Naval and Guillermo float opt-in innovation zones with different rules, and Max Hodak dismantles the idea for drug discovery specifically. Access pathways already exist: right to try, and single patient IND applications which the FDA approves over 99 percent of the time, sometimes over the phone. The binding constraints are elsewhere, in who physically holds clinical-grade drug and in the global inference a regulator draws from a bad outcome.

Right to try and single patient IND already grant access at over 99 percent approval
Dosing requires clinical-grade drug, which only the IP owner in an active trial holds
That owner will not share, because an adverse event contaminates their whole trial
Adverse inference is global, so a local innovation zone cannot contain the risk
Max's fix is narrow: prohibit the regulator from drawing adverse inferences across different users
Europe's notified bodies and China's CFDA are the two comparison systems worth studying

“an innovation zone would not solve the problem in drug discovery”
Max Hodak · 39:00

“still need clinical grade drug and the only entity with that is typically the IP owner who's in the middle of running a clinical trial”
Max Hodak · 39:30

#fda#drug-discovery#regulation#policy

Hot Take· 4

Hot Take01:30

The 1000x engineer is real, but token leaderboards do not measure it

Guillermo Rauch argues the job of an engineer has shifted from shipping output directly to building the factory that produces output at multiples. In intellectual and digital domains the spread was never 10x, it was 100x or 1000x, and choosing the right problem is an infinite multiplier on top. The new confusion is companies reading token consumption as evidence of leverage.

Judge engineers on whether they build systems that multiply output, not on the output itself
The 10x engineer claim was always understated in virtual and idea domains
Choosing what to work on is a larger differential than execution speed
Token spend is the new lines of code: a consumption metric masquerading as productivity

“the way that I'm judging you as an engineer is like are you producing the factory that would produce multiplicative outputs B through Z”
Guillermo Rauch · 01:30

“It's like the old measuring lines of code, you know, token consumption with lines of code feel like similarly not direct paradigms.”
03:00

#engineering#leverage#productivity#ai

Hot Take09:30

Is pure software dead, and where does the moat go now?

Naval raises the uncomfortable question: if models speak fuzzy human English and write the code, is pure software engineering obsolete and is a pure software company still investable? The answers on the table are that hardware founders are the big winners, model training and post-training may be the new software engineering, and the durable software value sits in reusable infrastructure building blocks that agents assemble rather than reinvent.

We learned to code to talk to machines; the machines now speak English back
Hardware founders benefit most, since good software is no longer the bottleneck
Mitchell Hashimoto's building block economy: agents need powerful reusable primitives
Shared dependencies have civilisational value even when a bespoke build would be marginally better
Existing infrastructure acts as a token cache so agents do not regenerate the universe

“is pure software dead like is pure software engineering like an obsolete thing”
Naval Ravikant · 09:30

“So where's the moat like for a founder? Hardware it's a boon”
Naval Ravikant · 09:30

#software#moats#infrastructure#agents#hardware

Hot Take29:00

The regulations are not the problem, and they might be the test suite

Max Hodak pushes against the Silicon Valley consensus that regulation is simply an obstacle: unsmoggy cities and swimmable rivers were progress, and the actual problem is the human cost of understanding and complying, plus months of latency every time you exchange a letter with government. Guillermo Rauch extends it: if agents succeed through explicit exit criteria, a regulation set is a rather good test suite, provided it is internally consistent and reasonable.

Overregulation is real, but many individual regulations encoded genuine progress
The binding cost is comprehension, compliance friction and months of round-trip latency
Regulations map cleanly onto exit criteria and guard rails for agent work
Blake's counter: pre-approval regimes are guilty until proven innocent, enforcement-based rules would be better
Naval's counter: this becomes a red queen race as agencies get flooded with agent-generated filings

“a lot of the regulations themselves are not the problem”
Max Hodak · 29:30

“So there's totally a world in where we say like the regulations are great. They're like our testing, our test suite.”
Guillermo Rauch · 31:30

#regulation#agents#policy#innovation

Hot Take44:00

Naval's third rail: make the first 20% of your income the deductible

Naval proposes an income-scaled healthcare deductible. The first 20 percent of your annual income is your deductible, so it is zero if you are broke and millions if you are rich, with insurance and government covering everything beyond it up to today's caps. The point is not the number, it is manufacturing a private market where patients pay directly and therefore vote with money.

