Add to Google Preferred Sources
Science Corporation CEO Max Hodak argues that the graveyard of deep tech is filled not with companies that couldn’t make their technology work, but with companies that outgrew their own management systems. In a Y Combinator Startup Podcast interview, the former Neuralink president details the unglamorous infrastructure — purchasing systems, hiring funnels, and a PageRank-style peer review mechanism he calls “eigen reviews” — that he says actually determines which startups survive. His core thesis: iteration speed compounds, and a company learning one thing per week will bury a competitor learning one thing per month, regardless of technical advantages. The interview comes as Science’s retinal prosthesis, already approved in Europe, enters commercial phase after completing major clinical trials in 2025 — with one patient reading a 300-page novel using the device. Hodak also offers contrarian views on fundraising, arguing that some ideas deserve $50 million or nothing at all, and that deep tech founders should push for profitability far sooner than convention dictates.
Key Elements
If you think deep tech startups die because the physics doesn’t work out, Max Hodak wants you to think again. “It’s uncommon that deep tech companies fail because the technology doesn’t work,” he said in an interview on the Y Combinator Startup Podcast. “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.”
Hodak is the CEO of Science Corporation, the neurotechnology company behind a retinal prosthesis that is now an approved medical device in Europe. One of his patients used the implant to read a 300-page novel. The results were published in the New England Journal of Medicine. And yet Hodak spent almost none of his hour-long interview talking about the implant itself — the hexagonal grid of photovoltaic cells slipped under the retina, the laser projector mounted in a pair of glasses, the six-country clinical trials. Instead, he focused on procurement. On hiring pipelines. On a custom-built feedback system that uses eigenvector centrality to detect toxic cliques inside a company.
The message was blunt: the unglamorous plumbing of company operations is the actual determinant of survival, and founders who ignore it are running on borrowed time.
Iteration speed is the only moat that compounds
Hodak opens with two borrowed aphorisms that frame his entire philosophy. The first is from Picasso: art critics debate form and structure, while artists talk about where to buy cheap turpentine. The second is a military maxim attributed to General Omar Bradley: “Amateurs talk strategy, professionals talk logistics.”
The point is that founders obsess over object-level technical brilliance while neglecting the “living organism” of company operations. When Science appears to move faster than competitors, Hodak insists, “it is mostly not that we are smarter. It is infrastructure like this. That is how speed is built.”
The stakes are quantified with compound-interest logic: “If you can learn one thing every week and there’s a competitor that’s learning a thing every month, they will never matter.”
This logic extends all the way down to technical choices. All else being equal, Hodak says, pick the approach with the shorter iteration cycle. The compounding advantage will eventually overwhelm whatever “redeeming characteristics” the slower approach offers. The founder’s daily reality is “a continual stream of facts” hitting the desk; infrastructure is what turns those local decisions into organizational learning. Without it, companies “end up spending $5 million a month and feel like you have very little control over it, and then you’re forced into coarser levers and harder decisions.”
Where deep tech dies quietly: purchasing and burn
The first concrete infrastructure Hodak describes is procurement, and he frames it as a direct rebuttal to the software-startup fantasy. The myth, as he tells it, is “me alone in an empty room with some computers writing software.” That fantasy is part of why VCs love funding pure software. But anything that touches the physical world means buying “many thousands of things” — computers, microscopes, electronics, 3D printers, resin, PCBs.
The naive approach is to give everyone a credit card and approve purchases as they come. That produces approval requests for a $3,000 power supply, followed by an internal debate: an auction in three days might get it at half price, but delivery takes two weeks. Meanwhile the company burns $100,000 a week in payroll, so any delay dwarfs the saving. “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,” Hodak noted.
The lesson is that transaction-level approval is the wrong place to scrutinize spending. The right approach is giving every employee a felt sense of their resource bucket so they make trade-offs within it themselves.
The professionalized alternative is almost worse. In a typical procurement system, placing an order takes two weeks — create a vendor account, handle insurance and certification paperwork, route quote to purchase order to invoice. Academics joining a startup find this baffling: “Why are there people whose job it is to purchase things? Surely I can just buy things.” But the function is real and requires active management.
The deeper problem is attribution. Science buys gases, resins, and media in bulk, distributing them across many experiments, which breaks any straightforward accounting of what a given experiment costs. “Experiments are free. It doesn’t cost dollars. It costs media. And media comes from the fridge,” Hodak explained. Pricing anything they manufacture requires “spreadsheets with opinions” — decisions about whether to include rent, tool depreciation, and assumptions about future volume.
