Everyone's Token-Maxxing: An Arms Race Nobody Dares to Stop | 5Y Capital's Xing Meng on AI

We spent a week in Silicon Valley and found that even the ones making the waves are getting drowned by them.

LatePost Column | Xing Meng, Partner at 5Y Capital

On the morning of March 24, 2026, sitting in the audience at YC's W26 batch Demo Day, I decided to stop taking notes by the time the fifth company took the stage.

Not because it didn't matter. Because I realized that whatever I was writing down would probably be obsolete next month.

Of the hundred-plus companies in this batch, their focus was remarkably narrow: roughly 80% were vertical agents — tools that help lawyers organize documents, help customer service triage tickets, help HR screen resumes.

If I'd seen these projects last October, I probably would've thought, "Not bad." But the problem is, in these five months, the world changed.

Claude Code evolved from a developer-centric tool into something almost anyone could use directly. When Opus 4.6 dropped, the barrier to vibe coding fell through the floor.

Those vertical agents, before they'd built any real business moat, could now be replicated by an ordinary engineer — or by me personally — over a single weekend. They'd already lost their investment value.

A YC batch runs three months. This cohort entered in December, and with pre-screening, these "promising companies" were essentially selected five months ago. And five months, at today's AI iteration speed, is enough for several paradigm shifts.

When I first started a company in 2012 and got YC's Fly Out invitation, the accelerator space had almost no real competition. YC's picks often signaled "the next big thing." But the competitive landscape has shifted. These past few years, YC seems to have inverted into a lagging indicator.

YC's batch system — apply, screen, enter, refine, pitch — ran successfully for over a decade in the mobile internet era. But that rhythm was designed for a slower world.

In the year and a half since returning to venture capital, I've come to Silicon Valley roughly once per quarter; last time was October. On previous trips, things always felt fast, but that "fast" was measured in months.

This time, it's measured in weeks.

Over dinner one evening, a friend working in post-training mentioned casually:

"I've realized that Silicon Valley itself can't keep up with itself anymore."

Token-Maxxing for All: An Arms Race Nobody Dares Stop

If someone had told me six months ago that tens of thousands of Meta engineers were all using a competitor's product to write code, I'd have thought they were joking.

But it's true. All of Meta is using Claude Code. Not a startup. Not some experimental team. A trillion-dollar company.

Code security? Abandoned. Token budgets? Blown up. Leaderboards? Gamified. All of Silicon Valley is throwing money at AI with abandon. But then what?

Let's start with code security. Six months ago, this would've been unthinkable — code is a company's core asset, how could you let some other company's API touch it? Meta initially thought the same; they built something internally called myclaw to address this. A Meta friend told me they produced a coding product, but "it wasn't good, nobody used it." After nobody used it, the company had to loosen up: as long as no customer data was involved, use Claude Code if you want.

Then departments started holding internal meetings on "how to become an AI-native organization," running trainings, implementing performance reviews. Code security, usage security — red lines that used to be sacred — all got deprioritized. Boost productivity first, figure out the rest later.

For security reasons, Google prohibits most employees from using competitor tools like Claude Code or Codex, but DeepMind is an exception. Several teams responsible for Gemini models and internal applications are using Claude Code.

Google hasn't been idle either: they launched an internal coding tool called Antigravity, and in February claimed that roughly 50% of the company's new code was already AI-written.

But even so, DeepMind people still use Claude Code. One key reason DeepMind can do this is that Anthropic set up a private deployment for them — after all, Anthropic's inference and training already run heavily on Google Cloud TPUs, so there's mutual trust. But Meta and other tech giants don't have that relationship. They're genuinely throwing code security out the window. Everyone is betting on the same thing: speed first.

Code security was the first flag to fall. The second is token budgets.

Among the AI-native startups I talked to in Palo Alto, an engineer's annual token budget runs around $200,000-plus. The number itself isn't remarkable; what's remarkable is that it means the AI costs for a top engineer now approach the engineer's own salary. It looks like companies are using AI to cut headcount and save money, but total costs may not actually drop — they've just swapped people costs for token costs.

