NNaval
← All episodes
02 July 2026

Live in the Future

5Frameworks
15Insights

Listen

Frameworks in this episode

Strategy6 steps

Live in the Future, Then Work Backwards

Buy tomorrow's cost curve today, watch what becomes possible, then build backwards.

Read →
Innovation6 steps

LSD Mode: The Vector-Space Collider

Cross unaligned vector spaces over your own corpus, then re-rank the collisions for ideas.

Read →
Productivity6 steps

The Compounding Error Rule

Buy maximum accuracy wherever steps loop — error compounds, price does not.

Read →
Communication6 steps

The Consumer Test for AI Output

Ask who reads it: machines get generated output, humans get your compressed thinking.

Read →
Leadership7 steps

Total Information Awareness for Operators

Wire an agent into every company system so the founder sees each team unfiltered.

Read →

Insights & moments

The myth-busts, hot takes, explainers, and tools worth keeping.

Myth Buster· 1

Myth Buster

The data-centre backlash was never about water

The panel dissects the attacks on data centres as a stated-reason-versus-real-reason problem. The water argument, they say, was debunked; it functions as a socially acceptable cover for a rational underlying fear of displacement. Naval's point is that the stated grievance is interchangeable — remove water and the movement finds an endangered bird instead — because the driver is the fear of being replaced, not the specific environmental claim.

  • The water objection was debunked but persists as the public reason
  • The underlying driver is fear of replacement, which the panel calls rational
  • The stated grievance is fungible — take away water and another one appears
  • Framed as a modern version of smashing the spinning looms
  • Loom-smashing did briefly slow adoption, so dismissing the backlash as futile is too easy

“So they're saying it's because of water. It's not about water. That was a lie that was debunked.”

“They'd find another one if it wasn't water, then it'd be some endangered egret”

#backlash#data-centers#luddism#politics

Hot Take· 4

Hot Take

90,000x more inference in three years — and why Nvidia may be underpriced

Gary Tan lays out the compute forecast the rest of the conversation rests on: roughly 90,000 times more inference available within twenty-four to thirty-six months, already baked into the chips and data centres being built. The panel notes this is around five orders of magnitude, and that each order of magnitude has historically produced new emergent capabilities rather than just more of the same. The provocative corollary is that the consensus view of Nvidia being overpriced may have the sign backwards.

  • Roughly 90,000x more inference within 24-36 months, or about five orders of magnitude
  • The build-out is already committed via chips and data centres
  • Every order of magnitude historically unlocks new capabilities, not just volume
  • If the curve holds, Nvidia may be underpriced rather than overpriced
  • Naval admits he missed AI from 2020 to 2022 by assuming it was perpetually ten years away

“we think like 90,000 X or so, there will be 90,000 X the amount of like inference from here to like three years from now”

“a lot of people say that Nvidia is way overpriced, but what if it's way, way underpriced by like several orders of magnitude?”

#compute#scaling#nvidia#forecasting
Hot Take

Gatekeeping has already started — and concentration is scarier than diffusion

The panel treats nationalisation as a process already underway rather than a hypothetical. They cite a frontier release being rolled out first to partners approved by the US government, and read it as a lab pre-emptively cooperating to avoid being shut down. Naval's position is that a small group controlling AI is far more dangerous than everyone having it, because control brings cyber capability, surveillance and the usual 'for your own good' justifications.

  • A frontier model release gated behind government-approved partners is cited as the tell
  • Read as pre-emptive cooperation by a lab that fears being shut down
  • Scaremongering about safety is the mechanism that produces licensing and approval
  • Naval: concentrated control is scarier than universal access
  • Counterpoint acknowledged: universal access has its own tail risks, though bio-printing labs remain locked down

“We just saw Sam Altman said that codex five point six will be rolled out slowly to partners approved by the US government first.”

