A Working Theory of AI
Where the value is going, from power plants to apps.
I update this regularly. As my thinking changes and the facts move, I revise it here instead of writing a new piece. This is the third version, last updated November 2026. Earlier versions were a Google Doc titled “Thoughts on AI and Opportunities.” A log at the end shows what changed each time.
In September 2025, I realized I didn’t know whether modern AI was another hype cycle or a fundamental shift in technology. To figure it out, I quit my job, went deep into the research papers, watched technical breakdowns, tested every AI product I could find, studied how it’s changing businesses, and started building. This note is a running synthesis of what I’ve learned since then, to check my own understanding and to share with anyone bored enough to read all of it.
Modern AI is a step-change in computing. Historically, computers were logic machines. They executed rigid, deterministic rules (if X, then Y) on structured data. They only “knew” what a programmer explicitly encoded. Modern AI flips this model. LLMs don’t just execute rules; they process knowledge. By ingesting vast amounts of unstructured information, they build probabilistic understandings of concepts and relationships, allowing them to synthesize new insights rather than just retrieving stored data. Diffusion models give us infinite visual creativity. By handling text, image, audio, and video inputs simultaneously, these tools allow us to approximate, and often exceed, human-level performance on tedious tasks, freeing us to focus on high-leverage work. However, they aren’t magic. They struggle without collaboration, especially in tasks requiring nuance (like negotiating a sensitive deal) or unavailable context (like a doctor diagnosing a patient without a history).
To understand where this is heading, and who stands to win, we need to understand three things: how the technology actually works (it’s simpler than you think), the industrial-scale buildout behind it, and why the products built on it are still so early. That last part is where the opportunity is.
Under the Hood: How AI Actually Works
At its core, AI is sophisticated autocomplete. The model predicts the next word in a sequence, one word at a time. But to predict well, it must understand concepts and relationships, not just patterns.
Consider this example:
Jack is playing with Jill. He picks up a stick and hands it to _ _ _ _ _ .
To predict the next word, the model must understand: Jack is male, Jill is female, “He” refers to Jack, and “hands it” involves two people. Based on these relationships, it assigns probabilities (“her” 90%, “Jill” 5%, “his” 5%) and continues from there. This simple mechanism, repeated trillions of times across the entire internet, builds up a web of conceptual understanding that can mimic human reasoning.
How models learn:
Building these models happens in three stages. First, pre-training: the model reads essentially the entire internet, learning how concepts relate to each other, the way you might connect “fire” with “heat,” “danger,” and “light.” At this stage, the model understands concepts but doesn’t know how to be helpful. It’s just a very sophisticated text predictor.
Second, fine-tuning: humans show the model thousands of examples of good question-and-answer pairs, teaching it how to package its knowledge into useful responses. For example:
- Q: “Give three tips for staying healthy.”
- A: “1. Eat a balanced diet with plenty of fruits and vegetables. 2. Exercise regularly. 3. Get enough sleep and maintain a consistent schedule.”
The model learns not just what to say, but how to format helpful answers, using numbered lists for “tips,” providing step-by-step instructions for “how to” questions, and so on.
Third, reinforcement learning: the model practices answering questions and gets scored on quality. Humans might see two responses to the same question and pick which one is better. The model learns to maximize this “human preference score.” For questions with objectively correct answers, like math problems or code that needs to compile, labs can skip human reviewers entirely and just check whether the solution works. This lets them train on millions of practice problems automatically.
Something surprising emerged from this process: models learned to “think step by step.” Consider this puzzle: “Mary’s father has five daughters: Nana, Nene, Nini, Nono. What is the fifth daughter’s name?” An early model might guess “Nunu” (completing the vowel pattern). A modern reasoning model works through it: “This is Mary’s father… he has five daughters… Mary must be one of them. The answer is Mary.” This chain-of-thought reasoning (now standard in every frontier model) emerged naturally from training models to solve problems correctly.
Beyond text:
Modern AI doesn’t just process words. It breaks images into patches (like cutting a photo into tiles), converts audio into patterns over time, and translates all of these into a shared “language of understanding.” This is why you can show Claude a screenshot and ask “What does this error mean?” or give Gemini a diagram and ask “How does this system work?” The model connects visual patterns with language concepts. It isn’t truly “seeing,” but it understands relationships between what things look like and what they’re called.