Meta is again the most extreme case. They created an internal token consumption leaderboard: high usage gets you on the board, low usage might get you laid off. Meta employees are even competing for an unofficial title called "token legend."

But at the same time, Meta has done two rounds of layoffs this year, totaling over ten thousand people. Everyone using Claude Code to pump token volume, while massively cutting headcount.

These aren't contradictory. They're two sides of the same coin.

I visited a Series C company where the head of engineering opened Slack to show me: nothing but agents running, a dozen Cursor agents working in parallel in the background, plus a Claude Code window for orchestration. The most popular anxiety among programmers right now: going to sleep without knowing what those dozen agents are going to do.

But has productivity actually increased proportionally? Since late last year, many CTOs at top inference engine and database companies excitedly told me about "100x engineers," "10x efficiency gains" — what used to take 60 people a year, now two people with Claude Code can do in a week.

I got excited with them at first. But when I calmed down, I'd ask one question: Okay, efficiency up 100x — has revenue grown 100x? Has the product line expanded 100x? You can't have "100x" improvement that just ends up optimizing away headcount, right?

I never got a straight answer. The reality is, 100x efficiency gains translate to maybe 50% or 1x revenue growth at the company level.

Where's the gap? Nobody can say yet.

"With this much token usage, companies should be mutating into something fundamentally different. But into what, I don't know either."

One founder with an enterprise sales background told me his team of 16 people, two in sales, went from zero to $30 million ARR in 12 months — all built on AI coding. You do see cases like this occasionally. But most of the time, what I see is startups building more things, and those things not having product-market fit.

Vibe coding to try 100 approaches and see what sticks, rather than just 10 — that's very trendy in Silicon Valley right now. But who will actually capture the next wave? Still hard to say.

The most striking counterexample came from inside Anthropic itself. I asked an Anthropic friend: what's the most painful internal use case for agents? He said oncall.

A typical oncall scenario: Claude's API suddenly slows down, an inference node goes down, users report abnormal outputs for a certain prompt class — the oncall engineer needs to quickly identify root cause, determine whether it's a code bug, compute allocation issue, or model-level anomaly, then decide how to fix it.

Anthropic is literally the best company in the world at building coding agents. This use case couldn't be closer to their core competency. And their internal oncall agent still doesn't work well.

That's the real state of affairs in April 2026: the steam engine has been invented, but sometimes it still runs slower than a horse. The key thing is, everyone knows the steam engine will eventually outrun the horse, so they're all throwing money at it wildly. Code security abandoned, token budgets blown, leaderboards gamified. As for when the steam engine actually outruns the horse? Nobody knows. But nobody dares stop and wait for that day.

Because the cost of stopping may be greater than the cost of burning the wrong tokens.

And token consumption is probably not growing linearly. This reminds me of my autonomous driving days: in 2021 in Shanghai, we first achieved five consecutive hours of hands-free autonomous driving. It felt like a major breakthrough. Before that, the test fleet might grow from 10 to 15 to 20 vehicles slowly; but past that inflection point, it quickly reached 100, then 1,000. Today's coding agents are at a similar stage.

In 2021, DiDi Autonomous Driving achieved five consecutive hours of hands-free driving in Shanghai — a milestone for autonomous driving in China. The photo shows Xing Meng, then COO of DiDi Autonomous Driving, in conversation with Sebastian Thrun, Google's "godfather of self-driving cars," in 2021.

METR is a California-based research institute specializing in evaluating AI coding capabilities. Last year they proposed a metric: measuring how long a task AI agents could complete with 50% success rate (calculated by human expert completion time). When first published in March 2025, Claude 3.7 Sonnet's number was 50 minutes; by the end of 2025, Claude Opus 4.6 had reached 14.5 hours. Over the past two years, the doubling time for this metric has compressed from seven months to four. Once agent reliability takes another step up, token consumption won't be a question of adding 50% per year — it'll jump an order of magnitude overnight.