“it is definitely much scarier that a small group of people have control over AI than everybody has it”

#regulation#nationalization#open-source#policy
Hot Take

Google lost it — and the diagnosis is PM slop

The panel's verdict on Google is blunt: despite a genuinely strong model moment, they never got the basics working. The examples are mundane and damning — a painful upgrade path across overlapping subscription products, and a mobile app that drops long-running queries when backgrounded while every competitor handles it. The structural diagnosis is organisational: too many overlapping products, too many product managers, a maze users cannot navigate. Naval extends the same critique to search, saying he now reads only the AI answer.

  • A strong model moment did not translate into a working product
  • Overlapping subscription tiers made paying them actively difficult
  • Long queries drop when the mobile app is backgrounded
  • The panel's diagnosis is organisational overlap, not model quality
  • Naval says he no longer looks at search results, only the AI answer

“Google, I think, has lost it.”

“It's PM slop. The whole company's PM slop at this point”

#google#product#org-design#gemini
Hot Take

Naval's radical take: there is no real competition with China

Naval argues the US is not actually in the competition it thinks it is. Taiwanese elites are dodging conscription and the second-largest party is pro-China; most people there expect a Hong Kong-style path of getting rich and assimilating over a generation or two. Militarily, land-based missiles and drones make carriers vulnerable, and the recent Iran episode showed thin missile stocks and no manufacturing base. His conclusion is that everyone is saving face while reunification happens slowly, and that picking a war while the US has low growth, high inflation and internal political conflict is the worst available move.

  • Rich Taiwanese families structure their children's travel to avoid conscription
  • The second-largest Taiwanese party is pro-China, so the position is already compromised
  • Carriers are vulnerable to land-based missiles and drones
  • Naval argues the US would run out of missiles within about a week
  • He supports symmetric tariffs and anti-subsidy barriers, but not conflict
  • Ricardo's simple trade story fails where scale economies and network effects exist

“I think everyone's trying to save face while like over 10, 20 years, Taiwan slowly reunites with China.”

“there's no reason for us to go to war with China”

#china#taiwan#geopolitics#trade

Explainer· 5

Explainer

The last creative mile — and whether the centaur model survives

Naval reframes the open question about AI capability. Two years ago the plausible ceiling was that AI makes you mid at everything; now it plainly makes you a pro at most things, and the remaining unknown is the last creative mile — stepping outside the system to make something genuinely new rather than recombining the training set. Gary raises the sharper version: chess went from human-plus-machine beating machines to machines beating the centaur, and it is unclear whether knowledge work follows.

  • The old ceiling was 'mid at everything, pro at nothing' — that has already fallen
  • The remaining question is originality outside the training distribution
  • Progress on a famous open maths problem unsettled the panel's mathematician
  • Chess is the cautionary precedent: the centaur model eventually lost to the machine alone
  • Naval's counter: spend limited time applying your humanness, not opining on the machine

“Now you're getting to the point where it'll get you the pro at everything, but we don't know if it'll get you the cre the…”

“are we gonna end up in a place where a machine is strictly better than a machine with a human?”

#agi#creativity#human-in-the-loop#chess
Explainer

Four theories for why the Chinese open models are catching up

Naval lays out the competing explanations without picking one. First, genuine independent pre-training on a fuller web corpus unconstrained by copyright. Second, mass distillation of American models, aided by grey-market resale of subsidised consumer plans. Third, weights leaking, since labs do not have the security posture of a national security facility. Fourth, real algorithmic breakthroughs plus a talent pool that dominates the relevant PhD and Olympiad pipelines. He also floats a policy fix: if you trained on the open web, you should have to open your model after twelve months.

  • Independent pre-training with fewer copyright constraints on crawl scope
  • Distillation at scale, funded partly by resold subsidised consumer plans
  • Weight leakage from labs without national-security-grade security
  • Genuine algorithmic contributions plus dominance of the STEM talent pipeline
  • Naval's proposal: train on the open web, open your weights after twelve months
  • Researchers move between labs constantly and take knowledge with them

“I think it's rich of anthropic to call that out when anthropic crawl the open web and distill the open web.”

