真正的 AI 繁荣还没开始：Andreessen 谈为何你不该成为可替代的人

Andreessen calls this a very historic moment and says 2025 was probably the most interesting year of his career, with 2026 likely to top it. He sees three forces colliding at once: collapsing trust in legacy institutions, a much freer global conversation, and large geopolitical shifts. Add AI and he compares the moment to the fall of the Berlin Wall or the end of World War II.

On the technology, he says it is now clear that AI really works. Three years after the ChatGPT moment, the open question of whether it could reason about medicine, science and law has been answered yes, and AI is even producing new math theorems. Over the holiday break, he says, AI coding hit critical mass and some of the world's best programmers conceded it now codes better than they can.

What he thinks people miss is the setting AI enters. Statistically, productivity growth in the US has run at roughly half the pace of 1940 to 1970 and a third of the pace of 1870 to 1940, so the sense of constant technological change is largely an illusion. At the same time fertility is below replacement in many countries, including China, which means populations will shrink over the next century. We need AI to lift productivity and to do work that will lack people to do it.

Andreessen and his wife homeschool their 10-year-old, and he says he thinks about AI's effect on individuals in two ways. It will make people who are good at something very good, and the truly great will become spectacularly great. His friends who are excellent coders, he says, tell him they are not twice as good as before but around ten times as good, so the goal is to get a child deep in a field in a way that fully uses AI.

Lenny suggests the key trait is agency, and Andreessen agrees while admitting he first struggled with the word. He reads it as initiative, the willingness to be a primary participant, which our rule-following culture and school system tend to discourage. He still tells his son that to lead you must first learn to obey, but he sees a huge premium on people who take responsibility and create something new.

He frames AI through alchemy. Newton spent decades chasing the philosopher's stone to turn lead into gold, and AI, he says, turns sand into thought, the most common thing into the rarest. He rejects the meme that Silicon Valley parents keep kids away from computers, arguing that the closer you are to the technology, the more you want your children to know how to use it.

Andreessen distinguishes between educating a whole nation, which needs large systems, and maximizing a single child. For the second case, he says, the ideal method has been known for centuries: one-on-one tutoring, as when Aristotle tutored Alexander the Great. He cites the Bloom two sigma effect, in which tutoring lifts a student by two standard deviations, from the 50th percentile to the 99th.

Until now only the richest families could afford that. With a large language model, a curious child can ask endless questions, get instant feedback, ask for simpler explanations and be quizzed. He suggests parents keep the traditional school but augment it with AI tutoring, mentions Khan Academy's push and the private school system Alpha that combines teachers with AI tutoring, and says now is a good time to look at the options.

Andreessen calls the job-substitution story reductive. After 50 years of slow change and little job turnover, even tripling productivity growth would only bring back the rate of turnover of 1870 to 1930, which he notes people then saw as a world awash in opportunity. Some task and job substitution will happen, he says, but the effects of economic growth and innovation will swamp it, and hiring booms should follow.

He adds that populations are shrinking and immigration may fall, so human workers in many countries will be at a premium rather than a discount. Without AI, depopulation would simply shrink the economy and remove new jobs, fields and sources of demand, a scenario he calls an economy euthanizing itself. In his view the timing has worked out miraculously well.

Mass job loss would require productivity growth of 10, 20, 30 or 50 percent a year, orders of magnitude beyond anything on record. Even in that utopian case, he says, output would surge and prices would collapse, which is equivalent to a giant raise for everyone and makes a social safety net much cheaper to fund. He stresses this is basic economics, not a bold prediction, and expects reality to be more incremental than either utopians or dystopians suggest.

Lenny lists Andreessen's correct calls, from the browser to software eating the world and five billion smartphone users, and Andreessen jokes that he buried the wrong ones behind the shed. He then says he has come around to Peter Thiel's side of their old debate. Thiel's real point, he now thinks, was that we have made lots of progress in bits but very little in atoms, and the built world looks much like it did 50 years ago.