There's a prediction my friends broadly agree on: by the end of this year, many companies (including major tech firms) will effectively need only 20% of their current headcount.

After the xAI Exodus, the Rocket Builders Started Building Models

At a steakhouse in Mountain View, past 9 p.m., a friend who had worked with Elon Musk for a long time sat down across from me. We talked for over three hours, and looking back, I don't think he said a single good word about Musk the entire time.

One detail: I asked him, you spent three years at xAI, what was the daily rhythm like? He said he basically lived at the company all three years, so he never really furnished his place — didn't even buy a bed. He slept in one of those sleeping pods at the office, like a youth hostel. I said, you're sitting on massive equity, you've already left, at least buy a bed. He smiled.

xAI's work intensity is notorious in Silicon Valley, but by now roughly 90% of the early team has left. They have an alumni group chat; people are joining every day.

The trigger was Tony Wu's departure, then a chain reaction. In the words of one insider, "What might take other companies six months to brew — a senior team exodus — xAI did in a month." Some people sensed Musk's dissatisfaction as early as October last year, but nobody expected such a thorough purge so quickly.

Now Musk is pulling people from SpaceX and Tesla to take over xAI. "The rocket builders are starting to build models."

Musk's frustration comes from pouring in endless capital and compute, yet Grok never broke into the top tier. But why? This is the question I ask everyone who left xAI. The answer is simpler than I expected. One friend put it directly: the team's combat effectiveness was extremely strong, their work ethic incredibly intense, but manufacturing-style management may not suit a large model company.

I spent eight years in autonomous driving, so I have some personal perspective on this. What Musk did with SpaceX and Tesla was fundamentally systems engineering: long chains involving software, hardware, supply chain, each with room for innovation, but ultimately an end-to-end engineering problem.

What he's good at is identifying the key leverage points in these long chains, then compressing timelines to the limit to crack them. Staged rocket engine combustion, reusable landing — these are products of that thinking.

But at xAI, what he's doing doesn't look like systems engineering. He's done three things: first, throw money at the world's largest GPU cluster (today people joke that xAI was supposed to be a neo lab, now it's more like neo cloud, basically providing compute for Cursor); second, set pulse-style deadlines for the team; third, personally dictate some product features. He's grabbing at a few points, not doing integrated planning.

Anyone who's done autonomous driving knows that once you get to later stages, "who leads whom" between the software team, infra team, and hardware team becomes the core tension. All three directions need CTO-level decision-makers, but no single person understands all three domains. The good approach is: even if the founder doesn't fully understand each piece, they know how to balance resources, set phase-by-phase priorities — software takes precedence this period, next phase push to infra. That's called having integrated planning.

xAI's problem is it doesn't have this integrated planning, only sprints. If the pressure weren't so intense, smart people could self-repair — given time, each direction would find its own collaborative rhythm. But Musk's hyper-intense management, combined with insufficient integrated planning, causes things to collapse under pressure. Every direction's leader is protecting their own priorities; nobody is doing global coordination.

A neglected reason why SpaceX and Tesla succeeded so spectacularly is that in both industries, Musk basically never faced competitors of equal magnitude. He was competing against himself. But AI is different. AI is a level of brutal competition where even OpenAI can get ambushed by Anthropic.

A friend at a top lab said last year there were two things he didn't expect: first, how brutal the competition would be; second, how few opportunities for application-level innovation there would be in the AI era — everything was getting eaten by the models.

Anthropic's rise was the most dramatic reversal in the AI industry over the past year. It also completely shifted the battlefield: a year ago everyone was competing for C-end user volume and video generation; now the (phase-dependent) decisive battlefield is B2B and coding.

Of course, the xAI story is simultaneously a story of "what happens when money comes too fast and too much."

I don't think the friends leaving xAI today regret their decision to join. xAI was arguably Silicon Valley's fastest wealth-creation myth. From its first funding round of several billion dollars to its merger with SpaceX into a $250 billion behemoth took just one year. And xAI's 11 cofounders — nearly every one became a billionaire. Core engineers also got tens of millions to $100 million. There's really just too much money in Silicon Valley. Today, if they start new companies, they'll have full confidence to pursue directions they're genuinely interested in, not just quick-money plays.

