Deep Tech Companies Lose to Their Org Systems, Not to Technical Failure
How fast a startup can iterate decides how long it can win; and iteration speed comes from "infrastructure" like procurement, hiring and performance review. Putting approvals at the budget layer and replacing reviews with eigenreviews is worth more investment than agonizing over the specific technology.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Line-by-line approval can't control money; budget buckets can
When he tried to personally gate every purchase, a $3,000 power supply was enough to stall the founder: should you wait three days for the auction and get it half price? But you're already spending $100,000 a week, and waiting a week to save half the cost of a power supply means the savings are already canceled out by that week's spending. So he offers a judgment: line-by-line approval is not the right place to control spending; review has to move upstream to the budget. Bucket the money first and let highly paid employees weigh trade-offs inside their bucket, rather than letting them get stuck on a $3,000 purchase.
— Max HodakExperiments look free because nobody has priced a single run
The hidden cost of hardware startups is that gases, resins and media are bought in bulk and then split across countless experiments; when nobody knows the true cost of a single experiment, experiments look free — they cost no dollars, only "the media in the fridge." Science built Helix for this, wiring procurement all the way through to manufacturing and batch records, and only then worked out that each wafer iteration costs $40,000. His arithmetic cuts deeper: even with a $20 million Series A, a 20-person team and 20,000 square feet, headcount eats half the burn and rent takes another $750,000 to $1 million, leaving only about $3 million a year that can actually go into research — gone in a few iterations.
— Max HodakResumes shouldn't be screened by HR first; the whole company should vote
He turned hiring into a four-step funnel: the first step is not HR screening resumes, but the system selecting seven or eight employees with similar backgrounds from across the company to vote, with options from strong yes to strong no — this avoids any one group becoming a bottleneck and replaces a single person's taste with averaged judgment. The actual funnel: 17% of applicants advance to a phone interview; after the phone interview about half get a take-home; the conversion rate from onsite to offer must stay at least 25%. The phone interview doesn't test specific skills but three things: judgment, horsepower and agency, and he tries to design AI-resistant assignments, such as the GPU kernel minimum-cycle challenge Anthropic has used.
— Max HodakAnnual 360 reviews should become a monthly re-vote
The annual 360 review, he thinks, is a low-feedback, high-trauma ritual. The replacement is an internal mechanism called eigenreviews: every four to six weeks, randomly have colleagues answer the same question — "knowing how this person has performed since, would you still vote to hire them today?" Weave the answers into a company-wide graph, then recursively weight it after the fashion of eigenvector centrality — people everyone endorses should have their votes count more. To prevent collusive voting, each run drops out a random subset of edges over a thousand rounds; if the scores come out bimodal, that suggests a possible voting clique that needs human investigation. The feedback lag drops from a year to about a month, and the annual HR process goes away with it.
— Max HodakWhat decides life or death isn't the technology, it's the weekly iteration rate
The convergence point of the whole piece is one sentence: iteration rate separates the winners from the losers. If you can accumulate one useful lesson a week while your competitor manages one a month, that competitor will never truly threaten you; between two technical paths, even if the other looks like it has plenty going for it, as long as this one has a clearly shorter iteration cycle it's worth prioritizing, because the compounding is too strong. Hodak specifically notes that deep tech companies rarely die because "the technology doesn't work"; more often, once they reach a few hundred people and hundreds of thousands of square feet, they haven't built the corresponding systems, so strategy can't connect to execution. He has seen too many teams of pedigreed scientists and engineers die precisely on execution not keeping up.
— Max HodakAI and regulation are a better combination than you'd think
Medical devices can't escape the quality system — a pile of standards detailed down to lithium batteries plugged into PCBs, circuit board insulation, even whether the corner of a shipping label peels up inside a vibration table. In the past you had to hire regulatory experts to find each applicable standard one by one and build compliance evidence tables, often burning months; AI can quickly find all the standards and generate the evidence tables. Hodak's counterintuitive conclusion: AI and regulation are a better combination than people think, since regulations are mostly "written in blood" and reasonable in essence — the difficulty is just that people are bad at executing them. He also stresses that companies need to be AI-native: hand all context, quality and batch records included, to agents; AI's current role for the team is a multiplier, not a replacement.
— Max HodakLong-horizon companies should chase revenue earlier than instinct says
"Some ideas are worth $50 million or $0 to try, but you can't try them for $5 million." His fundraising advice for long-horizon companies: first define the next key value inflection point, price the experiments needed to get there, then raise twice that amount and tolerate roughly 20-30% waste. More against the mainstream: even with a deep tech long-horizon label, pursue revenue earlier than instinct says. Hodak describes companies using round after round of financing to push "money cancer" back into remission, but that shouldn't last forever. With revenue, a company gets valued on its long-term roadmap rather than its probability of survival, and it opens up a set of investors who otherwise could never have come in.
— Max HodakThe strongest organizations hold two kinds of people who disdain each other
Observing from the periphery at SpaceX, he saw that about 20% of employees were "committed Mars colonists," while 80% thought those people were crazy and just wanted to build the highest-performance methane engine in the world — serious engineers. Both cultures are indispensable. Science deliberately replicated this: 30% are open transhumanists, 70% are clinicians and scientists who think the former are crazy. Hodak believes the company can simultaneously do frontier research at the edge of the Overton window and run clinical trials in six countries, win European marketing authorization and publish results in NEJM, precisely by getting these two types to tolerate each other. To truly reshape the future, you have to command both temperaments at once.
— Max HodakIn their own words · checked verbatim
It is mostly not that we are smarter. It is infrastructure like this.
Max Hodak6:14
if you can learn one thing every week, and there's a competitor that's learning a thing every month, they will never matter.
Max Hodak18:46
whenever you exert action into the universe, that creates information, like in a very fundamental sense.
Max Hodak22:51
If you get them wrong, you'll end up like spending five million dollars a month and feel like you have very little control over it.
Max Hodak23:54
I think that the combination of AI and regulation is, is a better fit than people think.
Max Hodak37:18
I think that people should push for profitability sooner than they often think that they need to.
Max Hodak44:36
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.
Max Hodak56:00
Figures
| Hiring funnel: share of applicants advancing to phone interview | 17% | 12:32 |
| Hiring funnel: share of phone interviews advancing to take-home | about 50% | 13:34 |
| Hiring funnel: minimum onsite-to-offer conversion rate | ≥25% | 14:38 |
| Runaway monthly spend when infrastructure isn't built (illustrative) | $5 million/month | 23:54 |
| Science internal culture mix | about 30% open transhumanists / 70% serious clinical staff | 52:51 |
| SpaceX culture mix (Hodak's observation) | about 20% committed Mars colonists / 80% serious engineers | 52:51 |
Glossary
- eigenreviews
- A performance review mechanism that periodically collects peer reviews and then weights them with a PageRank-like graph algorithm to determine who is truly trustworthy.
- AI-resistant homework
- Designing a written assignment so AI can't easily do it for the applicant and there's a clear optimization target, used to gauge an applicant's real ability.
- vibe coding
- Collaborating with AI in natural language to write software fast; this makes building internal tools in-house worthwhile for a startup for the first time.
- tropicalization
- The European term for the surface encapsulation technology of implantable devices, which must withstand the body-fluid environment long term without harming the body.
How to listen
Deep tech founders and early employees who are slowed down by procurement, hiring and performance processes; also investors watching the brain-computer interface space.
If you don't care about neural interfaces, you can fast-forward from 40:29 to 52:00; the cultural point at the end, 52:51, is worth pausing for.