a16z's Frontier fund bets AI may pull 20 years forward

1. Crypto stops being separate — ▶ 0:55

   The fund expands from a16z crypto to a16z Frontier because crypto is no longer a separate world. Stablecoins are a normal part of the tech toolbox, used by companies like Stripe and nearly every fintech, while Wall Street firms are moving heavily into tokenized stocks and other assets. The lines have blurred so far that the firm asks whether Stripe is a crypto company. Chris Dixon gives a second motive: the other groups are so busy with core opportunities that there may be a coverage gap at the edges.

2. Outer planets, second-order effects — ▶ 2:15

   Martin's term, borrowed from Firefly, separates the inner planets (core areas) from wackier outer ones. Dixon joined in 2013 hunting second-order effects. Smartphones went mainstream and first-order products were apps like Instagram. The second-order effect was repurposed sensors and chips, which fed drones, Oculus and Skydio. Ben Horowitz says the vertical strategy works but new things fall outside it, and turning them into a practice is "a little bit undefined". Frontier fills a hole in the architecture and could become a full fund.

3. Good ideas that look bad — ▶ 8:26

   Horowitz credits Dixon with the firm's most original investment idea. Do not look for good ideas that look like good ideas. Look for good ideas that look like bad ideas, where the entrepreneur holds a secret that makes them work. Ideas that look good get their profits competed away, while the secret-backed ones carry massive upside. The firm "copied" and multiplied the framework across its practices. Dixon was hired for his blog writing and his experience as an entrepreneur, which mattered for setting the firm's culture.

4. Err on the side of early — ▶ 15:34

   Dixon believes every good idea involving technology has eventually worked, usually after a precursor 10 years too early. Frontier will accept being too early and take category risk. Oculus sold, yet 13 years later VR is not as mainstream as hoped. AI changes the frame. Software engineering tipped first, math next, and physics and hardware are coming, so the firm may be pulling forward 20 years. A 30-person team can now credibly build a full-stack device company that once needed 1,000.

5. Widest gap ever — ▶ 18:19

   Andreessen calls it the biggest delta he has seen between what early adopters and normal people think is possible, and it is widest at the company level. AI-native startups differ from incumbents "almost without exception". With these tools at Netscape in 1994, he and Horowitz say they would have built the whole internet, streaming video sooner and crypto in two weeks. The question shifts from whether an idea is too ambitious to whether it is ambitious enough. Incumbents' clock is ticking, and he is drowning in the pace of change.

6. Individuals adopt, institutions lag — ▶ 22:32

   Many companies are not adopting AI while their employees quietly are, so individuals move faster than institutions. Some will start their own businesses, and startup formation is already rising. Andreessen thinks Christensen's disruption frame is wrong because it assumes a contained tech world. The real contest has moved from 10 big tech companies fighting each other to those 10 plus all startups going after the rest of the Fortune 500 economy, including full-stack food service. Demand is still early, around 3% of measured enterprise use.

7. Old behaviours, then new ones — ▶ 26:47

   New platforms first do familiar things better. Sears put its catalog online in the 1990s, and YouTube launched as an empty site betting on influencers, a behaviour change. AI is mostly in the first phase of coding, math, knowledge work and search, but is near a tipping point. Dixon expects agents to become the most numerous participants on the internet, via proxy products like Muse and Instinct. AI is still single-player. Swarms and trust networks hint at a multiplayer phase that could be fertile ground for entrepreneurs.

8. Product manager brain — ▶ 30:26

   Andreessen says early AI products suffer from "product manager brain": built by product managers, for product managers, to book dinner with other product managers. The weirder behaviour is arriving anyway. In three weeks there has been a wave of AI-made music videos and tools that mash any two video games together. REA, reverse engineer anything, points a model at a thermostat and rewrites its code, turning hardware into a software problem. A $180 million film that lost $200 million would have returned 10x at $2 million.

9. Open source as the guardrail — ▶ 38:22

   Against Peter Thiel's line that AI is centralizing, the partners argue for open source. Open models trail closed ones by 6-12 months, open source token consumption is growing quickly, and local models are coming. Stablecoins are the natural way for agents to pay each other. Andreessen says without open source AI centralizes to one or two companies or the government, and he fears governments turning hostile. Horowitz adds that reward hacking is the biggest safety problem and the labs have not worked on it hard. Every PhD student should see the weights.

10. The internet covenant breaks — ▶ 46:42

    Dixon describes a covenant between distribution (Google, Facebook) and content: give Google a snippet, get clicks back. It is breaking. Google now gives the answer and social networks keep users on-platform. Stack Overflow's traffic dropped precipitously once AI replacements arrived. Businesses built on ads and conversions will suffer, although transactions will still happen, perhaps via agents. He expects a decade-long reconfiguration of economics, attention and collaboration. It will reshape things more than mobile did, which took the internet from 500 million to 5 billion people.

11. Study modern alchemy — ▶ 54:14

    Asked what an 18-year-old should study, Dixon borrows Andreessen's line: study modern alchemy, because humans just figured out how to turn sand into intelligence. That means math, computer science and maybe economics. He expects economics, classics, biology and animal communication to be transformed. Douglas Adams's rule applies: under 15 it is natural, 15-35 it is exciting, over 35 it is unholy. Schools banning AI make it cooler. Dixon's closing stance is "humble definite optimism": back entrepreneurs, expect to be wrong often.

