July 2026
Watch what the frontier labs brag about now.
Two years ago, every model launch led with a benchmark table. MMLU, GPQA, SWE-bench — bar charts with your model in a slightly darker color, edging out the competition by two points. The launches of 2026 lead with something else. Price per million tokens. Latency. Efficiency curves. And since inference cost is mostly electricity plus depreciation, tokens per dollar is really tokens per joule.
Nobody announced the change. But the scoreboard is different. The race is no longer how smart. It’s how smart per watt.
That shift is worth staring at, because industries only change their headline metric when they’ve changed what they are.
Chips did this. The 1990s were the megahertz wars — raw clock speed printed on the box, bigger number wins. Then mobile arrived, batteries became the constraint, and the metric that mattered became performance per watt. Nobody puts clock speed on a billboard anymore.
Electricity did it first. The 1880s were pure spectacle — arc lamps, world’s fairs, Edison electrocuting an elephant to frighten people about alternating current. Then the spectacle wore off and electricity became cents per kilowatt-hour.
Here is the rule hiding in both stories: when a technology’s defining metric becomes output per unit of energy, the technology is confessing what it has become. An input. A commodity. A utility. Not a marvel anymore — an ingredient, priced like one.
Intelligence just made that confession.
Why now? Because the labs did this to themselves, and they had no choice.
Capability leads that used to last a year now last a quarter. Every trick discovered at the frontier leaks downmarket within months — distilled, open-weighted, replicated. The price of a fixed unit of intelligence has been collapsing roughly ten-fold a year, not because anyone wanted it to, but because five labs are competing like hell and none of them can afford to stop. When your product is interchangeable with your rival’s, and your rival’s marginal cost is the same electrons flowing through the same silicon, price doesn’t drift toward value. It falls toward cost. That’s not a strategic failure. That’s just what competition does to things that can be measured.
And intelligence, as the industry defines it, can be measured. That’s the trap. Anything with a benchmark can be matched. Anything that can be matched gets competed to cost. The benchmark culture the labs built to prove their superiority is the same machinery that guarantees the superiority is temporary.
So the labs are pouring hundreds of billions of dollars a year into building what is, structurally, a grid.
And the thing about grids: nobody got rich owning one. Electric utilities keep civilization alive and pay out like bonds. The grid electrified everything and enriched almost no one who built it.
So where did electricity’s fortunes actually go?
Radio. Hollywood. Refrigerators sold as status objects. Neon Times Square. Amusement parks strung with light bulbs. The money went to the people who took a cheap, abundant, boring input and turned it into something people wanted — which turned out to be mostly entertainment, comfort, and glamour. The frivolous layer, built on top of the serious one, captured nearly all the margin.
Telecom ran the identical script a century later. Carriers spent fortunes on fiber and spectrum and became dumb pipes. Netflix, YouTube, TikTok, mobile games — the trivial things riding the pipes — took the value. The pipe delivers bits at cost. The show on top charges whatever attention will bear.
It offends the engineer in me, but the pattern doesn’t care. The layer that touches human desire keeps the margin. The layer that delivers the electrons earns cost plus a thank-you.
The instinct is to resist this because it sounds unserious. Surely a trillion dollars of intelligence infrastructure pays for itself in cured diseases, written code, automated enterprises — utility in the noble sense.
Some of it will. But notice how each layer gets priced.
Usefulness saturates. A model that files your taxes correctly is worth exactly the price of the cheapest competent model that files your taxes correctly — and there is always a cheaper competent one next quarter. Correct is correct. Utility has a spec, the spec has a benchmark, and we already know what happens to things with benchmarks.
Delight doesn’t saturate, because it was never graded against a spec. It’s graded against your attention. The image of you and your friends rendered before you finish describing it. The companion that remembers what you said at 1 a.m. The character, the game, the feed. There is no eval for wanting more. You can watch the revealed preference in real time: companionship apps out-earn productivity assistants, image generation ships inside social feeds rather than office suites, and the labs that swore they were building scientists keep shipping entertainment. Nobody is confused about where the demand curve slopes. The superficial layer isn’t a distraction from the business. It is the business — the same way it was for electricity, the same way it was for bandwidth.
The one honest objection is winner-take-all: some lab crosses into recursive self-improvement, capability stays scarce, the grid owner becomes the everything-company, and none of this history applies.
Maybe. But every visible dynamic points the other way. Leads compress. Distillation leaks. And the binding constraint at the frontier is energy — the same electrons, the same physics, available to every competitor at the same price per joule. Intelligence per watt converges because physics is shared. Permanent scarcity of intelligence is a bet against everything currently observable, made mostly by people who need it to be true.
The datacenters are the grid. Necessary, colossal, and headed for utility returns. The scarce things are the ones capex can’t buy: distribution, attention, taste. The labs know it — that’s why every one of them is scrambling to become a consumer company before intelligence finishes commoditizing underneath them.
Intelligence will be sold by the joule. Joy will be sold at whatever the market will bear.
zguo0525@berkeley.edu · @Zhen4good
]]>May 2026
A language model, asked to predict the next token, computes probabilities across its entire training distribution and samples one.
It never decides. It calculates.
That distinction sounds small. It is not. It is the line between everything machines can do and everything humans are for.
Deep Blue beat Kasparov. AlphaGo beat Lee Sedol. GPT-4 passes the bar. Each milestone follows the same script: another domain that “required intuition” turns out to require only enough data and enough compute.
So the temptation is to draw a shrinking circle around what’s still ours. Consciousness. Creativity. Empathy. Pick your favorite. Then watch a model do a passable version of it next year, and shrink the circle again.
I think we are drawing the wrong shape.
The thing machines cannot do is not a capability. It is a stake — the act of binding yourself to an outcome you cannot verify. Marrying someone. Starting a company. Believing a paper is worth writing before anyone agrees. Voting. Praying. Choosing.
Pattern completion does not bind. It outputs and moves on.
When you commit to something real, you cannot possibly have the relevant information. You do not know who your partner will be in twenty years. You do not know if the company will work. You have not yet met the obstacles that will test whether you keep going.
You commit anyway. You choose to treat something as if it were certain, knowing it is not.
William James called this “the will to believe.” Not self-deception — practical necessity. When evidence is genuinely ambiguous and the choice itself helps determine the outcome, waiting for certainty is choosing never to act. The skeptic who refuses to believe “until all the facts are in” never participates in creating the facts.
A model can output a recommendation with 73% confidence. It cannot answer should I marry this person. Not because it lacks data — because the question presupposes a kind of stake the system does not have. You cannot wager what you do not own.
Here is the part that surprises people: human creativity thrives exactly where certainty ends.
Einstein published special relativity when the Michelson–Morley experiment was still contested. Barbara McClintock’s jumping genes were dismissed for thirty years. Every paradigm shift in science is somebody believing the answer before the evidence justifies the belief.
Art works the same way. The novelist does not know how the book ends. The painter does not know if the next stroke ruins it. They continue, guided by something that is neither logic nor randomness — a cultivated trust in judgment that no amount of training data can substitute for, because the training data does not exist yet.