Deductible scales with income, so it is zero for the homeless and large for the wealthy
LASIK, dental, veneers and plastic surgery advanced because they are private-pay
Without direct payers there is no feedback loop and no way to spend more into the system
Today wealthy patients cannot spend voluntarily because rate cards do not exist
Cash prices are sometimes quoted at 10x what the insurer is charged

“whatever your annual income is, the first 20% is your healthcare deductible”
Naval Ravikant · 44:30

“if you go shopping for medical care and you want to pay out of your pocket, sometimes they'll quote you a price that's 10x what…”
Naval Ravikant · 45:30

#healthcare#policy#markets#incentives

Explainer· 7

Explainer03:00

The model is roughly as good as you are in the domain

Max Hodak's observation is that frontier models mirror the capability of the person using them: a strong developer gets a powerful collaborator, a junior gets something junior-shaped. The small, sporadic corrections a user gives seem to disproportionately determine output quality. Guillermo Rauch adds that a new category of internal support has appeared, telling colleagues how to reprompt the model rather than solving the problem for them.

Frontier models reflect back the judgment the user brings to the conversation
Sporadic corrective feedback has outsized influence on output quality
Reprompting quality is now a coachable, high-leverage skill inside teams
Max expects this to fade as models get smart enough to need less from the user
Career progression for juniors is moving from writing implementations to choosing technologies

“if you're a really capable developer, then these things are really powerful”
Max Hodak · 03:00

“the feedback that you give them sporadically seems to be incredibly important”
Max Hodak · 03:30

#ai#engineering#prompting#management

Explainer06:00

Models stopped running away with your idea and started returning tradeoffs

Guillermo Rauch marks a concrete capability jump: models used to take a prompt and run with it in a straight next-token line. Now they come back unprompted with two or three routes and the tradeoffs between them, which is the behaviour of a principal engineer rather than a junior one. Their estimates are still frequently wrong, but the interaction has become a peer-level exchange.

Intuitive planning appeared without users having to ask for plan mode
Returning a set of routes and tradeoffs is what marks the graduation
Time and effort predictions from the model remain unreliable
Models now push back on bad architectural choices instead of complying
The open question is whether an experienced architect gets 10x while a junior gets 2x

“models now have been doing this like intuitive planning mode”
Guillermo Rauch · 06:00

“They used to be junior engineers now they're principal engineers because they come back to you with a set of tradeoffs”
Guillermo Rauch · 06:30

#ai#engineering#architecture#agents

Explainer18:00

China's open-source push is a hardware strategy, not a software one

Naval's read is that China is going all in on open models because it already holds hardware and supply chain superiority: if software becomes generatable on demand, its historic disadvantage against Silicon Valley disappears. Guillermo Rauch counters that the real chokepoint is frontier coding models, because without them you lose the ability to self-improve across every downstream domain. On actual usage, the panel reports frontier models dominate their gateway traffic.

Open models neutralise China's software disadvantage while preserving its hardware edge
The Chinese government has a history of funding ecosystem-wide, network-effect efforts
Without frontier coding models you lose self-improvement across the whole stack
Gateway data shows open model usage exists but frontier intelligence dominates the top
Naval's position: intelligence is an unalloyed good, so pay for the smartest model available
Gemini is singled out as the best industrial production model on cost and performance for non-coding tasks

“China is big into open- source models right they're basically going all in on it because they have hardware superiority”
Naval Ravikant · 18:30

“I think intelligence is an unalloed good you always want more intelligence and when these models make a mistake you don't know it”
Naval Ravikant · 20:30

#china#open-source#models#geopolitics

Explainer35:00

Why regulators block: the incentives are asymmetric and the voters agree

Max Hodak lays out the structural problem. Approve ten important drugs and nobody gives you credit; one patient dies and you are hauled before Congress. That asymmetry produces a systematic bias toward blocking, and it reflects genuine public preferences about risk in human subjects research. Naval pushes the point further: politicians are elected, so this is where the citizens actually are, and people cannot see the innovation they never received.

