out there, but there’s no company that loves their ERP system. I don’t know if there’s anyone who’s really like, “I want to spend more time in NetSuite.” On the contrary, there are a bunch of examples now of companies that grow up around a piece of software that’s really fit just for them. YC famously has a lot of internal software that really makes YC work. Facebook also very famously invested heavily in internal tools and now gets a lot of efficiency from that. SpaceX and Tesla internally have a pretty giant piece of software called Warp Speed that runs a lot of their manufacturing and R&D processes. So when one company grows up around a harness fit to it, it can be very powerful.
It is powerful in a way that the software you can buy isn’t. But this requires you to really look into the future because, certainly, especially at the seed stage, this is not the thing that you would think you should be focusing on. And historically, it has not been. I think this is a thing that has changed with agents. The fact that you can vibe code this now makes it a reasonable thing to think about. Historically, software has been so expensive, you would have had to buy it. And that’s what everybody did for a long time. That was, I think, a worse world, and that world has changed. And so now there are better options available. But like I said, we previously had used— we used Greenhouse. Greenhouse required us to have a small number of people as a bottleneck at that first funnel stage.
Replacing that with software, we were able to explore voting mechanisms and fairly detailed voting mechanisms that can make smart inferences about who would know about an applicant—things that you can’t really do with the commercial software. And so for a lot of these processes, you should think about how you want it to work for you.
These are human organizations, these human processes that have to be staffed. And if they aren’t done routinely, they’ll atrophy. And there are things that make sense for different teams and founders in the way that they view the world and think about it. It really is all very different. But if you build a thing that works for you and then you bake that into the company so when you put something there, it stays there, it can be very, very useful.
What changes should we expect in the world as BCIs start to work and get widely adopted? Do intelligence differences no longer matter?
So there’s this meme that BCI is an artificial intelligence-adjacent story. And there’s some of that. Eventually, if AI is building super intelligent machines and the BCI labs are building conscious machines and we’re building brain-to-brain connections so that the boundaries between those things become less meaningful. At some point, you want a super intelligent conscious machine that we can participate in, but that actually feels further away to me. I think in the near term, BCI is really a longevity story. And I view longevity as really just healthcare. Just biotech. It’s just that it hasn’t— I think it is not right to say that the pharma companies or any of these past healthcare companies are not interested in cures. I think that is what all of them want. It’s just that that’s been beyond our capabilities. And in neural engineering, when people hear BCI, I think they think of motor decoding.
I put some electrodes in motor cortex and now they can control it like a video game. I think neural engineering is much broader than that. We include our retinal prosthesis in that. We include cochlear implants in that. And this, I think, is a contrarian take on all of healthcare. It gives you these effect sizes that you just don’t really see in medicine. If you have a patient on a dopaminergic drug for Parkinson’s, that works for some period of time, but it’s a relatively small effect after a little while. You try on a deep brain stimulator and a patient goes from not being able to hold a cup of water to being able to write cursive in 10 seconds.
If you want to talk about strong patient testimonials, you should see a newborn having their cochlear implant turned on. When you deal directly with the brain as a computer, not only do you not have to solve some of these really hard biology problems that are just beyond humanity’s capabilities, but you get these results pretty readily that, again, are just like you can get an engineering gradient, you can get them more reliably, and they’re just large effects. And so I see this as a way to extend and improve the life of everybody. The brain is the thing that makes you you. It’s the only thing that in principle you can’t transplant. You can get a new heart or a new liver. You cannot even in principle get a new brain. And the brain is usually not the thing that fails. And so if you can deal with the brain directly, I think this is more of a radical longevity story than it is an AI one for the moment, although all of these things will come together over some period of time.
For your Eigen Review performance system, how do you prevent employees from colluding on their votes or downvoting somebody on purpose?
So as I mentioned, there are some tricks. For example, applying Markov chain Monte Carlo dropout allows us to detect things like voting cliques because now instead of seeing one peak, you’ll see two peaks. That is a clue to look in. I look into that. It’s designed to be tolerant of these things. I think it is really fairly transparent. It’s also not our only signal. It’s one of several. If anybody’s interested in this, send me an email and I will share a document with the specific tricks, but I want to understand a little more about how you were going to deploy it first. Some of this is tradecraft.
When you’re building something as long horizon as neurotech, how do you figure out how much runway you actually need to keep the company alive? And how do you get investors to fund that much?
Sometimes you see founders, especially more inexperienced ones, pitching VCs for what they think is reasonable to ask for rather than what they need to run the experiment. You’re raising some amount of money to go find out some answer. The answer to that might be no; the investors understand this depending on what business you’re in. But you have to actually run the experiment.