A diffusion model produces an image by denoising a random seed. It does not struggle. It does not wake at 3 AM convinced the project is worthless and return to it anyway. The machine outputs. The human endures.
I build agents for a living. So this is not an abstract argument for me — it is the question I sit with every time I wire one up.
When you give an LLM “agency,” what you have actually given it is a longer chain of pattern completions. The model picks a next action the way it picks a next token: by sampling from a distribution. Increase the horizon, add tools, give it memory, let it reflect — it gets more capable, sometimes dramatically. But at each step, it is still computing what the data would predict. It is not staking anything on the answer.
This is fine for most of what agents are useful for: scheduling, retrieval, code generation, research. Domains where “the right answer” is recoverable from prior cases.
It breaks at the edges where humans actually live. The agent will not start a company. It will not decide, at the cost of its own continuation, that a project is worth doing anyway. It will not refuse a profitable instruction because it violates a principle it cannot prove. It can imitate all of these. It cannot wager.
That gap is not going to close by scaling. Scaling makes the imitation better. The wager is a different category.
I should end here with a tidy conclusion. That would be dishonest to the thesis.
I do not know if this argument is correct. Maybe future systems will surprise us — some mechanism we have not imagined, indistinguishable from commitment, arising not by simulation but by a substrate change we cannot currently see. Maybe the line I am drawing is a temporary artifact of 2026’s architectures rather than a fundamental boundary.
Or maybe the line is sharper than I have suggested. Maybe there are souls, or sparks, or something else that guarantees the gap stays open regardless of what we build.
I have some of the answers. Not all of them. And yet here I am — staking my credibility on a view that may be wrong, sending it into a world that may not agree.
That is not a failure of rigor. It is the move the essay is about. Building cathedrals we may not see completed. Planting trees we will not sit under. Believing before the evidence is enough.
The machine outputs 2,400 words and stops. The human wonders if they were the right ones, and begins again.
If you’re working on agents or care about this question: zguo0525@berkeley.edu · @Zhen4good
]]>April 2026
Everyone is building org-chart agents.
Planner → Executor → Critic. Manager → Worker → Validator. Trees of LLMs wired into patterns that look suspiciously like the company structure on a whiteboard.
It feels right because it mirrors how we organize. But humans invented hierarchy to compensate for being human: slow sequential thinking, limited bandwidth, memory that doesn’t transfer between heads. Machines have none of these constraints, so when you wire LLMs into a fixed hierarchy you aren’t solving a coordination problem—you’re importing one.
I’ve been watching the 2025–2026 papers converge on the same conclusion from different angles. The pattern is hard to miss once you see it.
The empirical record keeps saying the same thing.
MacNet (ICLR 2025) scaled to a thousand agents and found irregular topologies beat regular ones—chains, stars, trees plateaued early. Kim et al. (2025) measured 17× error amplification in uncoordinated groups, with performance swinging ±80% on architecture alone.
Graph of Thoughts (AAAI 2024) delivered 62% quality improvement over Tree of Thoughts at 31% lower cost. Shen et al. (EMNLP 2025) showed query-adaptive topologies hitting 91.38% on MMLU and GSM8K, beating every fixed shape.
Even the field’s trajectory—Chain of Thought → Tree of Thoughts → Graph of Thoughts—points the same way. The next step isn’t another fixed structure. It’s no fixed structure at all.
Here’s the move: topology stops being input. It becomes output.
Static systems hardcode human assumptions about how agents should interact; dynamic systems let the task dictate the shape. A cheap LLM call—$0.001, 200ms—reads the problem and emits the graph: which models to call, in what order, with what parallelism, with back-edges where needed.
# Human designs topology
topology = org_chart(human_team)
result = topology.execute(task)
# Model designs topology
topology = llm("emit graph for this task", task, available_agents)
result = topology.execute()
Sometimes the model emits a chain (data pipeline). Sometimes parallel fan-out (brainstorm). Sometimes a loop with back-edges (design → code → critique → redesign). Sometimes a hierarchy—but because the task demands it, not because your framework ships it as a default.
The topology is not architecture. It’s output—chosen from a design space you define, not infinite chaos.
Dynamic topology doesn’t eliminate design. It moves it one level up.
You still design the agent pool—heterogeneous beats homogeneous, with X-MAS (2025) reporting +47% on AIME by mixing reasoners and chatbots instead of stacking identical models. You still design the edge vocabulary: what can flow between nodes, what kinds of back-edges are allowed, whether agents share state or only messages. You still design the density default—moderate sparsity with small-world structure (local clustering plus a few long-range shortcuts) is the strongest known inductive bias across Shen et al. and Wang et al.. And you still design the objective and constraints: whether the solver optimizes cost, latency, quality, or consistency, and what bounds it must respect—max depth, max cost per query, safety rules.
What’s still open: the right task-to-objective mapping, when to regenerate the graph versus reuse a cached one, whether topology should change within a single execution, how dynamic graphs degrade under adversarial agents, and what the theoretical frame is for scaling. The research is early. Treat the rules above as starting defaults, not a finished system.
This is more work than hardcoding a pipeline—much more—but the schlep is where the advantage is. Everyone else is wiring graphs by hand because the alternative is harder.
Compiler writers don’t write machine code—they design the rules that generate it. Agent builders should do the same, and be honest about which rules we don’t yet have.
Static agent topologies are training wheels for a bicycle that doesn’t need them.
The best agent system is not the one with the cleanest org chart. It is the one you cannot draw—because it vanishes into the problem it solves.
Stop wiring. Start compiling.
If you’re building agent compilers: zguo0525@berkeley.edu · @Zhen4good
]]>March 2026
I’ve never started a company. But I’ve read every word Paul Graham has ever published — all 231 essays. Here’s what I learned about what works, what doesn’t, and why most people get it wrong.
This is Graham’s most famous essay, and the one that changed how I think about building anything.
Successful startups require founders to do manual, labor-intensive work that cannot possibly scale. Airbnb’s founders went door to door photographing apartments. Stripe’s founders personally installed their payment system on users’ laptops. “Do Things that Don’t Scale” argues that none of this was optional — it was the reason these companies took off.
“Startups take off because the founders make them take off.” Not because they designed the perfect system. Not because they had the right architecture. Because they did the ugly, unscalable work nobody else wanted to do.
This connects to “Startups in 13 Sentences”: understand your users. That’s it. Everything else — cofounders, spending, morale — supports this one mission. Serve a small number of users intensely rather than many users superficially. You can be obsessed in a way that large organizations simply can’t — because you’re small enough to still hear the user’s voice directly, and you have nothing to lose by listening.
In “Schlep Blindness”, Graham explains why the best opportunities hide behind tedious, difficult work that founders unconsciously avoid. A company is defined by the schleps it will undertake.
Stripe is the perfect example. Thousands of developers knew payments were broken. They all chose easier paths. The Collison brothers didn’t — and built a $95B company.
I think about this constantly. Smart people consistently choose elegant problems over important ones. It’s a habit that feels productive but isn’t. The ugly problems — the ones that make you groan — those are the ones worth solving. Everyone else is avoiding them. That’s your opening.