Hodak’s solution was Helix, internal software where “almost everything you can do in the company is a button somewhere.” The system extends through manufacturing so every lab step sits in a database and can be correlated. The output is sobering: one wafer iteration for one protocol runs $40,000.
His burn arithmetic for a typical deep tech Series A paints an equally concrete picture:
| Line item in a $20M Series A plan | Hodak’s estimate |
|---|---|
| Headcount (20 people) | ~$3M/year — “about half the burn” |
| Facility (20,000 sq ft at ~$4/sq ft) | ~$750K–$1M/year |
| Remaining research budget | ~$3M/year, for 3–4 years |
The punchline: the team watches you raise more money than they have ever seen and concludes that every $3,000 power supply is essentially free. “This infrastructure actually determines success or failure in many companies.”
The hiring machine: a four-stage funnel with engineered conversion rates
Hiring, Hodak argued, “really separates the successes from the failures.” The best companies emerge from “scenes” — a nucleation moment plus an extended community that becomes your initial recruiting target — but eventually you have to hire from the general public. There is no single correct hiring process, “but a wrong answer for sure is not having something that you do very religiously as a company.”
Science’s funnel runs through Helix after the company replaced the commercial applicant tracking system Greenhouse. The architecture reveals a deliberate engineering of conversion rates at each stage:
The first-stage company-wide vote is the load-bearing design element. The system selects seven or eight current employees whose backgrounds resemble the applicant’s and pings them for votes on a five-point scale: known-good, strong yes, yes, no, or strong no. This distributes the top-of-funnel review across the entire company so no small group becomes a bottleneck. “If you’re doing anything cool, by the time you get a couple years into it, that top of funnel is overwhelming,” Hodak said. The process completes in 24 to 48 hours, with roughly 17% of applicants advancing to a phone screen.
The phone screen is a company-wide bar, not team-specific. Interviewers assess three qualities: judgment (thrown into a vaguely defined situation, do you make good decisions or create diplomatic incidents?), horsepower (basic technical competence plus demonstrated ability to learn), and agency (“are you effective at causing the world to look like you wish it were? Do you have specific ambitions for your life?”).
The homework stage is the most distinctive. Candidates receive tasks that “don’t saturate, have a very high ceiling, and are naturally scorable to two or three numbers” that can be plotted on a graph, making it obvious when someone beats the existing Pareto frontier. There is no penalty for using AI models. “We don’t care whatever AI models they use — like that can make you better,” Hodak said. The model example: GPU kernel optimization asking for the minimum cycle count, where for a period the bar was simply beating Claude Sonnet’s performance to earn an interview.
Interview conversion must stay at or above 25%. Below that threshold, too much team time gets consumed by on-sites that don’t convert to hires. Hodak insists these four steps represent “the minimum set of information that we need to make a full decision.”
Two auxiliary hiring doctrines round out the philosophy. First, the strongest evidence of exceptional ability is “winning at legible competitive games” — chess grandmasters, Design-Build-Fly champions, Formula SAE winners. “There’s a bunch of Silicon Valley deep tech companies that are basically built out of Formula SAE winners,” he observed. Second, engineering hiring should evaluate thinking, not programming. “We’ve never really used LeetCode… We try to evaluate thinking — can you understand the decomposition of the problem clearly? It’s really measures of can you think clearly rather than can you write code.”
The eigen system: reviews as a synthetic gradient
The market’s feedback on whether you hired well arrives far too late — “a very, very long feedback loop, a very poorly behaved loss function.” Annual 360 reviews, in Hodak’s experience, are disruptive and “don’t tend to surface issues that you don’t already know about but haven’t acted on.” The reason is simple: firing people is hard, and people drag their feet.
What Hodak wanted was a synthetic gradient — continuous, roughly unbiased signal about who is good and what is working. For six or seven years, he has run a mechanism that works as follows: every four to six weeks, employees get pinged through Helix with a single question — “knowing how this person turned out, would you vote again today for their hire?” — using the same five-point scale as the initial hiring votes.