The practical result: today’s AI can handle text, images, audio, and video simultaneously, making connections across all of them to answer questions and solve problems.
The Industrial Chain: Constraints & Commodities
In the summer before my sophomore year of college, I worked at Atlantic LNG in Trinidad. It is a massive industrial facility designed to do one thing: cool natural gas down to -162°C so it can be liquefied and pumped into ships. I remember standing in my hardhat and fireproof overalls amidst the roar of the turbines and the maze of insulated pipes, awestruck by the sheer physical scale required to change the state of a gas. Today, when I look at the data centers powering modern AI, I get that same feeling. We are building infrastructure on the scale of the oil and gas industry, not to move fuel, but to move thought.
A year later, that comparison feels less like a metaphor. Photos of the new AI campuses look like refineries: substations, cooling plants, rows of buildings the size of aircraft hangars. The money is industrial too. A recent Brookings paper by Columbia’s Stijn Van Nieuwerburgh estimates the US data center buildout at about 3.6% of GDP a year through 2032. The railroad boom ran about 2.2% of GDP at its peak, and the interstates and the fiber buildout about 1.1% each. Relative to the economy, this is bigger than any of them. Anthropic alone has signed up for about $900 billion of compute and infrastructure, with payments running out to 2045.
The scale comes from how these models are built. Training is a long search: feed in examples, compare the model’s answer to the right one, nudge billions of variables closer, and repeat. GPT-4 took an estimated 2 x 10²⁵ calculations, a job that would take a standard laptop a million years. OpenAI did it in about three months with around 25,000 specialized chips drawing 10–15 MW, roughly the output of one gas turbine. That was 2022. Today’s frontier campuses are roughly a hundred times bigger. By Epoch AI’s estimate, xAI’s Colossus 2 in Memphis holds the equivalent of about 1.1 million of NVIDIA’s H100 chips, and Meta’s Prometheus campus in Ohio is bringing its first gigawatt online this year, enough to power hundreds of thousands of homes.
The Chips
NVIDIA has owned this market for more than a decade, thanks to a bet it made in 2006 on CUDA, the software that made it easy to run AI math across many GPUs. It’s still the default, especially for training. But buying everything from one supplier at that supplier’s margins is a strategic risk, so the biggest players built their own. Amazon and Anthropic’s Project Rainier switched on in October 2025 with nearly 500,000 of Amazon’s Trainium2 chips. Google designs its TPUs with Broadcom, and Anthropic has signed up for as many as a million of them. Anthropic’s S-1 lists four chip families it runs on: Trainium, TPUs, NVIDIA and AMD. The deals have changed shape too. The suppliers now finance the labs: Amazon and Google will each own more than 5% of Anthropic once it goes public, and Broadcom has a similar arrangement. I dug into what that means in my piece on Anthropic’s S-1.
The Hard Limits: Energy and Data
The first hard limit is energy. Goldman Sachs expects US data center power demand to grow about 38% in both 2026 and 2027, adding roughly 12 GW this year and 17 GW next on top of about 31 GW in 2025. The grid can’t keep pace. In 2025, Texas was receiving tens of gigawatts of data center requests a month while approving about a gigawatt a year. With an AI cloud generating $10–12 billion per gigawatt a year, waiting five years for a grid connection is expensive, so the labs are bringing their own power. By mid-2026, about 2 GW of “behind-the-meter” generation was running at US data center projects, most of it xAI’s, whose Colossus 2 can run entirely on its own gas turbines. Meta is paying its Louisiana utility to build new gas plants for its Hyperion campus. The hyperscalers are becoming energy companies, and with solar and wind intermittent and nuclear years away, natural gas is filling the gap. The folks in Landman may be in a boomtown soon.
The second limit is data, and this is where my first version undersold how industrial things have gotten. The labs have already used up the public internet, so they went looking for more. Anthropic’s Project Panama bought millions of used books, cut off the spines and scanned the pages. (A judge ruled that was fair use; the $1.5 billion settlement was over pirated copies it had downloaded separately.) The bigger shift is data that doesn’t exist yet. Labs pay doctors, lawyers, mathematicians and engineers $85 to $125 an hour to write hard problems, worked solutions and grading rubrics. Mercor reported a roughly $2 billion gross run rate by mid-2026, and Surge AI, Scale AI and Handshake are each around $1 billion or more. Raw text is the crude. Expert-built data is the refined product: it teaches a model how to think, not just what to know.