Graham defines a startup not as a new company, not as a tech company, but as a company designed to grow fast. Growth rate is the compass for every decision. “Startup = Growth” argues that small differences in weekly growth — 1% vs. 5% vs. 10% — produce dramatically different outcomes over time because growth compounds.
“Superlinear Returns” generalizes this: small differences in quality produce dramatically disproportionate outcomes. If your product is half as good as your competitor’s, you don’t get half as many customers. You get none. The returns are exponential, not linear.
“Default Alive or Default Dead?” asks the question every founder should answer weekly: given current revenue, growth, and expenses, will you reach profitability before running out of money? Surprisingly many founders can’t answer this. The biggest killer is overhiring — confusing the correlation between successful companies having many employees with causation.
“Black Swan Farming” changed how I think about good ideas. The best startup ideas initially look like bad ideas. Airbnb — strangers sleeping in your house? Reddit — a link aggregator? If an idea seems obviously good, it’s probably too late.
“Frighteningly Ambitious Startup Ideas” goes further: the most transformative opportunities are psychologically repellent. They threaten your sense of identity and ambition, causing you to subconsciously filter them out. Pursue them indirectly — start small and build toward enormous goals.
“How to Get Startup Ideas” ties it together: the best ideas come not from brainstorming but from living in the future and noticing what’s missing. “Live in the future and build what seems interesting.” Founders who deliberately try to think of startup ideas usually generate bad ones. The good ideas emerge organically from solving problems you personally encounter.
Most people are allergic to ideas that look wrong. They optimize for looking right — polished pitches, safe bets, conventional wisdom. That’s a death sentence for a startup.
Graham nailed this in “The Hardest Lessons for Startups to Learn”: release early and iterate, because user feedback is invaluable. “If you’re not embarrassed by the first version of your product, you’ve launched too late.”
Perfection before launch is a disease. Startups that catch it usually die from it. You don’t know what users want until you put something in their hands and watch them use it — or don’t.
The startup’s only advantage is speed. You can outmaneuver, not outmuscle. You can iterate faster than incumbents because you have no legacy to protect, no committee to convince, no brand guidelines to follow. Use it. Ship ugly. Learn fast. Repeat.
“Maker’s Schedule, Manager’s Schedule” explains why this speed matters at the daily level too. Makers need long, uninterrupted stretches for deep work — a single meeting can destroy an entire afternoon. Protect maker time like a religion. The best teams do. The worst teams fill it with alignment meetings. The difference in output is staggering.
Graham’s essays on founders describe a specific kind of person.
“Relentlessly Resourceful” — the two-word description of the ideal founder. Not just determined (which implies brute force) but resourceful (which implies constant improvisation). The opposite is “hapless” — passively buffeted by circumstances.
“The Right Kind of Stubborn” draws a crucial line between persistence and obstinacy. Obstinate people cling to their initial ideas and resist feedback. Persistent people remain attached to the goal but continuously adjust their approach. True persistence requires five qualities simultaneously: energy, imagination, resilience, good judgment, and clear goal focus.
“Earnestness” — doing something for the right reasons while trying as hard as you can. In a world of cynics chasing exits, earnest people stand out. They exhibit a beneficial naivete about problem difficulty that helps them overcome obstacles cynics would never attempt.
“Mean People Fail” — being mean makes you stupid because it diverts energy from solving real problems. Kind founders who grow just one percentage point faster per week will crush aggressive competitors, because growth compounds but rapacity doesn’t.
“You Weren’t Meant to Have a Boss” — humans evolved to work in small groups of 8 to 20 people. Large organizations force hierarchical structures that progressively constrain freedom. Graham compares corporate jobs to junk food — superficially appealing but lacking essential qualities. In a startup, you don’t route around the bureaucracy. You just build.
This isn’t about startups. It’s about everything.
Graham’s deepest conviction, repeated across dozens of essays: writing is not just a way to communicate ideas — it’s a way to have them.
In “Putting Ideas into Words”: “Writing about something, even something you know well, usually shows you that you didn’t know it as well as you thought.”
In “The Need to Read”: complex problem-solving benefits from writing, writing skill depends on reading, therefore people who want ideas can’t afford to stop reading.
In “Writes and Write-Nots”: AI will split society into people who write (and therefore think clearly) and people who don’t. “If you’re thinking without writing, you only think you’re thinking.”
The difference between writing to communicate and writing to think is enormous. A pitch deck has a predetermined conclusion — you’re arguing for a decision already made. An essay has no predetermined conclusion — you’re discovering what you think by writing it down.
After reading 231 essays, here’s what I’d tell myself:
1. Solve one real problem. Not a feature. Not a platform. One problem that hurts someone enough to pay for it. Live in the future and build what seems interesting. Don’t brainstorm startup ideas. Notice gaps.
2. Ship before you’re ready. Perfect is the enemy of done. Ship ugly, learn fast, iterate. The startup’s only advantage is speed.
3. Do the schlep. The best opportunities hide behind work nobody wants to do. If it makes you groan, it’s probably worth pursuing. Everyone else is avoiding it. That’s your moat.
4. Care more than everyone else. You can’t out-scale the big guys. But you can out-care them. You can be the one who answers user emails at 2am. You can be the one who notices the detail everyone else missed. That’s the startup’s only real unfair advantage.
5. Stay weird. The best ideas start as bad ideas. Most people optimize for looking right. Startups should optimize for being right — even when it looks wrong. Embrace it.
6. Write to think. Don’t just build. Write about what you’re building and why. You’ll discover what you actually believe. And if you can’t explain it clearly in writing, you don’t understand it yet.
I’m still figuring this out. I have no company to point to, no exit to brag about. But I’ve read the closest thing Silicon Valley has to a philosophy of work — all 231 essays of it.
Here’s what I know:
Most people optimize for the wrong things. They chase prestige instead of genuine interest. They build what sounds impressive instead of what’s needed. They follow fashion instead of curiosity. They hire for credentials instead of ability. Graham wrote about this for 25 years.
The antidote is simple. Pay attention to reality. Talk to users. Do the ugly work. Ship early. Write to think. Stay weird.
The next great product won’t come from someone who played it safe. It’ll come from someone obsessive enough to care more, fast enough to iterate, and weird enough to see something others missed.
Maybe that’s you.
zguo0525@berkeley.edu · @Zhen4good
]]>March 2026
I read all 231 of Paul Graham’s essays. Every single one, from “Programming Bottom-Up” (1993) to “The Brand Age” (2025). Here’s what he said, what he meant, and what I think matters most.
Paul Graham is one of maybe five people who shaped how Silicon Valley thinks. He cofounded Viaweb (sold to Yahoo for $49M), created Y Combinator (which funded Airbnb, Stripe, Dropbox, Reddit, Coinbase), and wrote essays that became the intellectual operating system for a generation of founders.
His essays are not business advice. They are a philosophy of work, thinking, and life — disguised as startup wisdom.
PG’s most famous essay, “Do Things that Don’t Scale”, argues that successful startups require founders to do manual, labor-intensive work early on — recruiting users one by one, providing personal attention that cannot possibly scale. “Startups take off because the founders make them take off.” Airbnb’s founders went door to door photographing apartments. Stripe’s founders manually installed their payment system on users’ laptops.