That is also why he does not expect AI to change everything overnight. Much of the world is wrapped in red tape, unions, cartels and monopolies that resist change. In healthcare, he says, ChatGPT is almost certainly a better doctor than many real ones, but it cannot get a license, prescribe or perform procedures. He hopes AI might force a rethink of those assumptions and speed up building the future.

Asked about the three core product roles, Andreessen describes a three-way Mexican standoff: every coder thinks AI lets them do product and design, every product manager thinks they can code and design, and every designer thinks they can do the other two. He says they are all kind of correct, and that all three will eventually realize AI can manage better too, so they will aim their guns up the org chart.

The likely result is that stovepipe roles blur and talented people become superpowered across all three. A coder must learn to orchestrate a dozen coding bots, while PMs and designers pick up coding and AI design tools. The people who can build and design new products from scratch, he says, will be the most valuable of all.

He borrows the economists' view that the atomic unit of work is the task, and a job is a bundle of tasks. When email arrived, executives took over typing their own messages and secretaries moved on to planning travel and events, yet both jobs survived. Jobs persist longer than their tasks, so the advice is to swap tasks out quickly, adopt new tools and pick up more scope.

Andreessen traces the history of computing as repeated redefinition of the programmer's task. Originally a calculator was a person, and machine code gave way to punch cards, assembly, C and then scripting languages such as JavaScript, Python and Perl. Scripting was once mocked as not real programming, and now most code is written in it. AI coding is the next layer, and the best programmers now orchestrate around ten coding bots in parallel and argue with them.

For that reason he tells his 10-year-old, who vibe-codes games on Replit for hours, to still learn to write and understand code. If the code is wrong or too slow, you need the depth to see why. He says that if your goal is mediocre coding, let AI do it, but if you want to be one of the best you should understand the stack down to the chip and network, and also how the AI itself works.

AI also makes that deeper learning easier, since you can ask it to teach you memory management in ten minutes. Andreessen recalls that when scripting arrived, skeptics said it was slow and would not scale, and they were right at first and wrong in the end. People who understood the lower layers were the ones who made scripting languages work, and he expects a more dramatic version of that story now.

On design, Andreessen says AI will happily generate a thousand icons, but the capital-D questions remain: what is the thing for, will it make people happy, how does it fit into their lives. Designers will spend more time on those higher-level questions. He adds that an aspiring designer who fully harnesses AI might become the best the world has seen, because Jony Ive did everything he did without it.

He ties this to Scott Adams's career advice: being a decent cartoonist who also understood business made Adams spectacular at making Dilbert, and the value of two skills is more than double, of three more than triple. Hollywood has the same standoff among directors, writers and actors, and its superstars are those who can write and direct. Lenny proposes an E-shaped or sideways F framework, with breadth across several skills on top of depth in at least one.

Andreessen adds Larry Summers's rule for career planning: don't be fungible, meaning don't be a replaceable cog. Being only a designer, only a PM or only a coder makes you swappable, but an unusual combination of skills makes you one of few people who can do it. AI, he says, makes becoming that person easier than anything we have seen before.

The benefit he thinks people are not getting enough from AI is that it will teach you. Most people ask what the model can do for them, but Andreessen says you should spend every spare hour asking it to train you, for example to teach a coder how to be a product manager, assign exercises and evaluate the results. It will do that as happily as it does work for you. Lenny adds two tricks: watch what the agent is doing and thinking, which teaches you architecture, and when you get stuck, ask what you could have said to avoid the error. Andreessen agrees that you need grounding to understand where an agent went sideways, much as you need theory of mind to give a colleague useful feedback. He also notes you can have one AI critique another, which Lenny says people call LLM councils.

Andreessen sees three layers in how AI-forward founders operate. The first is how AI redefines the product: sometimes a new technology is merely added to existing products, and sometimes it replaces a category, as the internet did for on-premise software. He asks whether AI is just a feature in Photoshop or whether people will stop editing images once tools like Nano Banana can generate new ones.