Anxious Engineers, Even More Anxious Researchers

Talking with engineers today, there's a strange tacit understanding: everyone admits they barely write code anymore, but everyone pretends this is no big deal, because they'll be the ones armed with AI who eliminate the engineers who haven't AI-ified themselves.

Today, 80% of software engineers' core skills have already been replaced by models. The only reason they're still around is that models occasionally do something stupid and need human oversight. But the "oversight" itself may soon not be needed either.

Think more radically: today's so-called "AI-native organizations" sound sexy — have every department map out workflows, digitize the parts where AI can intervene, write them as skills. But essentially you're distilling yourself manually: you turn your capabilities into machine skills, the company gets your skills, and it's effectively already AI-ified. Whether to lay people off after that becomes an ethical question. This is what Meta is doing today.

Although everyone's competing on token-maxxing right now, you can still feel a pervasive underlying anxiety throughout Silicon Valley.

What surprised me more is that this anxiety is now spreading to researchers.

Researchers are the most pyramid-apex talent. It doesn't broadly mean "research personnel" — it means the people at large model companies (OpenAI, Anthropic, DeepMind, etc.) responsible for model training and algorithmic innovation. The difference from engineers: engineers "build things" — write code, deploy, optimize performance; researchers are further upstream, "figuring out what to build" — proposing new training methods, designing model architectures, running experiments to validate hypotheses.

And now, even researchers' work itself is being automated. This is what DeepMind colleagues are doing — using models to train models, also known as the AI self-improvement that's been on fire in Silicon Valley this year. This year it's engineers being eliminated; by year-end, researchers will start being replaced too.

This isn't a new concept. Andrej Karpathy's auto research opened the door. Today various AI scientist tools and harness frameworks are moving in this direction. But most current closed loops only go as far as "publish a paper" — AI helps you run experiments, write papers, but humans still make the judgments.

What OpenAI, Anthropic, Google want to do is more radical: they want the closed loop to go directly to model upgrades themselves, not just incremental improvements, but letting AI find the next paradigm-level breakthrough on its own. If this can be achieved, that's truly replacing researchers. Google DeepMind started working on this internally over a year ago — letting models decide what experiments to run next, evaluate which path is more promising after running them, then follow that path. This is models training their own next generation.

And researchers have more motive to be cut, for a brutal reason — because they're expensive. There are maybe only a few thousand researchers globally, with annual compensation ranging from millions to tens of millions, even hundreds of millions of dollars.

"The future scenario might be: 10 people doing the work of 100, taking 20 people's pay, and 90 people unemployed."

And the real layoffs are larger than the surface numbers suggest. The first cuts many companies make aren't on their own financial statements — they're to outsourcing providers. This means India and the Philippines, countries that once took on customer service, data labeling, and financial back-office work from Europe and America, may be hit first. The "services ladder" that some developing countries relied on to upgrade their economies may be getting pulled away by AI.

All of Silicon Valley is watching Meta. If its experiment succeeds — revenue doesn't drop, efficiency actually improves — other major firms will quickly follow, and layoffs will shift from isolated cases to industry norm. And layoffs have a brutal self-accelerating mechanism: at first nobody dares to cut, fearing morale damage; once it becomes normal, the cuts come faster and hurt less.

But as old positions are eliminated, new ones are also emerging.

Many startups are now hiring for a new role called "AI builder" — a hybrid that combines product manager, frontend engineer, and backend engineer into one person. There's also a composite role merging data scientist and machine learning engineer, and another that fuses writing, paid acquisition, and operations into a unified content operator.

Silicon Valley companies are ravenous for these new roles, but the core problem is: nobody knows how to hire for them. You can't screen by résumé because the role didn't exist before; the candidate's abilities might be buried entirely in personal projects. You can't test them with live coding either, because the core competency is a combination of taste + AI fluency. So startups are already building solutions: generating simulated environments based on employer needs, where candidates complete tasks using AI tools on the spot. Kind of like the old coding test, but measuring something entirely new.