This connects to “How to Get Startup Ideas”: the best ideas come not from brainstorming but from living in the future and noticing what’s missing. “Live in the future and build what seems interesting.” Founders who deliberately try to think of startup ideas usually generate bad ones. The good ideas emerge organically from solving problems you personally encounter.
“Schlep Blindness” explains why the best opportunities go unexploited — they hide behind tedious, difficult work that founders unconsciously avoid. Stripe succeeded because thousands of developers recognized the payments problem but chose easier paths. A company is defined by the schleps it will undertake.
“Black Swan Farming” reveals why the best investments look like bad ideas at first. Effectively all returns in startup investing come from a few massive winners, and those winners initially seemed implausible. Airbnb — strangers sleeping in your house? Reddit — a link aggregator? The pattern is consistent: if a startup idea seems obviously good, it’s probably too late.
“Startup = Growth” defines what a startup actually is — not a new company, not a tech company, but a company designed to grow fast. Growth rate is the compass for every decision. Small differences in weekly growth (1% vs. 5% vs. 10%) produce dramatically different outcomes over time because growth compounds.
“Superlinear Returns” generalizes this: in most domains, performance returns are not linear. Small differences in quality produce dramatically disproportionate outcomes. If your product is only half as good as your competitor’s, you don’t get half as many customers — you get none. Two mechanisms drive this: exponential growth (success compounds) and thresholds (winner-take-all dynamics).
“Default Alive or Default Dead?” asks the question every founder should answer weekly: given current revenue, growth rate, and expenses, will we reach profitability before running out of money? Surprisingly many founders can’t answer this. The biggest killer is overhiring — confusing the correlation between successful companies having many employees with causation.
“Relentlessly Resourceful” — if you had to describe the ideal startup founder in two words, these are it. Not just determined (which implies brute force) but resourceful (which implies constant improvisation against novel obstacles). The opposite is “hapless” — passively buffeted by circumstances.
“The Right Kind of Stubborn” draws a crucial line between persistence and obstinacy. Obstinate people cling to their initial ideas and resist feedback. Persistent people remain attached to the goal but continuously adjust their approach. True persistence requires five qualities simultaneously: energy, imagination, resilience, good judgment, and clear goal focus.
“Founder Mode” distinguishes between two fundamentally different approaches to running companies. Founders have been receiving advice designed for professional managers — “hire good people and give them room” — which is so ineffective for founders that it feels broken. Founder mode means maintaining deeper involvement across the organization, breaking traditional hierarchy constraints.
“Earnestness” — the quality PG most values in founders. Doing something for the right reasons while trying as hard as you can. Earnest people often exhibit a beneficial naivete about both human motivations and problem difficulty, which paradoxically helps them overcome obstacles that cynics would never attempt.
“Mean People Fail” — being mean makes you stupid because it diverts mental energy away from solving real problems. As society shifts from zero-sum competitions toward innovation-driven enterprises, the historical advantage of ruthlessness disappears. Nice founders who grow 1% faster per week will quickly surpass aggressive competitors, because growth compounds but rapacity does not.
This is arguably PG’s deepest conviction, repeated across dozens of essays.
“Putting Ideas into Words”: writing forces intellectual rigor. “Writing about something, even something you know well, usually shows you that you didn’t know it as well as you thought.” Unlike conversation, where tone masks unclear thinking, writing demands explicit clarity. “No one who hasn’t written about a topic has fully formed ideas about it.”
“The Need to Read”: reading is irreplaceable because it teaches writing, which is itself a thinking tool. “Writing is not just a way to convey ideas, but also a way to have them.” Complex problem-solving benefits from writing, writing skill depends on reading, therefore “people who want to have ideas can’t afford to” abandon reading.
“Writes and Write-Nots” predicts AI will split society into people who write (and therefore think clearly) and people who don’t. Drawing on Leslie Lamport’s observation: “if you’re thinking without writing, you only think you’re thinking.”
“Write Simply”: simple writing exposes weak thinking that fancy prose can mask. Simplicity is both an act of respect toward readers and a test of intellectual honesty.
“Write Like You Talk”: most people adopt artificially formal styles. The fix: revise by asking “Is this how I’d say it to a friend?” This puts you ahead of 95% of writers.
“Good Writing”: the two dimensions — sounding good and being right — are deeply interconnected. When writers tighten prose, they improve ideas. Like airplane design: if it looks good, it will fly well.
“How to Get New Ideas”: novel ideas come from identifying anomalies — things that appear unusual, absent, or broken. Knowledge expands in a fractal pattern where edges reveal numerous gaps on close examination. Creative breakthroughs come from investigating frontier gaps.
“The Bus Ticket Theory of Genius”: genius requires natural ability, determination, and — most underappreciated — obsessive, disinterested interest in a topic, pursued for its own sake. This obsessive curiosity lets geniuses persist down unpromising paths that more conventionally ambitious people would abandon.
“Beyond Smart”: intelligence is necessary but insufficient for generating new ideas. Society overvalues raw intelligence because it’s easily measured. The optimistic reframing: while intelligence is inborn, the other essential ingredients — obsessive interest, independent-mindedness, writing ability, work ethic — are all cultivable.
“How to Do Great Work”: choose meaningful work aligned with your aptitude, learn enough to reach the frontier, notice gaps others overlook, and explore promising ones. The engine is curiosity. “Writing a page a day doesn’t sound like much, but you’ll write a book a year.”
“Keep Your Identity Small”: discussions about politics and religion fail not because the topics are inherently hard but because people tie them to their identities. Once a belief becomes part of who you are, you can no longer think clearly about it. Minimize the labels you adopt.
“How to Think for Yourself”: independent-mindedness has three components — fastidiousness about truth, resistance to being told what to think, and curiosity. Cultivate it by surrounding yourself with independent-minded peers, seeking diverse perspectives, and habitually asking “is that true?”
“What You Can’t Say”: every era has moral fashions — invisible taboos mistaken for timeless truths. If you hold no opinions you’d hesitate to express, you’re likely just thinking what you’re told. Methods for identifying hidden taboos: notice what provokes disproportionate anger, track dismissive labels that shut down discussion, compare beliefs across times and cultures. Think freely but pick your battles about what to say publicly.
“The Four Quadrants of Conformism”: people vary along two axes — conventional vs. independent-minded, and passive vs. aggressive. Aggressively conventional people cause disproportionate societal harm. The customs protecting free inquiry have recently weakened, particularly in universities.
“Heresy”: the concept of heresy — punishment for forbidden beliefs — has reentered secular society. Two distinctive features: truth becomes irrelevant (labels end discussion without examining accuracy) and consequences are disproportionate (one controversial statement outweighs everything else). The origins: naturally intolerant people organized by ideology emerging from late-1980s universities.
“How to Work Hard”: three equally necessary ingredients for achievement — natural ability, practice, and sustained effort. The hardest part is honestly distinguishing real work from busywork and finding your sustainable intensity limit.
“The Top Idea in Your Mind”: everyone has one dominant idea that gets disproportionate mental energy through subconscious “ambient thought” — the thinking you do in the shower. Guard what becomes this idea. Disputes over money can displace creative work, and “it is nearly impossible to do a great job on anything that is not the thing you think about in the shower.”
