Patients Are Reading Novels With Retinal Chips — The CEO Is Talking About Procurement

Source: Y Combinator | Published: 2026-08-07T21:30:32Z

A Science patient read a 300-page novel using a retinal chip and mailed it to the company. CEO Max Hodak says the real bottleneck to iteration speed isn't the tech — it's procurement processes and experiment cost transparency.

Science CEO Max Hodak spent five years at Neuralink before founding a company that makes retinal prosthetics. Last year, one of their patients read a 300-page novel using an implanted chip, then mailed the book to the company. Science has completed Phase 3 clinical trials, conducted clinical research across six countries, and published results in the New England Journal of Medicine.

This is not an ordinary startup — yet Hodak spent 30 minutes of a recent talk on procurement, expense reimbursement, and performance reviews.

The Cost of a Power Supply

Early in the company's life, an employee messaged Hodak asking to buy a $3,000 power supply.

His first instinct: Is that too much? There's an auction in three days — maybe they could get it for half price, with delivery in two weeks. But then he caught himself: he'd hired a team of well-paid engineers burning $100,000 a week in labor alone. Waiting two weeks to save $1,500 would cost him over $200,000 in burn.

The math is simple. He admits he actually had to sit down and work through it before it clicked.

The problem wasn't whether to buy the power supply. It was that the question of "who decides what to buy" was itself broken. If the CEO is approving every equipment purchase one by one, the approval process becomes the biggest waste of all. Spending control belongs at the budget level, not at the moment of purchase.

So they built a procurement system. But the moment it went live, a new problem emerged: engineers were spending hours navigating enterprise B2B software, and equipment that should arrive in two days was taking two weeks — because vendors needed to verify insurance credentials, sign agreements, issue quotes, generate purchase orders, and then invoice. The more formal the process, the slower everything moved.

Their fix: bring the procurement workflow in-house. They dedicated staff to running the process — tracking every order, proactively following up on progress. From the outside, this looks strange. "You have people whose job is just to buy things?" But that's where speed comes from.

Opaque Experiment Costs Are a Management Failure

Science is a hardware company. Large quantities of materials — argon, silane, nitrogen, culture media, resin — are bought in bulk and consumed across experiments. They're drawn from shared inventory and used in different runs.

The result: nobody knew what any given experiment actually cost. Media pulled from the fridge looks free.

The consequences cascade. Teams develop no sense of resource constraints. You can't compare the cost-effectiveness of experiments. And you can't price your product without knowing what it costs to make — you need to know the manufacturing cost of an implanted chip before you can decide what to charge for it.

To fix this, they built their own internal management software called Helix. The system captures every step in the lab — materials consumed, time spent, labor involved — and links it all together. They eventually discovered that running a wafer process costs $40,000.

Hodak walked through a rough model: Series A raised $20 million, planned for a four-year runway, 20 people, with roughly half going to labor — about $3 million per year in personnel costs. Add 20,000 square feet of office space at $4 per square foot, and you're spending $750,000 to $1 million annually on rent. After everything, the actual research budget is around $3 million per year, stretched over three or four years.

The team sees the fundraising announcement and thinks there's plenty of money. There isn't. Without this system, they would never have known the true cost of each cycle — and without that, you can't make any rational tradeoffs.

Hiring Funnel: 17% Make It to a Phone Screen

There's a common failure mode in tech recruiting: the founder or an HR person sits at the top of the funnel, personally reviewing every resume. When the company is tiny, this is manageable. Once you have any reputation at all, inbound volume overwhelms that bottleneck.

Science distributes the initial screen across the whole company.

When someone applies, the system automatically selects seven or eight current employees whose backgrounds seem relevant to the candidate and asks them to rate: strong recommend, recommend, don't recommend, strong don't recommend. Results are aggregated within 24 to 48 hours. No single gatekeeper. Fast. And judgment is distributed across more people.

Their funnel numbers: 17% of inbound applicants advance to a phone screen, 50% of those move to a take-home, and on-site conversion after the take-home must stay above 25% — otherwise something in the earlier stages has broken.

The phone screen isn't a technical test. It looks for three things: judgment, learning ability, and initiative. What decisions do you make under complex, ambiguous conditions? Do you have concrete evidence of what you've learned? Does the arc of your life suggest you're actually pursuing something?

For take-homes, they prefer problems with high ceilings that AI can't trivially solve — things like: how few cycles can you get a piece of code down to, and can you beat a model's benchmark performance? If a task can't be AI-proofed, they substitute a technical phone screen or an in-person exercise.

They explicitly don't use LeetCode. For software, the evaluation is "can you decompose a complex problem" — not "have you memorized this question."

PageRank Logic for Performance Reviews

Hodak is blunt about the problem with traditional annual performance reviews: they usually just confirm what you already know but have been avoiding. The problem has been there for months. Firing is hard, so you wait until December.

