Max Hodak: Average Is Not Good Enough

Direct Answer

Max Hodak's central argument is that a startup's speed comes from its operating infrastructure, not only from the brilliance of its product team. Strategy becomes real through the systems that let employee 17 buy a piece of equipment, show what an experiment actually costs, move a candidate through a rigorous hiring funnel, surface performance problems early, and keep decisions attached to a human owner.

The talk is especially useful because Hodak rejects the usual promise of a universal founder checklist. Science Corporation builds medical devices, semiconductors, biological systems, and clinical programs. Its exact processes will not fit a two-person SaaS company. The transferable lesson is narrower: find the repeated delays that prevent your company from learning, then design an operating system that removes those delays without hiding cost, risk, or accountability.

Watch Max Hodak at Startup School 2026

Video credit: Y Combinator and Max Hodak. Hodak is co-founder and CEO of Science Corporation. Product and clinical claims below are checked against Science's first-party material, ClinicalTrials.gov, and the peer-reviewed PRIMAvera paper. Internal costs, hiring conversion rates, and company processes remain Hodak's account unless stated otherwise.

What Science's Retinal Implant Has Demonstrated

Hodak opens with the product that gives the operational discussion its stakes. PRIMA is a visual prosthesis for people who have lost central vision because of geographic atrophy from age-related macular degeneration. A small photovoltaic array is implanted under the retina. Camera glasses capture the scene and project a processed pattern of near-infrared light onto the implant. The array converts that pattern into electrical stimulation for surviving retinal neurons, bypassing photoreceptors damaged by disease.

The pivotal PRIMAvera study was an open-label, multicenter, prospective, single-group study with 38 participants. Its peer-reviewed New England Journal of Medicine paper reported that the system restored useful central visual function for many participants, including meaningful visual-acuity gains and reading tasks. Science says 80% of participants demonstrated a clinically meaningful improvement. The device supplements remaining peripheral vision; it does not recreate normal color, field of view, or natural visual acuity.

Hodak says one participant used the system to finish a roughly 300-page novel and sent the book to the company. That is a powerful first-person outcome, but it is an anecdote rather than the trial's primary endpoint. Science's March 2026 update said it had submitted regulatory applications in Europe and the United States and was funding commercialization. Availability and approval status can change, so patients should use current regulator and clinical sources rather than this article for medical decisions.

The Claim Ledger: Evidence, Internal Data, or Founder Doctrine?

The Company Operating System

A product strategy says what the company wants to make true. A company operating system determines whether people can repeatedly do the work required to learn if that strategy is correct. Hodak's examples form seven connected layers.

The layers must connect. Fast procurement without experiment attribution creates uncontrolled spend. Strong hiring without feedback allows drift. Rapid iteration without safety gates can produce faster harm. An operating system is useful only when it increases speed and preserves visibility.

Procurement Creates Speed Before It Creates Savings

Hodak's test is wonderfully concrete: how does the seventeenth employee buy a $3,000 power supply? A founder can use a credit card. A growing company introduces permissions, budgets, vendor records, insurance requirements, purchase orders, invoices, and accounting. Each control has a purpose, but unmanaged controls can create a queue longer than the experiment itself.

The key design move is to put the economic decision earlier. Leadership sets budgets, permitted categories, risk tiers, and approval ceilings. Inside those boundaries, routine purchasing should be fast. Review unusual vendors, large commitments, security exposure, regulated materials, and irreversible contracts. Do not force a fresh strategy meeting for every ordinary component.

Cost of delay

blocked people x loaded hourly cost x blocked hours
+ idle equipment or facility cost
+ expiring materials and rescheduling
+ missed revenue or milestone value
+ risk created by the delay
= total delay cost

Decision rule
If delay cost is greater than purchase saving,
the cheaper purchase is economically more expensive.

Measure procurement like an operational system: median request-to-order time, 90th-percentile time, percentage of requests returned for missing information, emergency purchases, budget variance, and delay cost. The goal is not maximum spending freedom. It is the shortest accountable path from a legitimate need to the work.

An Experiment Is Not Free Because the Invoice Is Shared

Deep-tech teams often buy gases, resins, media, wafers, animal work, external assays, and equipment in bulk. When those costs sit in a shared account, each experiment feels free to the person choosing it. Hodak says Science connected purchasing data to its manufacturing and laboratory records and discovered that one wafer iteration could cost roughly $40,000.