When AI can do everything, human value is shifting from "what you can do" to judging "what's worth doing and what isn't."

One Round, Two Valuations: NVIDIA Securing Chips at Every Table

I've spent all this time talking about who's being replaced — engineers, researchers, finance professionals. But one role isn't being replaced at all. Instead, it's becoming more like the shadow boss behind this entire reshuffle.

This world that looks like distributed innovation is, at its foundation, radically centralized.

That center is NVIDIA.

I had thought the scarcity of GPUs had eased over the past year. And it did, briefly. Around mid-2025, some NVIDIA-backed neo clouds — the "new cloud providers" that emerged during the AI wave offering GPU compute — struggled to raise funding. Some saw flat growth, and a few even sold themselves around that time. But what I found this trip is that scarcity is back, and it's more extreme than before.

One concrete signal: if you can deliver an API service with 99th-percentile stability today, you can charge two to three times the official API price.

After Anthropic's demand surge, API outages are becoming more frequent — a real problem for many Agent products built on top of Claude

The router business used to work like this: "I'm cheaper than official, so I get traffic." Now the logic has completely inverted: stability itself has become the scarce resource. A wave of startups is making serious money off this, and mini versions of CoreWeave / Nebius are springing up across Silicon Valley.

And this time the compute bottleneck isn't just about GPU allocation. Elad Gil recently made an assessment I agree with: upstream memory manufacturers (Hynix, Samsung, Micron) need at least two more years to expand capacity. This means no AI company can pull significantly ahead by simply stacking compute before 2028. The compute constraint is objectively reinforcing an oligopolistic structure in the large model market — it's not that anyone isn't trying hard enough, it's that physical-world manufacturing cycles are simply that slow.

The power structure behind this is clear: whoever has the GPUs is powerful, and who has the GPUs is decided by NVIDIA. CoreWeave, Lambda, and Nebius, all now public, all have NVIDIA standing behind them.

NVIDIA's positioning runs deeper than I previously understood. An investor at Reflection told me that when this neo lab first went out to raise funding, it was doing coding. Then the founder met with Jensen Huang, who told him: stop doing coding, come out and build "America's DeepSeek" for me — an American open-source model. I'll give you money and GPUs. Reflection did a 180-degree pivot.

This has also produced some unusual structures in U.S. capital markets: the same funding round, two valuation tiers. Investors with good relationships who got in early enter at the lower valuation; NVIDIA, the deep-pocketed leader, and latecomers get pushed to the higher valuation tier. This structure has started appearing in China recently too.

But however much NVIDIA wants to control allocation, it can't manufacture what doesn't exist.

Across American society, protests against data centers are escalating. Roughly 100 data center projects nationwide are now facing resistance, with 40 expected to fail outright. Maine just passed a bill banning data center construction entirely. One town approved a $6 billion data center project, then half its members were voted out overnight; the replacements had one purpose — reverse that decision.

Compute isn't scarce because the product isn't good enough or users aren't numerous enough. It's scarce because the physical world can't keep up with the digital world's appetite.

This is another dimension of "falling behind."

Silicon Valley's Valuation System Is Being Rewritten

Look at one number.

U.S. GDP is roughly $30 trillion. OpenAI and Anthropic each currently have revenue run rates around $30 billion — meaning each company alone already accounts for 0.1% of U.S. GDP. If both hit $100 billion by year-end, plus cloud services and other AI revenue, AI will represent roughly 1% of U.S. GDP. From virtually zero to 1% in just a few years.

This speed is unprecedented. But the strange thing is: the faster the growth, the less investors know how to price it — faced with such velocity, Silicon Valley's valuation framework is collapsing.

After several deep conversations with public market friends this trip, one phrase kept coming up: "re-rationalization."