“Maker’s Schedule, Manager’s Schedule”: there are two fundamentally different ways of using time. Managers operate in one-hour blocks. Makers need long, uninterrupted stretches for deep work. A single meeting can destroy an entire afternoon for a maker. Most organizational friction comes from managers not understanding this cost.
“Good and Bad Procrastination”: procrastination is unavoidable. Three types based on what you do instead: nothing (bad), less important work (mediocre), or more important work (good). The most productive people deliberately avoid trivial tasks to pursue significant work.
“Life is Short”: not a cliche but a measurable reality — you get about 52 weekends with a two-year-old. Ruthlessly eliminate “bullshit” (unnecessary meetings, online arguments, addictive pastimes). Act now rather than assuming future opportunities.
“How to Do What You Love”: doing work you love is essential for both success and happiness, yet most people are misled by the false dichotomy that “work equals pain.” Two forces derail you: prestige (social approval pushing toward impressive-sounding careers) and money. The path forward requires discipline, experimentation, and honest self-assessment.
“What Doesn’t Seem Like Work?”: the best way to identify work you’re suited for is to notice what feels effortless to you but burdensome to others. “What seems like work to other people that doesn’t seem like work to me?”
“A Project of One’s Own”: people working on self-directed projects feel “awake and alive” in ways that obligatory work never produces. Prior independent projects were far more predictive of startup founder success than academic grades.
“When To Do What You Love”: if you need moderate wealth you typically cannot afford to follow passion, but at the extremes — modest income or aiming for enormous wealth — passion becomes strategically sound. The best startup ideas emerge from passionate exploration.
“How to Raise Money”: treat fundraising as a distinct, time-limited activity. “Talk to investors in parallel, prioritized by expected value, and accept offers greedily.”
“How to Convince Investors”: focus on building something genuinely worth investing in, then convince yourself of its merit before approaching investors. Authentic confidence grounded in truth proves far more persuasive than salesmanship.
“Investor Herd Dynamics”: investor decisions are heavily influenced by other investors. Initial funding commitments make subsequent fundraising easier. A stampede can start from one strong commitment.
“The Fatal Pinch”: recognize when your startup has depleted its runway despite cash in the bank. Take immediate action — cut expenses, pivot, or shut down — rather than remaining in denial about fundraising prospects.
“A Fundraising Survival Guide”: fundraising can destroy morale through prolonged uncertainty and rejection. Maintain low expectations, keep building your product during fundraising, accept reasonable offers, and work toward ramen profitability so you’re never desperate.
“The Hacker’s Guide to Investors”: most VCs practice “momentum investing” — noticing what’s already taking off rather than predicting winners. They are dealmakers who read people and structure deals, not technologists who evaluate products. Prioritize angel investors, maintain momentum, and never appear desperate.
“Ramen Profitable”: earning just enough to cover founders’ basic living expenses fundamentally changes trajectory — it shifts the default from dying to surviving, gives leverage with investors, and frees founders from the distraction of fundraising.
“Beating the Averages”: startups can gain decisive competitive edges by adopting powerful but unconventional technologies. The “Blub paradox” — programmers systematically underestimate languages more powerful than what they know, so competitors will neither recognize nor replicate the advantage.
“Revenge of the Nerds”: programming languages differ enormously in power, yet most managers treat them as interchangeable. Companies using genuinely more powerful languages develop software dramatically faster.
“Succinctness is Power”: a language’s power should be measured by how succinct it makes programs. Since programmers produce roughly the same volume of code per day regardless of language, a more succinct language directly translates to more functionality per unit of time.
“Hackers and Painters”: hacking is fundamentally creative, more akin to painting than science. Hackers are makers who learn through practice and iteration, not researchers expected to publish papers.
“Great Hackers”: the productivity variation among programmers is far larger than in most fields. Great hackers are motivated by interesting problems, superior tools, quiet environments, and talented collaborators — not money.
“The Other Road Ahead”: server-based web applications represent the future, offering users easier access while freeing developers from desktop distribution constraints. This enables startups to compete with incumbents.
“The Acceleration of Addictiveness”: technological progress inevitably creates increasingly addictive products. Society develops defenses only slowly, leaving individuals vulnerable to engineered compulsions. Actively resisting new temptations is now necessary for living well.
“How to Make Wealth”: wealth is not money — wealth is tangible value people want. Startups offer a compressed path to creating wealth through technology and leverage. Getting rich requires measurable performance plus leverage.
“How People Get Rich Now”: in 1982, 60% of the richest Americans inherited wealth; by 2020, that dropped to 27%, with three-quarters of new fortunes from founding companies. The shift back to founder wealth is the historical norm, not the anomaly.
“Economic Inequality”: inequality stems from multiple causes — some harmful (rent-seeking, corruption) and some beneficial (innovation, wealth creation). Conflating all forms leads to counterproductive policies. Target poverty, social mobility, and corruption directly.
“Mind the Gap”: increasing income inequality often reflects genuine differences in wealth creation capacity. People carry a “Daddy Model” where wealth appears distributed by authority rather than created through effort.
“The Refragmentation”: mid-20th century American cultural cohesion was an anomaly caused by WWII and national oligopolies. Natural fragmentation returned as those forces faded. Rather than restoring artificial unity, acknowledge fragmentation as the natural state.
“Cities and Ambition”: different cities send distinct messages — New York says “make more money,” Cambridge says “be smarter,” Silicon Valley says “be more powerful.” These ambient signals are powerful enough to override willpower.
“The Origins of Wokeness”: traces wokeness as a modern manifestation of an ancient tendency — moral priggishness. “The performativeness, not the social justice” is the real problem. Elaborate, constantly changing rules became a substitute for actual virtue.
“Lies We Tell Kids”: adults systematically deceive children across many domains. The most damaging institutional lie is that success comes from following rules, when rules exist to serve institutional efficiency. Adults should actively unwind misleading narratives they absorbed as children.
“The Lesson to Unlearn”: school teaches “hacking bad tests” — gaming evaluation systems rather than genuinely learning. This habit becomes a liability when real results matter more than credentials.
“How Y Combinator Started”: YC’s most transformative idea — funding startups in synchronous batches — emerged almost accidentally from an educational experiment. What began as a way to learn angel investing became a revolutionary model.
“What I’ve Learned from Users”: most startups have the same problems in different forms. Founders frequently misdiagnose their actual problems. The fundamental formula: “Speed defines startups. Focus enables speed. YC improves focus.”
“The Airbnbs”: when Airbnb applied to YC, the company was nearly dead — maxed credit cards, universal investor rejection. What set them apart was relentless energy (Brian Chesky: “The Tasmanian Devil”) combined with genuine conviction born from firsthand experience.
“What We Look for in Founders”: five key qualities — determination, flexibility, imaginative intelligence, willingness to bend rules creatively, and strong co-founder relationships. Determination and adaptability outweigh raw intelligence.
“Before the Startup”: the way to get startup ideas is not to try to think of startup ideas. Develop deep expertise, work on genuinely interesting problems, and let business opportunities emerge naturally.