The votes form a directed graph across the company. Crucially, “your vote should be weighted more highly if everybody else has rated you highly” — a PageRank-like eigenvector centrality scheme that Hodak calls “eigen reviews.” Robustness comes from Markov chain Monte Carlo dropout: the system runs roughly 1,000 iterations, randomly removes a percentage of edges each time, and inspects the resulting score distribution. Extra peaks in that distribution reveal voting cliques that need investigation — which is also the answer to concerns about collusion and targeted downvoting. “It’s not our only signal. It’s one of several,” Hodak noted.
| Dimension | Annual 360 process | Eigen reviews |
|---|---|---|
| Cadence | Once or twice a year | Every 4–6 weeks |
| HR burden | Forms, meetings, coordination | Automated software pings |
| What it surfaces | Known issues not yet acted on | Continuous signal, ~1-month lag |
| Bias governance | None | Centrality weighting + MCMC clique detection |
| Track record | — | In use 6–7 years |
Hodak offered to share the detailed implementation document
What the machinery is for: a retinal prosthesis that works
The infrastructure exists to advance a specific product frontier. Science’s retinal prosthesis operates on a deceptively simple principle:
Each hex on the implant grid is essentially a solar cell. The laser projection both powers and drives the device wirelessly, eliminating any through-skin connector that would invite bacterial infection. Hodak is direct about the evidence: three clinical trials to date, major trials completed in 2025, device approved in Europe, results in the New England Journal of Medicine, a patient reading a full novel with the implant. “That is a real, not a marketing, data point.”
The technology origin story is itself a case study in deep tech M&A. The implant was invented at Stanford roughly 15 years ago and licensed to a European company. Science explored the full decision space: two retinal targets (bipolar cells or optic nerve) crossed with two modalities (electrical or optical). The electrical-bipolar route led to a French company that represented the state of the art — which Science acquired, gaining both the license and the technology. “Business is just a fancy word for talking to people and doing things,” Hodak said. They cold-emailed surgeons: “Hey, we have a weird surgery to develop. Do you want to be a consultant?” People replied.
The hardest engineering frontier is physical. Power and thermal constraints push toward implanting as little as possible, since skin is a critical immune barrier. The most underrated problem, Hodak says, is packaging — what his European colleagues call “tropicalization”: keeping the device in and the body out. “There are no truly passive surfaces anywhere in the body. Even bone is constantly getting remolded.” The classic solution, a laser-welded titanium can used in pacemakers and deep brain stimulators, cannot fit in the eye. A predecessor retinal device that strapped a titanium box and battery to the outside of the eyeball required a 4.5-hour surgery and “didn’t work.” Science’s wireless laser projection sidesteps that constraint, but next-generation conformal coatings remain an open materials-science problem.
The brain as a longevity play
Hodak’s framing of brain-computer interfaces departs from the AI-adjacent narrative that dominates Silicon Valley. “In the near term, BCI is really a longevity story — essentially healthcare and biotech,” he argued. “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.”
Neural engineering, he noted, delivers effect sizes that conventional medicine rarely sees. A deep brain stimulator takes a Parkinson’s patient “from not being able to hold a cup of water to being able to write cursive in like 10 seconds” — a contrast to drugs whose effects are “a little bit, dwindling over time.” A newborn’s cochlear implant activation, he said, is arguably the strongest patient testimonial in all of medicine.
The long-term vision compresses into two moves. First, consciousness is a practical problem, not a philosophical one: “The brain is composed of ordinary matter arranged according to the rules of chemistry, only things found on the periodic table.” With a sufficiently capable BCI able to read and drive every neuron, “I think we’d figure out consciousness pretty fast.” If the endpoint of AI is superintelligent machines, “the end of the BCI quest is conscious machines” — and eventually, superintelligent conscious machines that humans can participate in. “At some point, the boundary between these technologies becomes less meaningful.”
AI becomes the company’s operating system
The throughline that makes Helix coherent is AI. Hodak is explicit that bespoke internal software is only now a rational choice at the startup stage because agents and “vibe coding” have changed the economics. The design principle is AI-native: “Gather all of the context, all 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.”
He reports AI as a multiplier, not a replacement. The biggest impact areas so far are coding and regulatory compliance. On coding: “I’ve written a lot of code in my life. I don’t think I’ve looked at the source very much the last six months.” On compliance, the quality-system bureaucracy is “the quintessential heavy bureaucracy” — the idea of quality itself is sound, but humans are bad at reading and interpreting requirements. Standards cover everything from lithium-ion battery connectors on PCBs to electrical insulation to whether shipping-label corners curl inside a vibration box. Historically, you hired regulatory experts to enumerate standards and build evidence spreadsheets over many months. AI now generates the standards lists and evidence tables almost immediately. “Regulations are written in blood and largely good ideas — it’s just hard for humans to do it.”