Exact allocation will never be perfect. It does not need to be. A consistent approximation is better than a precise-looking zero. Use the same rules across candidate protocols, document assumptions, and show a range when uncertainty is material.

Build a Hiring System Before Founder Instinct Becomes the Queue

Early startups recruit from the network and technical scene that produced the company. That source can be excellent, but it eventually runs out and can reproduce the same backgrounds. Hodak says Science built a four-stage funnel: distributed application voting, a phone screen, a job-relevant homework exercise, and a full interview. The system is designed to make broad participation possible without sending every decision through the founder.

He describes three traits: judgment, horsepower, and agency. The second term is memorable but imprecise. A safer scorecard turns all three into observable behavior.

Hodak prefers work samples with a high ceiling, numerical scoring, and a moving frontier. AI use can be allowed when the job itself permits AI. The purpose is not to create an "AI-proof" puzzle. It is to observe problem framing, tool choice, verification, communication, and the accepted result. A candidate who uses an agent well may be demonstrating the job more faithfully than one forced to hide it.

Structured interviews and work samples still need governance. The U.S. Equal Employment Opportunity Commission says selection procedures should be job related, validated for their purpose, and checked for discriminatory impact. The U.S. Office of Personnel Management's structured-interview guidance recommends consistent questions, anchored rating scales, trained interviewers, and job analysis. Other jurisdictions impose different duties, so employers need local advice.

Eigen Reviews Are an Experiment, Not a Default

Hodak criticizes annual 360-degree reviews because they are disruptive, delayed, and often restate problems everyone already knew. Science instead asks employees every four to six weeks whether, knowing what they know now, they would vote to hire a colleague again. Its internal system weights responses through a graph inspired by eigenvector centrality: ratings from people who are themselves trusted carry more weight.

The idea is clever and the risk is substantial. A network score can confuse visibility with value, amplify dominant groups, penalize dissent, hide retaliation, and make social popularity look mathematical. Detecting cliques with repeated graph runs or dropout does not remove the underlying employment and power issues.

The more transferable alternative is continuous, specific feedback attached to observed work. Ask what outcome was expected, what happened, what evidence supports the conclusion, what support is missing, and what changes next. Shorter loops are valuable; opaque social ranking is optional.

Iteration Speed Is an Infrastructure Property

Hodak argues that a company learning weekly will separate from a competitor learning monthly. The compounding intuition is useful, but speed is not the number of tickets closed or experiments started. An iteration counts only when it changes what the team knows and improves the next decision.

Accepted learning loop

observation
-> explicit assumption
-> smallest reversible action
-> success and failure thresholds
-> controlled execution
-> independent result check
-> decision and owner
-> updated operating rule

Procurement shortens the wait before execution. Experiment costing helps choose the right test. Hiring provides the required skill. Quality systems define acceptable evidence. Feedback fixes recurring failure. The infrastructure is not overhead around iteration; it is what makes responsible iteration possible.

Track median idea-to-result time, percentage of experiments with predeclared thresholds, cost per accepted learning outcome, repeat-failure rate, time from evidence to decision, and rollback or corrective-action time. A team can then distinguish genuine learning speed from frantic activity.

You Cannot Delegate Founder Judgment

Hodak's most important warning is that founders cannot outsource judgment. Advisers can supply experience. Employees can own domains. Investors can identify patterns. AI can retrieve evidence, model scenarios, and challenge assumptions. None of them accepts the founder's accountability for the final company-level tradeoff.

This is not an argument for ignoring expertise. It is an argument against averaging advice until every decision resembles the market. Successful startups are long-tail outcomes. Copying an average process can be useful for commodity work; copying an average strategic opinion can erase the reason the company exists.

Hodak says action produces information. Once a team takes a reversible step, the world reveals constraints that discussion could not. The mature version of that rule is: act where the downside is bounded, instrument the action, and keep irreversible decisions behind stronger evidence and human review.

When to Build Internal Software and When to Buy It

Science built internal tools for purchasing, manufacturing data, recruiting, and reviews because its workflows did not fit ordinary software. Hodak argues that AI coding agents now lower the cost of creating tailored infrastructure. That expands the build option, but it does not remove maintenance, security, data governance, or continuity costs.

Start with a manual workflow and measure it. Build the smallest internal interface that removes the proven bottleneck. Preserve source data, logs, permissions, tests, export, and an exit path. Agent-generated code can make version one cheaper; it cannot make ownership free.