There are definitely some ideas that are worth funding with $50 million or $0, but not $5 million. You won’t run the experiment. It’ll be a really frustrating experience. You’ll get an ambiguous outcome. My first piece of advice is you should figure out what you think it’s going to take to actually run the experiment, which is not the whole company. That is your next value inflection point. No matter how ambitious and open-ended your plan is, you should have some sense of what is your next key value inflection point. What are the experiments that need to go into that? Price that out and then raise twice the money. There’s some amount of waste. I think if you can get waste down to 20 or 30%, that’s pretty good. Anyway, the advice is figure out what it costs to actually run the experiment.
Raise twice that and raise that or not. Beyond that, you’ll always discover new things. There’s usually some path through the mess. But when you start the company, you’re not going to get a guarantee that you won’t be on a bridge to nowhere or that it will work on the funding that you have. You’re going to have to get in there and figure it out halfway through. I think that people should push for profitability sooner than they often think that they need to. For us, even though we are seen as this very open-ended, deep tech company with a very long roadmap, which is true, we are also relentlessly focused on revenue at this point. We are trying to get to sustainability. It feels like the company is kind of constantly dying slowly of this money cancer that we can beat into remission every couple years with the fundraising, but then it eventually comes back and I want that feeling to be over.
No matter how big of a problem or big of a vision it feels, you do need to think about how do you get to revenue so that not just you can do it forever, but then you’ll be valued on your long-term roadmap, not valued on your probability of dying. It really opens up another set of investors that wouldn’t be relevant otherwise.
What is the best piece of advice you’ve received?
I don’t know. I’ve acquired way too much brain damage over the last 20 years to have a memory capable of picking that out. Other than speed being the basis of success and infrastructure determining your speed, it is important to appreciate that there are no general principles. People are looking for shortcuts. People are looking for a pithy set of instructions that are like, “Oh, I figured it out.” And that doesn’t exist. Every one of these things is different. When you get to that moment in history, you’re doing something new. We can reflect for a second on how crazy it is that this is possible. For the vast majority of human history, if you were a smart 20-year-old who had an idea to make your society better and you raised this to the people with capital, the reaction was, “You should pay attention to the harvest.”
The fact that it is not widely available—it’s not universally available—but it’s now widely available that if you’re a really smart 20-year-old, you can come to San Francisco and make the case. And if it’s an interesting idea, you’ll get millions of dollars to find out. This is not the case for most of the world today. And it’s certainly not the case for most of history anywhere.
But that shouldn’t feel normal. This is given to push the frontier out. And when you’re on the frontier, you’re figuring it out as you go. That is the job. So I would try to rely less on things that feel like startup advice and more on how good your judgment is. How well is that refined in your domain? And remembering that you have to think for yourself.
What are some of the hardest remaining engineering challenges involved in getting BCIs to work?
So in BCIs, we often feel very limited by power and thermal constraints on the implants. This creates a strong pressure to implant as little as possible and do the rest off the body. You can’t pass a wire through the skin because the skin is a very important immune barrier. The skin won’t fully heal around it. If you have any connector through the scalp, you’re constantly at risk of a bacteria crawling down that and into the brain, and then the patient’s going to have a really bad time.
And so you really have to be able to close the skin. That requires you to have implanted a radio or transceiver of some sort. Getting the power on that down—there’s a frontier at low power electronics, which is really important. As I mentioned earlier, a lot of this is now becoming increasingly biology as our biological engineering capabilities increase. But then on those implants, ironically, one of the harder, more open problems is what we call packaging. Our colleagues in Europe call it tropicalization. This is your ability to keep your device in and the body out of an implant that you put in the body. There are no truly passive surfaces anywhere in the body. Even bone is constantly getting remolded. So if I put a device in, it’s going to be attacked by the body and it’s not regenerating itself.
You need a material that is going to survive that for an extended period of time. The classic example of this is the laser welded titanium can, which, if you’ve seen a pacemaker or a deep brain stimulator, they’ve got this big titanium box. Obviously, we can’t put a big titanium box in the eye. Interestingly, one of the earlier retinal prostheses before us, about 10 years ago, was a device that did have a titanium box that they attached to the eyeball. It was a four and a half hour surgery. They had a little belt that went around the eyeball with a little titanium box on the side of the eye with a battery and a little PCB. This didn’t work. This was not good enough. They needed to get rid of that somehow. In our case, we’ve solved this with the laser projection trick where we power it wirelessly.
But having this next generation packaging—some type of conformal coating that we can use to protect the implant that is not degraded by the body, is also not harmful to the body, and is resistant to all of the ways that the body will try and kill it—that material science is a very open-ended field. If you’re interested in material science, that is a thing that we need progress in.