For the past few years, AI investing logic ran on future cash flows: it's fine that you're losing money today, I'm betting on your ARR three years out, five years out. But now that framework is broken.

The problem lies in DCF, the most basic valuation model. Normally you project cash flows for 10 years, then add a terminal value — assuming the company eventually stabilizes and operates steadily, bundling remaining value into one lump sum. Typically terminal value accounts for 70-80% of total valuation.

But two things have changed simultaneously: first, you might only be able to project 3 years instead of 10, because what this industry looks like in 3 years (sometimes even 1 year) is fundamentally unknowable; second, terminal value becomes impossible to calculate, since its premise is that the company eventually stabilizes — but if AI can disrupt everything at any moment, the "stable operation" assumption doesn't hold.

A public market investor friend gave me this analogy: companies not on AI's main trajectory today are more like waiting for a "nuclear bomb" — you know they'll be disrupted, just not when. So the evaluation focus shouldn't be "what if they're not disrupted," but "how fast can they respond when disruption hits." This is an entirely different valuation logic.

SaaS was the first to be repriced by Wall Street. In 2023, Snowflake would have needed nearly 100 years of free cash flow to pay back its valuation; it's now been cut in half. ServiceNow, Workday — same trend. This is just the beginning.

Conversely, the companies that might actually still suit DCF valuation may only be the leading large model companies, because relatively speaking, their futures appear to be stably growing in a positive direction. They won't "get nuked" — they're seeing how far their boundaries can expand.

The old startup recruiting pitch was: "Lower salary, but here's equity, it'll be worth a lot someday." But that pitch assumes the company still exists and is valuable in 15 to 20 years. If that premise no longer holds, the rational employee response becomes — "Forget the equity, just give me cash."

Which in turn changes company cost structures and financing logic.

The VC side is hurting too. Over the past 3 to 6 months, nearly every fund in Silicon Valley invested in at least one neo lab — researchers from prominent AI labs, raising hundreds of millions on their ideas. Now, in retrospect, everyone feels a bit impulsive, a bit like they overpaid. But why did they still invest? Because if the company actually works, growth will be so fast that the original valuation will look cheap.

One investor friend put it bluntly: either way it's zero to 100 or zero to zero. Rather than investing in an expensive Series A to earn "hard money," better to buy a ticket for a neo lab with unlimited upside.

People used to think $1 of ARR was $1 of ARR, whether you were building models, applications, or infrastructure. Now that equation is broken.

Vertical agent companies trade at the lowest multiples (around 5x), general agent companies higher (around 10x), model companies highest (20-30x ARR — Anthropic at $30B ARR, $800B valuation, 26.7x). A year ago I thought you could just apply a uniform multiple to ARR. That approach is completely wrong now.

Lime Trees and the AI Assassination List

Silicon Valley is experiencing a deep crisis of security.

Throughout this trip, I kept hearing friends seriously discuss the same things: buying Bitcoin, building bunkers, installing bulletproof glass at home — and they weren't joking.

Lime trees have become genuinely popular in Silicon Valley lately, because their branches grow four-inch thorns that make anyone trying to climb over pay dearly.

The Wall Street Journal even reported on a $15 million "fortress mansion": concrete planters ringed with lime trees, behind them a moat, behind that a laser intrusion detection system, the front door a three-inch-thick solid steel plate with thirteen deadbolts, inside a safe room behind a 2,000-pound door — even the landscape architecture is defensive fortification.

Residential security firms serving CEOs are seeing their highest growth since 2003. The trend accelerated sharply after the UNH CEO was shot and killed on a Manhattan street.

Then the gunfire reached AI leaders' doorsteps.

At 4 a.m. on April 11, a 20-year-old boy in a Champion hoodie flew from Texas to California, carrying a fuel can. He stood before Sam Altman's $27 million mansion, lit a Molotov cocktail, and threw it in.

Ninety minutes later, he appeared at OpenAI headquarters, grabbed a chair, smashed a glass door, and shouted at security: "I'm going to burn this place down, kill everyone inside."