Approving something bad ends a career; blocking something good goes unnoticed
This is the single most important problem to solve in the regulatory state
The preference is real and polled, not an elite conspiracy
If you are seen as a bad actor working around it, society rejects you
Voters cannot perceive the counterfactual prosperity they gave up
The NRC example: nothing is safer than a nuclear plant that was never permitted

“One patient dies and they get hauled before Congress and yelled at.”
Max Hodak · 35:30

“If you approve a bad thing, your career is over. If you block a good thing, nobody notices, right?”
Naval Ravikant · 36:00

#regulation#incentives#policy#fda

Explainer41:30

Healthcare's fixed bucket problem, and why laptops are the contrast

Max Hodak's sharpest structural argument. Twenty years ago we bought fewer, pricier laptops and phones; now they are cheaper, better and we spend more in total, and everyone is happy. Healthcare cannot do this, because reimbursement makes the bucket of money roughly fixed and tied to tax receipts. Spending ten times more on AI in two years would be exciting; spending ten times more on healthcare would be a catastrophe, which is fundamentally at odds with being a growth industry.

In technology, better products expand total spend and everyone benefits
Reimbursement turns healthcare into an enterprise sale with a fixed budget
Healthcare spending grows at roughly the rate of tax receipts, not of capability
New capabilities like restoring sight or extending life have no funding path
The way out is lowering the cost to bring things to market, not single payer
The restaurant analogy: send every receipt to your insurer and quality stops improving

“the bucket of money that we use to buy healthcare is basically fixed”
Max Hodak · 42:00

“if in two years we're spending 10 times as much on healthcare, this would be a catastrophe”
Max Hodak · 42:30

#healthcare#economics#reimbursement#innovation

Explainer55:30

A 100x increase in coders still leaves 99% of people never writing code

Naval reckons the share of people writing code has gone up perhaps 100x, from around 0.01 percent of the population to maybe 1 percent, and Guillermo confirms signups are through the roof with a new class of users who are not engineers. But the majority never start, because to them software was always a black box, so nothing visibly changed. Meanwhile vibe coding is displacing video games for those who did start.

Roughly 0.01% to 1% of the population, a 100x increase, with 99% still never coding
A new user class has appeared that uses infrastructure without being engineers
Non-coders do not perceive the change because they never understood the old process
Vibe coding replaced FPS gaming for Naval: same feedback loop, real output
The addictiveness is described as lottery-like and unlike programming for a decade

“I'll bet you we all know a lot of people now who are coding who weren't coding before including many cases ourselves right”
Naval Ravikant · 55:30

“So we might have taken you know 0.01% of the population writing code to maybe now it's 1% call it a 100x increase but 99%…”
Naval Ravikant · 57:00

#vibe-coding#adoption#software#culture

Explainer59:00

What humans uniquely do: surprise, intent, and stepping outside the system

The philosophical core of the episode. Naval argues creativity is what surprises you from outside the system, out of the training distribution, and offers Godel's incompleteness theorem as the archetype of stepping outside a formal system to break it. Max defines art as meaningful out-of-distribution behaviour, where meaning is anything that changes your future trajectory. Naval's competing definition centres intent: art transmits a felt emotion, so attribution matters and a computer is almost definitionally excluded.

Max's definition: art is meaningful out-of-distribution behaviour
Naval's definition: art conveys an emotion one person felt to another, so intent is load bearing
Attribution matters; the same photo means more if a human took it
Once a style saturates, like Ghibli imitations, it is in-distribution and stops surprising
The framing is not human versus computer but human with computer versus computer alone
Open question: can humans go out of distribution without randomness, and where do new ideas come from

“I think that creativity is still the thing in the environment that surprises you. You step out of the system and do something that wasn't…”
Naval Ravikant · 59:30

“It's not going to be human versus computer. It's going to be human with computer versus just computer.”
Naval Ravikant · 1:00:00

#creativity#art#agi#philosophy

Story· 3

Story22:30

Why Science bought its own MEMS foundry

Max Hodak explains that vertical integration at Science is a consequence, not a preference. They would rather buy, and do buy commodity parts like PCBs cheaply from Asia, but the components needed to push their brain interface toward a single block of covalently bonded matter simply do not exist for sale. Owning a captive MEMS foundry on the east coast was the only way to do the packaging and assembly they needed.

Buy wherever a vendor exists at a good price; PCBs are effectively free
Integration density drives lower power, smaller size, higher performance and longevity
Assembling only off-the-shelf parts caps how far you can innovate
The foundry is instrumented so improving models show up immediately in cell engineering and materials work
The largest current AI impact inside the company is regulatory: tracing which of thousands of ISO standards apply

“the closer that our products get to being like a single block of coalently bonded matter the better they'll be lower power smaller higher performance…”
Max Hodak · 23:00

“we own a captive MEMS foundry on the east coast which we bought because there was really no other way to do the type of…”
Max Hodak · 23:30

#hardware#vertical-integration#manufacturing#biotech

Story27:30

Certifying an airplane: 200 pages and two months becomes minutes

Blake Scholl gives the concrete case. Proving an aircraft can withstand a lightning strike requires a test plan document running to roughly 200 pages, historically written by an engineer hired to be a keyboard monkey over a couple of months, with full rework every time the design changed. Boom built a retrieval system that prompts its way through the work in minutes, and the third-order effect matters more than the time saved.