His alternative is called IGEN Reviews, used for six or seven years now.

Every four to six weeks, the system randomly sends employees a question: "If you were voting again today, would you still support hiring this person?" Same rating scale as the initial hiring screen. Not HR-organized, not a form a team lead fills out — just evaluations that emerge naturally from daily collaboration.

All votes form a graph, and weights are determined by the graph's structure: if someone who rates you highly is themselves highly rated, their assessment carries more weight. The logic is close to Google's original PageRank algorithm — eigenvector centrality.

To detect cliques inflating each other's scores, they use a technique: randomly drop edges in the graph, run a thousand iterations, and check the distribution. Multiple peaks signal a possible voting cluster that needs manual review.

The result is a continuous signal with roughly a one-month lag, not a once-a-year event. Hodak compares traditional annual reviews to a loss function with enormous noise and terrible latency. IGEN Reviews is something like a synthetic gradient — early, continuous, telling you where something is going wrong before it's gone all the way wrong.

Judgment Is the One Thing You Can't Outsource

Hodak says that in year four or five of running a company, you hit an inflection point: there's a high-stakes decision that only you can make, and when you go looking for advice, you find there's no one to ask.

That's not a tragedy. That's the job.

He compares a startup to an exam — you can't win by copying the person next to you, because that only moves your score toward the class average. Making good company-level judgments requires being, at certain moments, the one person standing alone while everyone else thinks you're wrong.

This demands that you constantly calibrate your own judgment: which domains can I trust myself in, and where do I go wrong? His five years at Neuralink shaped him more in this respect than any technical experience. What mattered was the repeated experience of making real-stakes decisions and watching the outcomes unfold. He calls it reinforcement learning — and says it's the most underrated part of a founder's education.

His advice: before starting a company, go work somewhere with a culture you genuinely respect. Startup culture is rarely invented from scratch — it's usually inherited. One company shapes the next. Watching someone with demonstrated judgment make decisions up close is worth more than any entrepreneurship course.

Speed Is Determined by Infrastructure

This is the core argument of the entire talk: iteration speed determines outcomes, and iteration speed is not set by your technical ability — it's set by your infrastructure.

What is infrastructure? How fast your procurement process moves. Whether you can actually calculate experiment costs. Whether your hiring funnel has bottlenecks. How often performance feedback updates.

He gives an extreme example: if you learn one new thing per week and your competitor learns one per month, they will never catch up. The effect isn't additive — it compounds.

He's seen many companies founded by world-class scientists and engineers ultimately fail — not because the technology didn't work, but because when the team grew to hundreds of people and the physical space grew to tens of thousands of square feet, there was no system to hold the organization together. The connection between strategy and execution snapped.

That's why they built Helix instead of buying NetSuite or Greenhouse. Commercial ERP software is designed for the generic case. But every company's operations have their own shape, and when you build around software that's actually designed for you, you gain an efficiency advantage that can't be purchased.

He points to YC, Facebook, SpaceX, and Tesla as companies that have all invested heavily in internal tooling. Historically this was considered not worth the development cost. AI-assisted coding has dramatically lowered that barrier. This, he says, is one of the things that has genuinely changed in the AI era.

"Someone Did Something" Is Evidence

Someone asked Hodak what counts as proof of exceptional ability.

His answer: being able to point to one specific thing that sets this person apart from their high school classmates.

The best evidence is winning in a competitive game with real feedback. Chess master. Formula SAE champion. He mentions several deep tech companies in Silicon Valley whose core teams were recruited almost entirely from university Formula SAE championship teams — people who spent four years of college building race cars, competing, and figuring out what went wrong.

There's a deeper logic here: you need to have experienced making a judgment with real stakes, then waiting for the outcome, then getting real feedback. Part of what draws smart people to software, he notes, is exactly this — you have an idea, and within hours you know whether it works. In biology, the same idea might take months to test. Fast, high-frequency feedback loops are accelerators for developing competence.

How a Retinal Implant Company Sees AI

Hodak says he's barely looked at source code in the past six months. He's largely moved to AI-assisted coding.

But the area where he thinks AI has a bigger impact on their company is regulatory compliance.

Medical device development requires tracking a massive number of applicable standards — from PCB interfaces for lithium batteries to packaging labels with corners that can't curl and peel off. In the past, this meant specialized regulatory consultants spending months compiling evidence matrices. Now AI can quickly search all relevant standards and generate compliance documentation. His read: the combination of AI and regulation works better than most people expect. Most regulations exist for good reasons, paid for in blood — humans are just inefficient at executing on them. AI fills that gap.

On whether AI will replace human researchers: his practical answer is that for now, AI is a multiplier for the team, not a replacement. Their approach is to record everything that happens at the company in Helix — not just for internal management, but so that context can eventually be handed off efficiently to AI agents. The companies that thrive in the AI era are the ones that got their data structures in order ahead of time.