The strangest new source is the inside of real companies. In August, Google paid about $10 million in a bankruptcy auction for Spirit Airlines’ internal records: roughly 100 million emails, hundreds of millions of Teams chats and 30 million lines of code, with passenger data left out. Labs and data vendors are also buying the leftovers of failed startups (Slack archives, Jira tickets, codebases) to build simulated workplaces where agents can practice real work. If the next frontier is agents that do jobs, the labs need the messy record of how jobs actually get done.
Are We Overbuilding?
Probably, at least in places. Almost every big US infrastructure buildout has come with some: railroads, highways, broadband. On the Brookings podcast about Van Nieuwerburgh’s paper, the host noted that expansions like these usually end with some overinvestment and some sort of recession, and Gartner estimates that more than $3.5 trillion of AI infrastructure spending could be exposed to impairment by 2030.
But overbuilding usually hurts the people who build the infrastructure more than the people who build on top of it. The fiber laid in the ’90s eventually got used; many of the companies that laid it didn’t survive to see it. What survived was cheap bandwidth, and a generation of products ran on it. If AI compute gets overbuilt, it gets cheap, and that’s good news for anyone building products.
That’s why, after digging into Anthropic’s S-1, I think the compute capacity question matters less than the production adoption question: will enough products get built to fill all this capacity? For all the money going into the ground, we’re still remarkably early in building the products that use it. That’s the rest of this note.
The Interface Gap: Why Chatbots Are the MS-DOS of AI
So, we have a new type of computer that understands the concepts and relationships we use in everyday language, can recognize objects in images and connect it all to answer questions.
But most of the time we’re using it in a chatbot.
I remember my first computer, a 286 running MS-DOS. Confronted by the flashing cursor, I naively typed, “What’s the temperature in Antarctica?” The machine’s cold reply: “Bad command or file name.” To make it useful, I had to memorize the manual and teach myself BASIC. The magic quickly faded into a mild disappointment: this “smart” device wasn’t smart at all. It was just a vessel for the rigid logic and painstaking instructions I fed into it.
Windows and the graphical user interface made computers accessible to everyone, but for decades we’ve still had to learn the computer’s language: hunting for the right icon, memorizing gestures, typing on a keyboard layout designed for mechanical typewriters. AI should flip that. Why type when voice can be 3x faster, especially on a phone? Why click around Figma or PowerPoint when I could sketch what I want and have the computer build it? I have my answers to what apps would look like if they were designed for this new kind of computer, and I’ll tell you this…
It’s not a chatbot. In fact, Nick Turley (Head of Product for ChatGPT) admitted on Lenny’s Podcast that the interface was merely a “throwaway research demo” built in a 10-day sprint to test the model, an accidental interface that stuck.
What’s wrong with chatbots?
A chatbot is great for a single turn: you have a question or a task and want to hand it to something that behaves like an expert.
Here I delegated the research to the AI and kept control of the programming myself. Ethan Mollick calls this clear division of labor “Centaur” work. Creative work is different. Writing, designing a mechanical part or sketching an image means shaping the file directly, and the person brings context no prompt could hold: taste, spatial reasoning, and the financial, legal and cultural constraints. Mollick calls this back-and-forth “Cyborg” work. It needs an interface built for AI from the ground up, where you work on documents, slides, CAD drawings and video with your hands, voice or a stylus, hand bigger tasks to the AI, and then edit what comes back. That’s how I wrote this note: highlighting a section, asking a model to find a citation or tighten the wording, then reworking what it produced.
Coding tools got there first. Cursor gave engineers the cyborg version, with AI suggesting and editing inside the file they’re working in. Claude Code pushed toward delegation: I hand off a task, let the agent work through it for an hour, and review what comes back. Outside of code, we mostly don’t have this yet. Google Docs and Office have bolted AI into sidebars, and Anthropic’s Cowork has started bringing the Claude Code pattern to everyday knowledge work, but we’re a long way from tools that work this way across everything we use every day.