The second layer is the job: if you have budget for 100 coders, you ask how to make them superpowered AI coders and whether you need 10 or still 100 doing ten times more. The third layer, which he says has not fully dropped, is the idea of the company. A founder might oversee an army of AI bots, pursuing the one-person billion-dollar outcome, and some are even chasing autonomous AI businesses that pay dividends to a human.

Lenny is skeptical that a true one-person billion-dollar company is feasible given the endless support tickets and edge cases. Andreessen points to Bitcoin, which was created by Satoshi, and to tiny teams behind Instagram and WhatsApp, while admitting he has no firm answers. He says many of the smartest people he knows are thinking hard about it.

Andreessen says big technological transformations take a long time to play out, and early on people state confident predictions about who will win that are mostly wrong. He notes the media rewards definitive answers over open questions, and that coverage of the internet from 1993 to 1997 and beyond shows how badly those confident statements aged. Today there are many unknowns, from how AI intersects with politics and unions to what China will do.

On moats, he argues both sides. Model building costs billions, needs scarce engineers paid like NBA stars, and might yield an oligopoly. But within a year or so of ChatGPT, several American companies, several Chinese ones and open source models offered comparable products, and apps built on models might hold the value instead. He points to Claude Code building Cowork in about a week and a half as proof of both capability and thin barriers.

He describes the industry as a complex adaptive system in which technology, regulation, entrepreneurs' choices, economics and investor capital all interact, and he says he puts a big discount on his own forecasts. For a VC, that justifies making bets across strategies. Many insiders also tell him there really aren't many secrets among the big labs, though a dozen nascent ideas inside any one lab could become hard-to-replicate breakthroughs.

Andreessen recalls Peter Thiel's two-by-two of optimism and pessimism, determinate and indeterminate. Thiel regarded Elon Musk as a determinate optimist with concrete plans, and VCs as indeterminate ones. Andreessen places himself firmly in the indeterminate camp: Silicon Valley has thousands of determinate founders, and the best way to deal with an unknowable future is to run as many experiments as possible. Founders, he says, need a specific plan, and history remembers Henry Ford rather than his seed investor.

On AGI, he finds both common definitions lacking. The cosmic version, a singularity where human judgment becomes irrelevant, he does not think we live in, and the practical version of doing a basket of valuable economic tasks as well as a human understates things because human skill is capped by biology. Human IQ tops out around 160, he says, and current models test around 130 to 140, so he expects models at 160, 180, 200 and beyond.

He thinks the world would be better off with more Einsteins, and that machines exceeding human ability in coding, medicine and law is great news. Human-equivalent AI will be a footnote, he says, and what matters is the process of exploring beyond human capability. He adds that he often feels the frustration of his own limits, and finds the prospect of machines without them much more exciting than people give credit for.

Andreessen describes his media diet as a barbell: he reads X for what is happening right now and old books that have stood the test of time, and he is skeptical of everything in the middle. Rereading last week's newspaper, he says, shows that none of the predictions played out as claimed. He also thinks content from practitioners, through newsletters, podcasts and Substack, is still dramatically underrated, since experts like to explain what they do.

He ties this to Silicon Valley's open culture, calling AI the ninth major technology platform in its history. He notes the region is still called Silicon Valley although it no longer makes silicon, and that its flexibility, rather than any plan, let it move from chips to the internet, smartphones, cloud and now AI. His film pick is Eddington, which he calls the best movie of the decade for how it handles 2020 and life lived through the internet.

On products, he says his 10-year-old is obsessed with Replit and vibe-codes Star Trek simulators using the LCARS design language. He is also excited by AI voice, such as Grok's Bad Rudy avatar and Sesame, by voice input apps like Whisper Flow, and by a coming wave of wearables and pendants. He closes by pointing listeners to a16z's YouTube channel.