“if you train on open web and open data, you have to open your model after X months, after 12 months or something”

#open-source#china#distillation#policy
Explainer

Software ate the world, AI ate software — so what is left uncommoditised?

Naval's central strategic argument. Hardware was always a commodity, which is why VCs demanded a software moat; his old line was that hardware buys you time to build a software castle. Coding agents burned the castle down: once you can specify software, a model can produce it, so software is commoditised too. Hardware remains commoditised and is largely owned by China's supply chains. What is left is AI research itself — and unlike software, it is not democratised, because it needs enormous clusters, proprietary data and, increasingly, regulatory permission.

  • Hardware was the moat that bought time to build a software castle
  • Coding agents commoditise software the moment it can be specified
  • China owns the hardware supply chain via scale economies and subsidy
  • AI research is now the only uncommoditised layer, and it cannot be done in a garage
  • China's labs are told they need not profit, because commoditising software pushes margin to the hardware they own
  • Naval's irony: Western access to capable open models is underwritten by that subsidy

“Software was eating the world, and AI ate software.”

“So what's not commoditized? It's actually just AI research.”

#commoditization#moats#hardware#strategy
Explainer

Five kings became two: why OpenAI and Anthropic pulled away

Two years ago the panel would have named five credible frontier contenders. They now count two, and give two structural reasons rather than one. First, both earn revenue directly from their models rather than cross-subsidising from another business, so cash recycles into the next model instead of being borrowed from an ad or hardware line. Second, large active user bases feed reinforcement learning with the guidance and trajectories that drive improvement — a compounding advantage. Elon is granted a possible extra bite via a SpaceX war chest.

  • The field narrowed from five plausible contenders to two
  • Direct model revenue means no cross-subsidy dependency
  • Active user bases supply RL guidance and trajectories that compound
  • Elon is given one more shot on the back of SpaceX capital and orbital data centres
  • The panel's verdict on Google's position is notably harsher

“There were five kings. There's two kings now.”

“OpenAI and Anthropic are the only ones that are A, making revenue off of their models directly”

#frontier-labs#competition#business-models#rlhf
Explainer

The 2027 harness wars — and why a duopoly protects startups

Gary predicts that the defining fight of 2027 is the harness: which agent shell people actually live in day to day. Naval's structural point is about how many winners there are, not who wins. Silicon Valley's mobile era worked for startups because two platforms existed and had to behave; a single platform owner at the top of the chain would have been a true monopolist. If the harness layer collapses to one, that is a bad outcome for startups — and nationalising it would slow it down and hand the lead abroad.

  • The harness, not the model, is Gary's pick for the 2027 battleground
  • Two mobile platforms were what made the post-2007 startup wave possible
  • A single winner at the top of the chain is a true monopolist over startups
  • Naval doubts nationalisation, noting search and the big platforms were never nationalised
  • A nationalised lab would slow down and lose to faster competition abroad

“I think that like 2027 will be the year of the uh AI harness war.”

“If it boils down to just one, it's a very bad situation for startups.”

#agents#platforms#competition#startups

Story· 1

Story

Cost, not intelligence, is the bottleneck: $100 a month to $2.84

One of the founders on the panel runs a dedicated agent for every single user of their product, which forced them to confront unit economics before capability. Starting on a frontier model, each user cost around a hundred dollars a month. Three to four months of engineering — an eval harness plus an elastic agent fleet that spins up and down — drove that to two dollars and eighty-four cents. His conclusion is that intelligence is not the constraint right now; cost is.

  • Every user in the product gets their own persistent agent
  • Initial cost was roughly $100 per user per month on a frontier model
  • Three to four months of work brought it to $2.84 per user
  • The stack is an eval harness plus an elastically scaling agent fleet
  • The framing: intelligence is available, affordability is the open problem

“I don't think intelligence is the bottleneck, cost is the bottleneck”

“We spent three, four months driving that down to two dollars and eighty-four cents.”