How did you approach interacting with the medical field to build your retinal implant?
Business is just a fancy word for talking to people and doing things. You talk to them, you send them emails. For our retinal implant, it was originally invented by a professor at Stanford almost 15 years ago, I think. It was licensed to a European company that we were tracking. Let me back up a second. When we started the company, I came from Neuralink. Four of my five co-founders came from Neuralink. We took a look around the world in early 2021 and asked what is the most valuable thing that we can do that would be likely to work in the near future, that may have a big impact to patients and allow us to be the foundation for the type of scalable medical device company that we wanted to build.
And we came to the conclusion that restoring vision to the blind by stimulating the retina was the thing. In that, you have a choice of two types of cells in the retina that you can stimulate: these things called bipolar cells or the optic nerve. And you could do that electrically or you could do that optically. We explored all four quadrants of that. We developed internally a state-of-the-art gene therapy that optically stimulated one of those cells. And we identified this French company as being the state of the art in electrical stimulation. It’s a small community. You can meet people, you can talk to them.
It eventually made sense for us to acquire them. We ended up with the license to the technology, and we work with surgeons and doctors all the time. If there’s a new surgery that you want to figure out, typically this is best going through networks so that people are more likely to respond to your email. But we cold email surgeons all the time saying, “Hey, we have a weird surgery to develop. Do you want to be a consultant?” And people reply. Before I get to the next question, there is a real cultural thing here. In my time hanging out around the periphery of SpaceX, I observed that at least circa seven or eight years ago, probably like 20% of that company is what you might characterize as committed Martian colonists, and 80% are serious engineers. I think that those people are lunatics and they just want to work on the highest performance methalox engines in the world.
And you need both of those cultures to be really successful long term. And that’s especially tricky in medicine, because that’s a very, very conservative, arguably very authoritarian culture for the most part. Similarly, at our company, we have, I’d say, 30%—I mean, it’s an overtly transhumanist mission—and then 70% serious clinicians and scientists and researchers and people who think that those guys are crazy, but we’re going to build some really valuable medical devices for critical unmet needs in the process. I think one of the things that makes Science, the company, very special is that it has both of those cultures and is able to integrate them. We’re able to simultaneously do some really cool research that I think is really at the edge of the Overton window, while simultaneously running clinical trials in six countries, now with an approved medical device in Europe and clinical trial results in the New England Journal of Medicine.
You have to be able to navigate both of those things, I think, to really reshape the future.
Has biotech gotten easier to break into for earlier-stage founders?
Biotech remains capital intensive. I don’t know that I’d recommend biotech if you have a choice of other stuff to do. For me, I realized almost 30 years ago that if you could alter the brain, you could alter reality. This was one of the biggest missions of the next 30 or 40 years—building these things. Every now and then, I think my life would be way easier if I had just gone into AI instead of BCI, but somebody has to do it. I think it’s important to—
Biotech is hard. It is a much harder path than many other things you can do, but when you’re successful, it has an impact that you don’t really see elsewhere. Increasingly, there’s—Paul Graham wrote a long time ago that you get vibes in different cities: the vibe in Cambridge, Massachusetts is you should be smarter; the vibe in New York is you should be wealthier; the vibe in San Francisco is you should be more powerful. Especially with the rise of things like artificial intelligence, I think people realize that this isn’t just about money. For many of the most effective startup founders, it’s not about the money. It’s about changing something. There’s some way in which you want the world to be different. It just turns out that for that project, the for-profit company is an incredibly powerful way to marshal the resources required to cause the world to be different in that way.
This is not about money; this is about power. There are many different types of power. There’s economic power, there’s military power, but the power to heal the sick is a very dramatic one. When you get that, not only is that a real force to reshape the world, it’s one that can be shared very readily. You can’t share military power or economic power, but you can share the power of restoring sight to the blind or giving life to the cancer patient. The world is getting more complicated, and there are big impacts from all the things being worked on by the people in this room. Biotech is hard. It’s very capital intensive. It’s a long road. When you start a company in this space, you’re committing to a decade of your life that you will never get back, no matter how it turns out.
But the results of that, when it works—the impact that this has on patients and their families—is really unlike any other sector.
What’s a popular belief in tech that you think is wrong?
And I don’t even know what the popular beliefs in tech are now. Well, okay. Even the whole basis of building Helix is contrarian. I think that if you raise a Series A and then you tell your investors that you’re going to vibe code a purchasing system, I think that any reasonable board is going to ask you what you’re thinking. And we were able to do that because I never got those questions, because I control the company. But that’s one narrow example, I guess.
All right. Thank you.