On build-versus-buy, Hodak is blunt: “There’s no company that loves their ERP system. I don’t know there’s anyone who’s really like, I want to spend more time in NetSuite.” His precedents include Y Combinator’s internal software, Facebook’s internal tools, and the “pretty giant” piece of software called Warp Speed that runs much of SpaceX and Tesla’s manufacturing and R&D. A company growing up around software fitted to its own shape is powerful in ways off-the-shelf software is not. Historically, custom software was too expensive, so everyone bought. “That was I think a worse world and that world has changed.” The concrete failure of buying: Greenhouse forced a small human bottleneck at the top of the hiring funnel and foreclosed the voting mechanism. Custom software enabled “smart inferences about who would know about an applicant.” His rule: build what works for you, and “bake it into the company so when you put something there it stays there.”
Fundraising, judgment, and the apprenticeship no one talks about
Hodak’s closing arc is about judgment — developing it, funding it, and being forced to deploy it alone. On fundraising, his core accusation is that founders, especially experienced ones, pitch VCs for what seems reasonable to ask rather than what the experiment actually costs. “You’re raising some amount of money to go find out some answer. The answer to that might be no.”
Some ideas “are worth funding with $50 million or zero dollars, but not $5 million” — the middle amount produces an ambiguous outcome and a frustrating experience. His rules of thumb: define your next value inflection point, price the experiments required to reach it, and “raise twice the money.” Accept that 20% to 30% of capital will be wasted — “that’s pretty good.” There is no guarantee you won’t end up “on a bridge to nowhere.”
Most provocative is his stance on profitability: “I think that people should push for profitability sooner than they often think that they need to.” He describes a company as “kind of constantly dying slowly of this money cancer that we can beat into remission every couple years with the fundraising.” Revenue changes what investors value you on — long-term roadmap instead of probability of dying — and unlocks a different set of investors.
His years at Neuralink were, in retrospect, an apprenticeship. He worked alongside someone with “empirically excellent judgment,” where a problem would arise with two plausible solutions and the answer was “oh, it’s definitely option B, the problem would never recur.” The value was reinforcement learning on decisions with real stakes and delayed feedback — “that is an essential part of the education of an entrepreneur that I think many people underrate.”
Beneath this sits a physics-flavored theory of action. A thrown ball’s ballistic trajectory is “information-minimizing” — it’s the path of least action. Whenever you exert action on the universe, you create information. When stuck, you “have to start injecting action, producing entropy.” The counterintuitive corollary: a company trapped in a deep local minimum can sometimes be unblocked by removing someone who is individually strong but wrong-fit, because the action itself reshapes the system. “The action space is always larger than it appears.”
His final position is almost anti-advice: “There are no general principles. People are looking for shortcuts. That doesn’t exist. When you get to that moment in history, you’re doing something new.” What should not feel normal, he says, is that a smart 20-year-old can come to San Francisco and get millions of dollars to test an interesting idea — “that shouldn’t feel normal.”
He borrows Paul Graham’s map of city vibes — Cambridge tells you to be smarter; New York, wealthier; San Francisco, more powerful — and reframes ambition accordingly: “This is not about money. This is about power,” including the shareable power of restoring sight or “giving life to the cancer patient.” That path costs a decade of your life “that you will never get back no matter how it turns out.” He admits he sometimes thinks life would be easier if he worked on AI instead of brain-computer interfaces, “but somebody has to do it.”
The interview closes on one telling confession: his most contrarian infrastructure choices were never scrutinized by his board “because I control the company.” A reasonable board, he notes, would question a founder who announced they planned to vibe-code their purchasing system. He never got those questions.
For investors and operators watching the deep tech landscape, Hodak’s framework offers a specific lens: the companies worth betting on may not be the ones with the most elegant technology, but the ones that treat their own operations as an engineered product — with measurable throughput, deliberately designed feedback loops, and a founder who understands that judgment, unlike purchasing, can never be delegated.
Full content available at:Max Hodak: “Speed Is Determined by Infrastructure”
- US Stocks Surge as Sector Rotation Intensifies; Hon Hai ADR Jumps Over 4% While AMD Plunges Nearly 6%
- A-Share Markets Open Lower Across the Board; Semiconductor Materials and Memory Sectors Lead Declines
- China A-Shares Plunge at Open, Zhongji Innolight Tumbles 13% Leading CPO Rout
Once added, BigGo Finance appears first in Google Search Top Stories, so you get the broadest, most up-to-the-minute, and most comprehensive global financial news first.