The FBI found a document on him. The title: "Your Final Warning." Inside were names and home addresses of multiple AI company CEOs and investors.

Two days later, early Sunday morning, Altman's home was attacked again: a Honda sedan pulled up briefly, the passenger window rolled down, a hand extended with a gun, fired a single shot at the house, and sped away.

This was no isolated incident. In late March, a massive anti-AI protest had already swept through downtown San Francisco, crowds carrying signs reading "Stop the AI Race" and "Don't Build Skynet," giving speeches outside the offices of Anthropic, OpenAI, and xAI. Senator Bernie Sanders warned in Congress: "Human beings may really lose control of this planet."

Talking with friends at xAI, I heard that Elon Musk is also deeply afraid of being assassinated — an open secret in the circle.

The fear behind all this is actually quite simple: if AI takes over most production, and humans are no longer necessary participants in economic activity, then all past social contracts about "how much you contributed, how much you deserve" simply collapse. What remains is a radically simplified power structure: whoever controls the GPUs and electricity controls everything. Class isn't being stretched apart — it's being flattened: a tiny minority on one side, everyone else on the other.

"In the US presidential election two years from now, the hottest campaign issue will definitely be AI's relationship with society. We might even see a Luddite movement for the AI age."

Inflation in America remains severe. Having lived in California for many years, I'd never seen gas prices hit the sevens. Right around late February, Citrini published Global Intelligence Crisis — an "AI doomsday report" that scenario-planned an economic crisis in 2028 caused by AI being "too successful"...

Epilogue

On the flight back to Beijing, I flipped through my notes from the past two weeks and realized I'd been writing the same word from start to finish: "can't keep up."

YC can't keep up, Meta's code security rules can't keep up, xAI's management can't keep up, researchers can't keep up, compute can't keep up, valuation frameworks can't keep up, society's psychological capacity can't keep up... to the point that Silicon Valley itself can't keep up with itself.

But here's what I want to end with. A friend at Anthropic mentioned that Dario Amodei said something internally: with AI's help, cancer has in some sense already been solved — not that it's disappeared, but that it could become a chronic disease that doesn't kill you, though treatment remains too expensive and widespread adoption will take time.

I'm not sure if Dario's claim that "cancer is already solved" is overly optimistic, but during this trip to Silicon Valley, the startup direction we saw most frequently was AI4S, AI for Biotech. Many people coming out of large model companies don't understand medicine, but they want to use AI technology to transform this industry.

I saw so much "can't keep up" during these two weeks, and that is genuinely anxiety-inducing. But if AI really does turn cancer into a chronic disease within a few years, or fast-forwards materials science by two decades, then this "can't keep up" may be the greatest acceleration in human history.

My baby is two years old this year; we may have a second child next year. I have absolutely no imaginative capacity to construct what kind of world their generation will face.

But I hope that in the world they grow up in, there will be more people healed because of AI, and fewer Molotov cocktails and gunshots crashing against AI practitioners' front doors.

In Paul Graham's 2008 essay Cities and Ambition, there's this passage: "Though people in Silicon Valley are very respectful of intelligence, the message Silicon Valley sends is: you should be more powerful. That's not quite the same as New York's message. New York is impressed by wealth, especially if it's in the billions. But Silicon Valley doesn't care about that. Except for a few real estate agents, no one in Silicon Valley cares about that. What matters in Silicon Valley is how much effect you have on the world. The reason people there respect Larry and Sergey is not their wealth but the fact that they control Google, which affects practically everyone." Today AI has pushed this atmosphere to a new peak.

LatePost columnist Xing Meng: Partner at 5Y Capital, former COO of DiDi Autonomous Driving. This is the first installment of his AI investment observations, which he will continue to update on LatePost.

5Y Capital seeks out, supports, and inspires lonely entrepreneurs, providing them with support from the spiritual to all operational aspects. We believe that if the "crazy" you in others' eyes begins to be believed in, the world will become a different place.

BEIJING · SHANGHAI · SHENZHEN · HONG KONG

WWW.5YCAP.COM