“The 18 Mistakes That Kill Startups”: the single most critical factor is effort — the biggest mistake is not trying hard enough. Many potential founders never commit because they secretly lack confidence.
“Is There Such a Thing as Good Taste?”: yes. Denying good taste requires rejecting good art entirely, which requires denying that anyone can be skilled at any artistic pursuit — an absurdity. Good taste is achievable through deep knowledge and clarity of mind.
“Taste for Makers”: good taste is not subjective preference but a learnable skill following universal principles. Simplicity, timelessness, solving the right problem — these recur across mathematics, engineering, art, and writing.
“How Art Can Be Good”: since humans share biological and perceptual traits, objective standards for art do exist. Good art achieves its purpose of engaging its audience particularly well.
“Copy What You Like”: develop taste by imitating things you genuinely enjoy, not things that seem impressive. Examine your guilty pleasures as pure indicators of taste.
After reading all 231 essays, a single pattern emerges. PG keeps circling the same core insight from different angles:
Most people optimize for the wrong things. They chase prestige instead of genuine interest. They build what sounds impressive instead of what’s needed. They write to sound smart instead of to think clearly. They follow fashion instead of curiosity. They hire for credentials instead of ability. They seek comfort instead of growth.
The antidote is always the same: pay attention to reality. Talk to users. Write to discover what you actually think. Notice what’s broken. Do the tedious work nobody else wants to do. Measure what matters. Stay curious. Ship early. Iterate.
It’s not complicated. It’s just hard.
This summary covers all 231 essays published at paulgraham.com/articles.html. Each essay was read in full and distilled to its core insight.
]]>March 2026
This essay is not advice from someone who’s made it. It’s observations from someone who’s been paying attention.
The easiest way to kill a startup is to build something that already exists.
Me-too products are seductive because there’s a proven market. You can point to a competitor and say, “We’re like that, but better.” That’s the trap.
When your product is not fundamentally new, users compare. They’ll stack your feature list against the incumbent’s. They’ll ask, “Why should I switch?” And suddenly you’re in a game you can’t win — a game of incremental improvements, cost reduction, and scale that big companies play much better than you.
Big companies can win that game. They have distribution, brand recognition, and margins to subsidize features. If you’re a startup chasing a me-too product, you’re signing up for a race where the finish line keeps moving. You’re not competing on product. You’re competing on capital.
The only way out is to build something different enough that comparison becomes irrelevant. Not “better.” Different. Something that creates its own category, solves a problem others aren’t solving, or solves it in a way that feels fundamentally new.
This doesn’t mean you need to invent new technology. It means you need to find a unique angle, a unique audience, or a unique philosophy that becomes your moat. The question isn’t “How are we different?” The question is “Why would someone choose us when the other option exists?” If you can’t answer that honestly, don’t start.
If I learned one thing at Meta, it’s this: speed is a feature.
Startups die from perfectionism. They spend eighteen months polishing a product nobody asked for. They ship late, launch stale, and wonder why users moved on. The truth is, you don’t know what users want until you put something in their hands and watch them use it — or don’t.
This sounds obvious. Everyone says “move fast.” But what it actually means is uncomfortable: you have to ship things that embarrass you. You have to accept that your first version will be wrong in ways you can’t predict. You have to listen to users tell you your baby is ugly, and then fix it fast enough that they stay around to see version two.
At Apple, I saw the opposite extreme — and it taught me the same lesson. Apple’s philosophy is to wait until something is perfect before shipping. That works when you’re Apple. You have years of runway, a brand that survives missteps, and a distribution channel that can make up for late delivery. For a startup, that patience is a luxury you can’t afford.
The startup’s only advantage is speed. You can outmaneuver, not outmuscle. You can iterate faster than incumbents because you have no legacy to protect, no committee to convince, no brand guidelines to follow. Use it. Ship ugly. Learn fast. Repeat.
Here’s the part that’s hard to teach.
The best product doesn’t win because it has the most features. It doesn’t win because it’s cheapest. It wins because someone cared more about the user’s problem.
I think about this every time I use a product that clearly wasn’t built by anyone who’d actually use it. The features are there. The UX is “fine.” But something feels off — like no one ever sat with a real user and felt their pain. These products are technically adequate and fundamentally broken.
The products that break through are the ones where you sense, even before you understand why, that someone behind the screen gives a damn. They anticipated your problem. They sweated the edge cases. They didn’t cut corners because “good enough” wasn’t good enough for them.
This is the startup’s real unfair advantage. Big companies optimize for engagement, retention, metrics that proxy user value but aren’t user value. Startups can optimize for actually solving someone’s problem. You can be obsessed in a way that large organizations simply can’t be — because you’re small enough to still hear the user’s voice directly, and you have nothing to lose by listening.
The product that solves the problem most urgently, most empathetically, most relentlessly — that’s the one users stick with. Not because it’s perfect. Because they feel seen.
These three ideas reinforce each other: build different → iterate freely → learn what users need → care more → find your angle.
Don’t play the big players’ game. They have more money, engineers, and data. You have something they don’t: the freedom to be weird, to be wrong, and to care more about a problem they barely notice.
That’s how you make a great product.
zguo0525@berkeley.edu · @Zhen4good
]]>Every AI startup in 2025 follows the same nauseating playbook: raise $10M+ seed, hire 20 engineers, build slick demos in public, hope Sam Altman doesn’t clone you, then get acquired or die. Peter Steinberger said fuck that. He built OpenClaw — originally just a prototype in one hour — at night and on weekends, while working as a software engineer. Takeout ramen for company. No pitch deck. No CTO. No equity comp for early believers. Just code, conviction, and a lobster that made people smile.
That’s the whole story.
The industry doesn’t get it. They obsess over scalability, moats, and growth metrics. OpenClaw has none of that: not scalable (a personal tool, not a platform), no moat (open source, anyone can fork), slow growth (word of mouth, not growth hacks). Yet it’s worth more than half the YC batch combined. Why? Because it solves a real problem, built by a real person, and doesn’t feel designed by a committee optimizing for engagement.
Peter wasn’t trying to build a company. He had already exited PSPDFKit for €100 million in 2021 — a B2B PDF company that employed 70 people. After that, he felt lost. Partied. Moved around. Over three years, he worked on 43 experimental AI projects, rediscovering his love for coding. In November 2025, he connected a chat app to Claude in one hour, as a “simple local AI assistant toy.” He expected big companies to replicate it immediately. They didn’t. The community did.
The tech is not the point. Agent frameworks are a commodity — LangChain, AutoGPT, Claude Agent, Manus all work. None of them went viral. OpenClaw did for three reasons:
First, one man who actually ships. In a world of 50-person teams and $100M valuations, Peter shipped alone. His philosophy: “I ship code I don’t read.” No pivots, no corporate speak, no “we’re a team of 10 but actually 3 founders and some contractors.” Just one dude shipping at 2 a.m., iterating based on community feedback, releasing “rough, dangerous” code. That resonates. Deeply.
Second, local-first data ownership. This is the real differentiator. OpenClaw runs on your machine. Your data stays on your device. Not on Anthropic’s servers. Not on OpenAI’s cloud. Not training their models. Every other “personal AI” — Pi, Rabbit, PIN AI — still owns your data. OpenClaw hands it back. In an era of privacy anxiety, that narrative is unbeatable. It’s not a feature. It’s a philosophy.