A 30-Day Startup Operating-System Audit

Week 1: purchasing and delay

Trace ten recent requests from need to order. Record handoffs, waits, rework, approval value, and cost of delay. Set budget bands and risk tiers, then remove one redundant approval while keeping audit logs and escalation rules.

Week 2: experiment economics

Choose three representative experiments or product releases. Allocate direct materials, labor, equipment, shared costs, quality work, failure, and lead time. Compare cost per accepted learning outcome and decide which protocol should change.

Week 3: hiring and feedback

Define four job outcomes, create anchored criteria, and build one realistic paid work sample. Measure funnel conversion and time by stage. Replace one vague performance question with a specific expectation, observed evidence, next action, and review date.

Week 4: decision loops and internal tools

Map one recurring decision from signal to owner. Add success and failure thresholds, a human gate, a worklog, and a review cadence. Only then decide whether a lightweight internal tool would remove enough repeated friction to justify ownership.

Copy-Ready Operating Worksheets

Experiment brief

Decision this experiment must inform:

Assumption:

Success threshold:
Failure threshold:

Direct materials:
Labor:
Equipment:
Shared facilities and consumables:
Quality / regulatory / external testing:
Expected failure and rework:
Lead-time cost:

Total expected cost:
Expected time to accepted result:

Owner:
Decision date:

Structured hiring scorecard

Role outcome:

Criterion 1: judgment
Observable behavior:
Evidence:
Anchored score (1-5):

Criterion 2: learning capacity
Observable behavior:
Evidence:
Anchored score (1-5):

Criterion 3: agency
Observable behavior:
Evidence:
Anchored score (1-5):

Criterion 4: role craft
Work-sample result:
Verification quality:
Anchored score (1-5):

Accommodation offered:
Conflicts or bias risks:
Decision and rationale:
Independent reviewer:

Founder decision record

Decision:
Accountable owner:

Evidence for:
Evidence against:
Unknowns:

Reversible or irreversible:
Maximum acceptable downside:

Options considered:
Chosen action:

What would change our mind:
Review date:
Observed result:
Operating rule updated:

Video Chapters

Bottom Line

Max Hodak's talk is not really about procurement software or a clever review algorithm. It is about the distance between a founder's intention and what the organization can do on Tuesday morning. That distance is filled by budgets, permissions, cost models, scorecards, quality evidence, feedback, and decision ownership.

"Average is not good enough" should not become permission for arbitrary processes or heroic founder instinct. The sharper lesson is to stop borrowing operating assumptions without testing them. Measure where learning stalls. Build the smallest system that removes the constraint. Keep costs visible, employment decisions fair, medical claims precise, and accountability human.

A startup does not need enterprise bureaucracy. It needs an operating system proportionate to its work. When that system is good, the company buys faster, learns what experiments cost, identifies talent with better evidence, surfaces problems earlier, and gives founders higher-quality information for the judgments only they can own.

Sources and Link Map

• Y Combinator: Max Hodak - Average Is Not Good Enough - the embedded Startup School 2026 talk and Q&A.
• Science Corporation - company, technology, patient, and research information.
• Science: PRIMA retinal implant - first-party description of the implant, glasses, and intended patient population.
• New England Journal of Medicine: Subretinal Photovoltaic Implant to Restore Vision in Geographic Atrophy Due to AMD - peer-reviewed PRIMAvera results.
• ClinicalTrials.gov: NCT04676854 - study design, sponsor, eligibility, outcomes, and publications.
• Science: preliminary PRIMAvera results - company summary of the 38-participant trial and letter-acuity results.
• Science: PRIMA CE mark application - June 2025 regulatory-submission update.
• Science: $230 million Series C - March 2026 financing, trial summary, expansion, and regulatory application status.
• Science: acquisition of Pixium Vision's PRIMA assets - technology origin, mechanism, and trial portfolio.
• U.S. EEOC: Employment Tests and Selection Procedures - job relevance, validation, protected groups, and adverse-impact responsibilities.
• U.S. OPM: Structured Interviews - job analysis, consistent questions, rating scales, and interviewer training.
• JQ AI SYSTEMS: Jensen Huang's Startup School Founder Mindset - reality, learning loops, systems thinking, and the limits of delegated judgment.
• JQ AI SYSTEMS: Sam Altman's Startup School Founder Playbook - ambition, agent leverage, founder networks, and safety boundaries.
• JQ AI SYSTEMS: Personal AGI for Founders - owned context, reusable skills, deterministic tools, governance, and a 30-day build plan.