First-order effect: the documentation time collapses
Second-order effect: changing the aircraft spec no longer costs months of rework
Third-order effect: change aversion disappears and iteration rate rises
You can then staff a small number of creative engineers instead of a large compliance team
Effectively, much of the regulatory burden on iteration speed drops away

“the regulatory documentation for the test plan for such a thing stretches on for say 200 pages”
Blake Scholl · 28:00

“we can build a rag that will enable us to basically prompt our way through all of that work you know in let's call it…”
Blake Scholl · 28:30

#regulation#aerospace#rag#iteration

Story45:30

Sid from GitLab and the case for n-of-1 medicine

After a successful IPO, GitLab's Sid was diagnosed with a rare cancer, exhausted frontline chemotherapy and the one alternative available, and was told there was nothing left. He took it into his own hands and built a personalised treatment plan; six or seven companies have come out of it, there are now twenty or thirty drugs in his escalation ladder, and he has lived far past prognosis. The panel treats this as a preview and names its ugliest constraint.

Escaping the standard pathway required resources and refusing to deal with insurance
Six or seven companies and twenty to thirty candidate drugs emerged from one patient's effort
n-of-1 medicine may become a rich research source for translatable treatments
The cruel irony: it demands maximum agency from a patient at their weakest
The gap is knowledge access as much as money, and that is where AI should shine

“there's now 20 or 30 drugs in his escalation ladder. He's still alive.”
46:00

“It requires a ton of agency from the patient in a moment where they're at their weakest”
Guillermo Rauch · 47:00

#healthcare#cancer#personalised-medicine#agency

Takeaway· 2

Takeaway24:30

Junior engineering got automated; junior engineers got promoted

Naval notes he has stopped using lawyers for routine documents entirely, because basic legal work is now generated. He reframes the labour story: the same event can be read as paralegals being fired or as paralegals being promoted to senior lawyers who now spend their time thinking about the law. The parallel with software is exact, and both professions rest on trusted authority rather than inspected work product.

Routine NDAs, agreements and legal research no longer justify a lawyer
Law is described as spaghetti code written in English with no APIs
The same shift reads as firing or as promotion depending on framing
What you buy from a lawyer is a trusted authority putting their reputation on the line
That trust relationship is exactly what code review must now replicate

“it's been a while since I've generated a basic legal document using a lawyer”
Naval Ravikant · 24:30

“junior engineers basically got a promotion to senior engineers and junior engineering got taken over by agents”
Naval Ravikant · 25:00

#labour#legal#automation#careers

Takeaway1:07:30

A larger number of much smaller teams, and what to do about it

Blake Scholl's hypothesis is that headcount per task collapses, so the first-order read that 998 of 1000 jobs vanish is wrong; you simply get many more jet engines, many more companies. The panel converges on an explosion of entrepreneurship built on very small teams, where credentials and memorised expertise depreciate and creativity, taste and judgment do not. The closing advice is concrete and unglamorous.

Fewer people per task means more tasks attempted, not fewer jobs
Expect an explosion of founders and a large number of very small teams
Hiring bar shifts to juniors and super seniors who are genuinely good with agents
AI supplies base intelligence and domain knowledge; agents supply much of the agency
Generalists benefit, since you no longer need twenty years in a field to contribute, while credentialed expertise without taste or judgment is the most exposed
Best single action: get good with the tools and learn where their boundaries currently sit

“My my hypothesis is we end up with a larger number of smaller teams.”
Blake Scholl · 1:07:30

“the single best thing you can be doing right now for yourself is just getting really good with these tools”
Naval Ravikant · 1:09:30

#entrepreneurship#hiring#careers#leverage

Frameworks in this episode

Agent-First Incident Response

Sign Off on Consequences, Not Lines

The Company-Wide AI Build Week

The Internal Software Factory

Waste Tokens, Save Time

Insights & moments

⚡Myth Buster· 1

Why an innovation zone would not actually fix drug access

◆Hot Take· 4