Third, comes to you. OpenClaw doesn’t ask you to download a new app or create a new account. It connects to WhatsApp, Telegram, Discord, Slack — platforms you already use every day. You don’t go to AI. AI comes to you. This seems minor. It’s everything. Pi had to build a product. Rabbit had to build hardware. OpenClaw just needed to integrate. That’s leverage.
Compare this to the competition:
PIN AI — $10M from a16z, “all-star angels,” enterprise landing page, ecosystem diagrams. They did everything “right”: platform thinking, moats, growth metrics, investor logos. And they went nowhere.
Pi AI — $1.5B+ raised, co-founded by Mustafa Suleyman (DeepMind), positioned as “emotionally intelligent” personal AI. Massive funding, famous founder, enterprise polish. Still just another chatbot — and it owns your data.
Rabbit R1 — $20M+, sleek hardware, huge hype, sold hundreds of thousands of units. Then the reviews came in. Returns piled up. The hardware play crashed.
The irony is brutal: every “correct” startup recipe lost to one guy in one hour. They had the money, the talent, the press releases. OpenClaw had a problem it actually solved.
The true alpha isn’t the most funding, the biggest team, the best model, or the flashiest demo. It’s solving a problem you actually have, shipping when it’s ugly, optimizing for use not exit, and being weird, memorable, and human. Peter didn’t set out to build a company. He set out to build a tool. The rest followed.
In February 2026, OpenAI hired him to lead their personal AI agents division. Headlines wrote themselves: “OpenClaw creator joins OpenAI.” But the project, the community, the idea — still open source, still running on people’s machines, still theirs. Peter left. The lobster stayed.
That’s the point.
zguo0525@berkeley.edu · @Zhen4good
]]>February 2026
The AI industry spends over a hundred billion dollars a year making computers talk to themselves.
Not to users. To themselves. The internal monologue — the “thinking” that reasoning models do before answering — is now the majority of all tokens generated on Earth. Thousands of words of self-talk that no human reads, no human asked for, and no human would pay for if they understood what was happening.
The smarter the model, the more it talks to itself. That should strike you as suspicious.
Here’s how we got here.
Pretraining was hitting diminishing returns. Bigger models, more data, same benchmarks. The industry needed a new story. OpenAI found one: let models think before they answer. Chain of thought. Deliberation. Self-reflection.
It worked. On benchmarks, thinking models crushed everything. So everyone followed. A year later, every major lab ships reasoning models. The meta is clear: think harder, score higher.
But something strange is happening. Ask a reasoning model a hard question and it might generate 5,000 tokens of thinking to produce a 200-token answer. You pay for all 5,200. Only 200 were for you. The rest was the model having a conversation with itself — rehearsing, second-guessing, rephrasing, going in circles.
We built an industry around AI talking to humans. Then we discovered it mostly talks to itself.
The obvious defense: “But it makes the answers better!”
Sure. But why does it work? Not because thinking is inherently good. Because the architecture has no other option.
A transformer can only carry information forward by generating tokens. Each token updates the hidden state. No token, no state update. The model literally cannot think without speaking. The monologue is not a strategy. It’s a prosthesis.
Humans don’t have this limitation. When you recognize a friend’s face in a crowd, there’s no inner monologue. When a chess grandmaster sees the right move, she doesn’t narrate the search tree. Most human cognition is silent — patterns firing in neural substrate, not words forming in sequence.
Language is how we communicate thoughts. Not how we produce them. The AI industry confused the two, and now spends $100 billion a year on the confusion.
It gets worse. The returns are collapsing.
Going from zero thinking to some thinking is transformative. Going from some to a lot is incremental. Going from a lot to a massive amount is almost nothing.
This is not a scaling law. Scaling laws are the magic of deep learning — more compute, reliably better results. What we have with thinking tokens is a logarithmic curve pretending to be a power law. The industry is extrapolating a trend that’s already flattening, and building infrastructure on the extrapolation.
Ten thousand GPUs generating text that nobody reads. That’s not the future of intelligence. That’s a very expensive diary.
You can see the problem most clearly at the edges.
Voice assistants. If thinking takes 15 seconds, voice is broken. Nobody waits 15 seconds in a conversation. So you either give up on voice, or you give up on thinking. Today, most products give up on voice.
Real-time agents. An agent that takes 30 seconds to decide its next action is not an agent. It’s a committee. Reasoning models are too slow for the agentic future everyone is betting on.
Embodied AI. A robot that pauses to monologue before picking up a cup is a robot that drops the cup. Physical interaction happens at the speed of physics, not the speed of text generation.
The monologue is not just expensive. It’s a ceiling on what AI can become.
Some people think the answer is making the monologue faster. Distillation, speculative decoding, efficient attention. Make the same thinking happen in fewer milliseconds.
This is like making a faster horse. You’re optimizing within a paradigm that’s the problem.
Others think the answer is making the monologue shorter. Prune redundant reasoning steps. Stop the model when it’s confident. Tax verbosity during training.
Better. But still treating symptoms. You’re still paying the thinking tax — just a smaller one.
The real question is: why does thinking need to happen in language at all?
It doesn’t. And we already know this.
A few research groups have shown that you can achieve equivalent reasoning performance in the model’s hidden states — continuous vectors, not discrete tokens. No monologue. No text. Just computation happening in the space where the model actually lives.
The implications are hard to overstate. If reasoning doesn’t require language, then the entire infrastructure we’ve built around generating, storing, and billing for thinking tokens is unnecessary overhead. Not a moat. Not a feature. Overhead.
We are probably living through the vacuum tube era of AI reasoning. Everything works. Everything is expensive. And something much simpler is about to make it all obsolete.
I don’t know what replaces the monologue. Nobody does yet. The research is early, the results are promising, the engineering is hard. It might take two years or ten.
But I’m fairly confident about one thing: we will look back at this era and find it absurd. A hundred billion dollars a year, generating text that no one reads, in a medium the model doesn’t natively think in, to solve a problem that doesn’t require language.
The best thinking is the thinking you never see. The best AI is the one that just knows.
We’re not there. But the monologue is not the way.
zguo0525@berkeley.edu · @Zhen4good
]]>Most people are stuck in a rat race. They look around, see what others are doing, and try to do it slightly better—20% better, 50% better, maybe 100% better. This is a death spiral.
The only way to go up is to look up, not look around.
To escape the rat race, you need three things working together: help people solve real problems, love what you do enough to become the best, and be different enough to disrupt. They’re not separate principles—they’re a system. Each one reinforces the others.
Help people gives you direction. When you focus on solving problems others actually have, you’re not looking around at what competitors are doing—you’re looking up at what problems need solving.
But you can’t help people effectively unless you love what you do. When other people see it as a job, but for you it’s a game, you have a chance to be the best. The people who become the best aren’t just talented—they’re obsessed. They think about problems when they’re not working because they can’t help themselves.