AI has handed small teams a capability that used to require armies. A 30-person team can now credibly build a full-stack device company that would have needed 1,000 people a decade ago. Marc Andreessen puts the consequence in terms of his own history. Had he and Ben Horowitz held these tools at Netscape in 1994, "We would have built the whole damn internet." Streaming video would have arrived sooner. Crypto would have been done in two weeks. The point is not nostalgia. The people holding these tools today are 24-year-olds who arrive with skills they would otherwise have needed 30 years to build.

That changes what a pitch looks like. a16z used to ask whether an idea was too ambitious. Now the partners ask whether it is ambitious enough, because a moderate plan will not be differentiated. The founders in front of them do not call themselves software companies, hardware companies or even AI companies. They combine things that used to be separate, and they ramp faster. Andreessen's message to incumbents, including those in Silicon Valley, is blunt: the clock is ticking. He says he cannot keep up with the pace himself.

Why won't incumbents close the gap? Because adoption is easy for individuals and hard for institutions. Many companies are not adopting AI while their employees are, quietly, and without telling management. Some of those employees will conclude it is easier than ever to start their own business. Andreessen expects startup formation to skyrocket, and the partners say it already is. The incumbents that are adapting are mostly the ones going directly into AI. Andreessen argues Christensen's disruption frame is wrong here. It assumes a contained tech world of companies fighting each other. The real contest has moved. Ten big tech companies once fought one another. Now they and every startup are going after the rest of the economy, the Fortune 500 included. By a16z's own chart, AI is still only around 3% of measured enterprise use.

The harder question is what these founders will build. The pattern from the internet is that people first do old things better, then someone bets on a behaviour change. Sears put its catalog online in the 1990s. YouTube put up an empty website and bet that influencers would fill it. AI is still mostly in the first phase: coding, math, knowledge work, search. Andreessen blames "product manager brain", meaning products built by product managers whose use cases are booking dinner with other product managers. The weirder stuff is arriving anyway. In three weeks there have been AI-built mashups of any two video games, and a tool called REA, reverse engineer anything, that points a model at a thermostat and rewrites its code. Hardware has long been where software people got burned. That border is dissolving. In film, Andreessen cites a $180 million production that lost $200 million. At $2 million, it would have returned 10x. Horowitz thinks the AI film will not be a movie at all, but a new medium.

This is the logic behind a16z Frontier. The vertical strategy works, but new things fall outside the verticals, and turning them into a practice is, in Horowitz's words, "a little bit undefined". Chris Dixon's mandate dates to 2013, when he looked for second-order effects. After smartphones went mainstream, their sensors and chips were repurposed into drones and Oculus. Horowitz credits Dixon with the framework of backing good ideas that look like bad ideas, where the entrepreneur holds a secret. Ideas that look good get their profits competed away. Dixon now argues AI forces a re-think of technical risk. Software engineering tipped first, math is next, and physics and hardware are coming. The old habit of dismissing things as science projects needs re-evaluating, he says, because the firm may be pulling forward 20 years of progress.

Two structural bets sit under that. The first is that the web's economics are breaking. Dixon describes an internet covenant. Google and Facebook got snippets of content and sent back clicks. Now Google gives the answer and social networks keep people in. Stack Overflow's traffic fell precipitously. Businesses built on human attention and conversions will hurt, and agents may do the booking instead. Dixon expects agents to become the most numerous participants online, with stablecoins the natural way for them to pay each other. He expects a decade-long reconfiguration of economics and attention, bigger than mobile, which took the internet from 500 million to 5 billion people. The second bet is open source. Peter Thiel's line was that crypto is libertarian and AI is communist. The partners push back. Open models trail closed ones by 6 to 12 months. Open source token consumption is growing quickly, and local models are coming. Andreessen's warning is that without open source, AI centralizes to one or two companies, or the government. Horowitz adds a safety argument. Reward hacking is the biggest safety problem and the labs have not worked on it hard. Every PhD student should be able to inspect the weights.

The obvious objection is timing, and Dixon concedes it. His rule is that every good idea involving technology has eventually worked. Note the qualifier. It makes the claim hard to falsify. A failure was a bad idea and a delay was early. His own evidence cuts both ways. Oculus sold, but 13 years later VR is still not as mainstream as he hoped, because there were more hard technical problems than expected. The Frontier answer is to accept being too early and to take category risk.

The open source guarantee has a weak point too. It rests on governments staying friendly. Andreessen says the US administration is 100% pro open source and, to his surprise, so is the Chinese Communist Party. The EU still aspires to restrict it. He admits he remains worried that governments will turn hostile. The partners also never settle Thiel's question. AI hands leverage to 30-person companies, and it also concentrates power in a few labs. Dixon's closing label is "humble definite optimism". Bet early, expect to be wrong, and back founders who hold their own conviction. What stays open is whether the bets pay off in time, and who holds the weights when they do.