The meme says "nothing ever happens." Naval Ravikant thinks that's over, though he admits he hasn't pinned down why.
His best guess is that the world was in an unstable equilibrium and COVID broke it, producing a phase shift. Whatever the cause, he sees faster movement geopolitically, economically and technologically. Venture capitalists are being pushed to fund hardware, rockets, drones and other sci-fi technologies, while sci-fi engineers and scientists are in short supply. He calls it the Chinese curse made literal: may you live in interesting times.
Nothing ever happens has stopped being true
The sharpest open question, in his view, is the structure of the AI industry. Two to five companies (counting Nvidia) dominate it. Will that settle into a commodity business, an oligopoly or a monopoly? Does the data run out and the models plateau, or does the path run to AGI? Can training be distributed, or does it keep centralizing around data centers and power?
Naval says the emerging consensus on centralization looks right to him, though a distributed-training breakthrough would be an interesting contrarian bet. He doesn't see the evidence yet. On AGI he stays agnostic: the people inside the frontier labs have long believed in it, but the AI he sees has jagged intelligence and handles multimodal reasoning poorly.
He also draws a line around the term "world model." Systems that generate a navigable, world-like environment don't qualify. A real world model is an agent that holds a model of the world internally, takes actions, predicts their consequences and adjusts its behavior, much like a reinforcement learning loop. Yann LeCun's recent JEPA work points in that direction, and Naval expects new kinds of models and agents.
Impossible runs as a fully interconnected graph
At Impossible, his current company, there is no Slack and no project management software beyond GitHub. When people need each other, they text one on one. His co-founder, the CEO, acts as a light hub, the one product manager who carries the whole project in his head. Everyone else is encouraged to talk directly to whoever can solve their problem.
Naval frames this in networking terms. Communication overhead grows with size, and the traditional answer is a tree: a CEO, VPs, middle managers. It keeps everyone marching in one direction, and it also breeds politics and walls off anyone more than a level or two away. He admits to hating organizational management, and says he prefers small groups for the same reason.
The alternative is the fully interconnected graph, where anyone talks to anyone. It has a price, which is that every node must be highly intelligent.
If they can't navigate their way to the person they need to talk to, or if they can't cooperate or communicate with other people, then they don't belong in this kind of organization.
Those people, he says, should find a hierarchy where they'll be more comfortable. He's explicit that this is a requirement of size: the tree becomes necessary at scale, and Impossible isn't there.
AI makes the intranet and the org chart optional
The company hasn't reorganized around AI the way Jack Dorsey has at Square. Naval says AI helps implicitly, and gives two examples.
The first is retrieval. AI can read someone else's complicated code or a research paper and summarize it, or scan a code base and name who in the company is likely the expert on a topic. That removes much of the need for an explicit intranet or carefully maintained documentation. Pointed at hardware designs, supplier folders or company email, it can answer "how far are we from shipping?" and draw a Gantt chart of who's ahead and who's behind. Dashboards and business integration systems become reports generated on demand.
The second is that AI lets specialists reach 20 to 30 percent of each other's work. AI researchers can build their own test harnesses instead of waiting for a software engineer. Hardware engineers can write enough code to bring up a new device. Naval's conclusion is that everyone becomes a bit more of a generalist, which creates more touch points between people. An AI can discover an API, create one, or skip it and connect at the database level.
Drones push mutually assured destruction down to individuals
Naval thinks drones are still under-leveraged and that we're nowhere near their end state. Defense looks hard. An attacking drone has kinetic energy and surprise, and attackers can mass their drones in one place while defenders are spread thin. The defender's one edge is short range.
His larger argument is historical. The rifle let a former peasant bring down a feudal knight, which required factories and drilled musket infantry, and nation states displaced feudal ones as the right structure for organizing violence. After 1945, nuclear weapons became the new logic: seven to nine truly independent sovereign nations, with everyone else under someone's nuclear umbrella. Drones, he says, bring that mutually-assured-destruction logic down to the individual. If you hate someone enough, a drone will eventually be able to reach them.
He doesn't know which way it breaks. A few powerful countries might control all the drones, or drones might spread so widely that any individual can be deadly.
Biology faces the same pattern, with a regulatory handicap
Naval applies the same reasoning to biological weapons. Smart people could always have worked one out, but few had both the expertise and the access. Just as vibe coding multiplied the number of people who can write software by hundreds or thousands of times, AI could multiply the number who can reach dangerous biology. He says he doesn't want to alarm anyone, and calls it a pretty scary thought.
The same AI could help design vaccines and countermeasures, but he argues the good-faith research is gated by regulation, and few regulations are as bad as medical ones. His prescription is anonymized, cleaned-up data sets released at scale, plus a right to try for volunteers. He fears this will only happen in an emergency, and even COVID didn't force it. He asserts that the vaccines "turned out not to be that effective anyway" and blames bioethicists and bureaucracy for slow trials, but offers no evidence for either claim in the conversation.
Software stops being the bottleneck for hardware
Hardware's chronic weakness has been software. Apple succeeded by doing both well, while most companies manage one or two things. Google is strong in cloud and AI and weak in hardware.
That constraint is loosening. A maker of security cameras, kids' toys or programmable lamps can now have a bright kid with Claude Code write the software. Or the maker may skip custom software entirely, because a user's AI agent can interact with the device directly.
Naval sees incentives pushing the same way. China is behind in AI models, and it dominates consumer electronics manufacturing, so open-source models commoditize its complement. Nvidia wants as many models running on its cards as possible. Hyperscalers lean the same direction. Open models commoditize software, software unlocks hardware, and he expects more interesting, usable hardware as a result.
Doom is easy to imagine because optimism takes creativity
Naval says he doesn't get worked up about the future, partly because he's a blind optimist and partly because he lives in the first world. He offers a mechanism for why pessimism feels more rigorous than it is.
Doom is legible. Job loss is the clean example: you can look at an existing job and see exactly how it disappears, but you can't see the next job before it exists. Nobody 200 years ago, when everyone worked a farm, could have imagined 10 percent of today's occupations. Every decade he's been alive has had its environmental catastrophe and its war that was supposed to end everything.
He grants that some of those scares came close. COVID could have been far nastier, and a nuclear exchange would be catastrophic. The point is that methods of destruction are easier to picture than methods of rising up, so attention goes to them.
His answer is to treat optimism as something to cultivate and reward, even irrationally, because he sees no other way out. The doom-sayers "might be right, but it's certainly not helping matters." Then comes the line that closes the conversation:
That's not the person you want to be in a foxhole with.