#unit-economics#agents#infrastructure#cost

Takeaway· 4

Takeaway

Once open source takes the lead in a category, it rarely gives it back

Drawing on Linux and other precedents, the panel argues that open-source leadership is unusually sticky, because an ecosystem forms around whatever is winning — especially in enterprise. Applied to models, they note some open categories are already number one, notably video. The strategic consequence for a closed lab is a clock: you must use finite cash and investor backing to jump meaningfully ahead again before the money runs out, and pouring precious resources into merely catching up to free is a bad use of them.

  • Ecosystem gravity makes an open-source lead hard to reverse
  • Some open models are already category leaders, video being the clearest case
  • Closed labs face a clock: get decisively ahead before cash runs out
  • Spending scarce resources to merely match free output is hard to justify
  • At the time of recording, the leading closed coding models were still ahead

“once something open source kind of gets in the lead, it rarely surrenders it”

“failing to catch up to open source with your precious resources is not a good use of them”

#open-source#ecosystems#competition#linux
Takeaway

You cannot buy a culture — the Meta critique

On Meta's aggressive acquisition of AI talent, Naval's objection is not the price but the assumption. Buying people from many different places and ramming them together does not produce culture or morale, and leaves you unable to tell missionaries from mercenaries. Culture is a slow-growing asset. The panel's constructive alternative is the same fine-grained visibility discussed earlier: know what is actually happening per person and per team, and tune deliberately, instead of blunt company-wide moves that read as indiscriminate.

  • Buying talent in bulk does not produce culture or morale
  • Mixing many origins at speed obscures who is a missionary and who is a mercenary
  • Culture takes time to evolve and cannot be shortcut with compensation
  • Blunt company-wide reorganisations read as indiscriminate and leak badly
  • The alternative is per-person, per-team visibility and deliberate tuning

“I don't think you can build a culture or morale by just buying people”

“Culture is a real thing that takes time to evolve and grow.”

#culture#meta#hiring#leadership
Takeaway

The bitter lesson for vertical AI: general models beat the specialists

The panel's scariest question for founders is whether a startup ecosystem survives when software is commoditised. Small teams can now reach large revenue with tiny headcount, which argues for more startups. But the counterweight is the bitter lesson: a general frontier model with good tool use tends to beat a model trained specifically for a vertical. The labs do not even need to launch a vertical product — customers simply stop paying for the thin wrapper once the general model does the job.

  • Commoditised software means small teams can reach large revenue
  • General models with tool use tend to beat vertical-specific training
  • The frontier lab does not need to enter your vertical to take it
  • Vertical defensibility depends partly on whether labs keep their best models open to paying customers
  • The buyer's question becomes why pay a large monthly fee for a wrapper

“one of the less one of the bitter lessons here is that the general models beat the specialized ones”

“Well, I think the big scary question is will there still be a startup ecosystem?”

#vertical-ai#bitter-lesson#startups#defensibility
Takeaway

AI handlers, not the unemployed — plus universal basic robot

The panel's closing optimistic case. If AI reaches expert level but not superintelligence, humans stay in the loop as the motivated agents supplying guidance, taste and creativity — closer to Pokémon trainers directing AIs and robots than to displaced workers. Naval's observed evidence is that people using AI today have more work than ever, and that the only people being displaced are the ones refusing to use it. The material version is universal basic robot rather than universal basic income: everyone gets a robot that does the work nobody wants.

  • If AI stops at expert level, humans remain the guidance and taste layer
  • The role is a handler or trainer of AIs and robots, not a manual labourer
  • People using AI today report more work, not less
  • Displacement correlates with refusal to adopt
  • Human desire is the one input AI cannot supply, so humans stay in the loop
  • Precedent: US farm labour fell from about 50% to 2% without 48% unemployment — the open question is the speed of transition
  • Universal basic robot is preferred to universal basic income; robotics is 2-10 years out depending on who you ask

“The only reason you're being displaced by an AI is because you refuse to use the AI.”

“We should do UBR. UBR is good. We need to have universal basic robot.”

#future-of-work#robotics#ubi#adoption