But loving what you do and helping people isn’t enough if you’re just doing what everyone else is doing slightly better. That’s still the rat race. You need to be different—fundamentally different. You can’t win by being 20% or 50% better. You need to be 10x better to disrupt. The world is winner-takes-all. Being mediocre at anything is a risk.
Here’s how they reinforce each other: when you help people solve real problems, you’re more likely to find something you love. When you love what you do, you’re more likely to find the edge that makes you different. When you’re different enough to disrupt, you’re more likely to help people in ways others can’t.
When all three work together, something happens: the exponential curve. The more energy you put in, the more it comes back to help you. And the more you put in, the more it accelerates.
Early on, progress feels slow. But if you’ve found the right problem, the right passion, and the right differentiation, the curve accelerates. The gap between you and everyone else widens.
This is how you break the glass ceiling. You see the bigger picture. You aim for the best and top, not the middle. Put in the energy. Watch the exponential curve work for you.
If you’re building something or navigating your own path, I’m always open to conversations: zguo0525@berkeley.edu.
Liked this? Follow on X: @Zhen4good. Collaborations/advising: zguo0525@berkeley.edu • LinkedIn
]]>In June 2025, Apple published “The Illusion of Thinking”, questioning if AI models actually reason. Days later, WWDC 2025 announced Liquid Glass—a visual redesign, not AI advancement. The stock dropped 1.2% (CNBC).
That’s when I realized: the company I joined to build AI products was moving in a different direction. I’d spent a year shipping Visual Intelligence to millions of users, but the strategic shift was clear. The question wasn’t whether to stay or go—it was whether I wanted to build the future or watch it get built elsewhere.
At Apple, I worked on Visual Intelligence—on-screen search and smart event creation that launched at WWDC 2025. We shipped features, but the deeper challenge was organizational: how do you maintain urgency when a company has been dominant for over a decade?
Apple’s smartphone dominance created comfort. When the AI revolution accelerated, the cultural readiness wasn’t there. The response was telling: rather than building AI that works, the company published research questioning whether AI works at all.
While Apple was publishing critiques, DeepSeek R1, OpenAI’s Gibili, and Gemini Nano Banana were pushing boundaries. The competition in AI is unforgiving: you either ship something remarkable, or you fall behind.
Visual Intelligence shipped, and I’m proud of what we built. But the signals were clear: when your company publishes papers questioning whether AI works instead of building AI that works, that’s not a technical problem—it’s a strategic one.
Meta’s approach was different. In 2025, Mark Zuckerberg entered “founder’s mode”—personally spearheading AGI and ASI (Bloomberg).
The commitment was visible: Zuckerberg established Meta Superintelligence Labs (MSL), recruited Alexandr Wang (Scale AI) and Nat Friedman (GitHub), and offered $200-300M compensation packages to elite researchers (Bloomberg, CNBC, AM World Group). Infrastructure matched: “Prometheus” (1 GW by 2026) and “Hyperion” (up to 5 GW) data centers, plus custom silicon via MTIA (Tom’s Hardware, Meta Engineering). The company is building models from scratch, training Llama 3, and aligning teams with AGI objectives (Axios).
I joined Meta’s Superintelligence Labs. At Apple, I learned what it takes to ship AI products at scale. At Meta MSL, I’m learning what it takes to push capability frontiers. This positions me for where AI is heading.
My experience spans research (MIT, MIT-IBM Watson, papers like JetMoE, API Pack) and product (Apple’s Visual Intelligence, notification summaries). What works:
1. You need both capability and integration, but most teams optimize for only one. At MIT-IBM Watson, I pushed capability frontiers. At Apple, I learned integration. The teams that win do both. Apple had integration but wasn’t pushing capability. Meta builds models from scratch while also building integration systems.
2. Constraints teach optimization, but they can limit ambition. JetMoE taught me efficiency (Llama2 performance with $0.1M). Apple’s on-device constraints forced similar optimization. But constraints can become excuses. The best teams use constraints to force creativity, not limit ambition. Meta’s approach—infrastructure first, then capability—avoids this trap.
3. Integration is harder than capability. Notification summaries seemed straightforward but exposed the real challenge: making AI work reliably across edge cases, privacy boundaries, and organizational silos. Demos wow, but shipping is harder. At Apple, I learned benchmarks don’t matter if features break in production.
4. The gap between research and product is organizational, not technical. At Apple, split ownership between research (training models) and product (shipping features) created delays. At Meta MSL, research and product are integrated—we’re building capabilities that will become products. This alignment matters more than model size.
5. The best career moves align with where the industry is heading, not where it is. I left Apple not because the work was bad—Visual Intelligence shipped—but because I saw the strategic shift. When Apple published “The Illusion of Thinking” and announced Liquid Glass instead of AI breakthroughs, the signal was clear. Meta was going all-in.
If you want to create something great, you need to train your own models and build new capabilities from the ground up. You cannot win by adding rewrite or summarization features to existing apps.
The market rewards first-movers who create eye-catching products. It does not reward incremental add-ons. In an exponentially growing industry, the gap between the best product and everything else widens with each passing month.
In AI, either you are the best, or you are nothing. Companies that build foundational capabilities capture value. Companies that add AI features to existing products get left behind. This is why Meta’s approach matters: building models from scratch, investing in infrastructure, and assembling top talent isn’t just about competing—it’s about having a chance to win.
At Meta MSL, we’re building capabilities that don’t exist yet, not improving features that already do. The difference between shipping features and building the future: are you training models and creating new capabilities, or adding AI to what already exists?
That’s what I’m betting on.
For founders: The alpha vs. beta play matters more in AI. Companies building foundational capabilities capture value. Companies adding AI features to existing products get left behind. The companies that win are building capabilities that don’t exist yet.
For career decisions: Look for strategic signals, not role descriptions. When deciding to leave Apple, I looked at three things: (1) What is the company actually doing? Apple published “The Illusion of Thinking” and announced Liquid Glass—actions signaling a shift away from AI. (2) Where is the industry heading? Meta was going all-in on AGI; Apple was questioning whether AI works. (3) Does the culture match the ambition? Apple optimized for polish and stability; Meta MSL optimizes for speed and exploration.
Signals to watch: Are they publishing papers questioning AI or building AI that works? Investing in infrastructure or cutting costs? Recruiting aggressively or slowing hiring? These signal strategic commitment.
For evaluating opportunities: Capability and integration both matter. When I evaluate AI companies, I look for teams that do both. Apple had integration but wasn’t pushing capability. Some research labs push capability but can’t integrate. The teams that win do both.
For timing: In an exponentially growing industry, the gap widens fast. I left Apple because I saw the strategic shift happening. In AI, timing matters more. The difference between joining a company going all-in versus one questioning fundamentals compounds over time. The question isn’t just “is this a good role?” but “is this where the industry is heading?”
For founders building AI companies or anyone making career decisions in this field: align with where the industry is going, not where it is.
These are my own views, not those of Apple or Meta. I’m grateful for both experiences and the exceptional people I’ve worked with at each company. If you’re building an AI startup or navigating a career transition in this field, I’m always open to conversations: zguo0525@berkeley.edu.
Liked this? Follow on X: @Zhen4good. Collaborations/advising: zguo0525@berkeley.edu • LinkedIn
]]>