What’s actually working
In business, most of the AI-native companies at real scale sell tools for building things, not the things themselves: Cursor (past a $4 billion run rate before SpaceX bought it), Replit, Lovable, ElevenLabs. In Menlo Ventures’ enterprise data, coding alone was more than half of departmental AI spend, and Anthropic said Claude Code passed a $2.5 billion run rate in February. The vertical products with real traction, like Harvey in law, Sierra in customer service and Abridge in healthcare, are the ones that went deep. And even the leaders are small next to the labs: Harvey’s roughly $400 million run rate is well under 1% of Anthropic’s.
Consumer is even sparser. Outside ChatGPT, nothing is at scale. From what I can see, the new consumer AI apps (Instinct, Dots, Town, Muse) all have weak retention. Meta will pump Muse’s headline user numbers with distribution across its apps, but those numbers hide missing product market fit.
So the map is one chatbot for consumers, coding plus a handful of vertical products for business, and not much else.
Code worked first for a reason. A coding agent lives where the work happens: the context is in one place, the files are the source of truth, and the tests tell you whether it worked. Code got lucky. Most work doesn’t come packaged that way. I went deeper on this in my Anthropic piece.
Why So Little Has Been Built
If the capability is this good, why is the landscape so thin? I see four reasons, and only half of the first is about technology.
We don’t know how to build with it yet.
That’s really two problems: building software with AI, and building AI into software.
Building software with AI has never been cheaper. With Claude Code included in a $20-a-month plan (heavy users like me pay up to $200), the barrier has dropped from “We are an experienced Consumer Product team” to “I have an idea.” But more code isn’t more value. A friend manages a company where a couple hundred engineers wrote 11 million lines of code in a month. That’s something like a fifth of Windows, which took Microsoft decades. They’re building a great product, but its economic value is nowhere near Windows. The engine got ten times more powerful and the rest of the car didn’t. Research, design and product management haven’t adapted fast enough to find the new ideas, features and interfaces that would use the extra capacity. Without bigger tires, better suspension and aero, all that power just makes tire smoke.
Outside the big companies, the tire smoke is a flood of thin apps attacking every conceivable niche. As I worked on a product to help families organize their lives, I found at least five venture-backed startups and a large incumbent each solving slightly different versions of the same problem. And with AI products carrying real costs per user (unlike traditional software), they need either a clear value proposition people will pay for, or mass appeal sufficient to power an ad network. Most won’t have either.
Get Adriel Frederick’s stories in your inbox
Join Medium for free to get updates from this writer.
Building AI into software is the harder problem. Most product teams spent their careers building software that does the same thing every time. AI products don’t produce the same output twice, so you have to design for checking the work, not just doing it, and for the moments when the model should stop and ask. A generic model is like a brilliant consultant who knows nothing about you: useless until it has the files, the history and the objectives, and assembling that context reliably is a big part of the work. Then you have to know what the model can actually do, when that changes every few weeks. Something that failed in the spring works in the fall, and the workaround you built for it becomes dead weight. Nobody teaches any of this yet.
Fear of the Eye.
In my first version of this note, I called this the chatbot consumption paradox. As I iterated on product concepts, I worried that most of what I was building would get absorbed by the chatbots themselves. Most of my product’s work was an agent doing web research, peeking at my documents for context and writing structured output to files. That felt like a small extension of a chatbot, not a new interaction paradigm.
I wasn’t alone. A lot of builders read the signals as “don’t build anything the labs might touch,” and the signals are real. Perplexity had a real consumer story going after web search, until Google folded AI answers into search itself. Its traffic has fallen by roughly 40% from its late-2025 peak. It’s still growing revenue by moving toward agentic products, but the consumer search wedge is gone. The labs keep moving up into products too, from Cowork to apps inside ChatGPT. Evan Tana’s Avoiding the Eye of Sauron got passed around as exactly that warning.
The context is locked up.
The most valuable AI products need what I think of as the “God View” of our lives and work: emails, calendars, messages, documents, photos, health records, the sales pipeline. Most of it sits behind someone else’s wall. On the phone, Apple owns iMessage and the button that summons Siri, and Meta owns WhatsApp. At work, Epic owns clinical records and Salesforce owns the sales pipeline. Even the best AI app has to pay the landlords (Apple, Google, Meta) to reach users, and the owners of these systems won’t open them up freely. They’re too attached to their data moats or stuck in internal politics. My family organizer runs into this every day. Without access to iMessage or WhatsApp, it can’t see a lot of what a family actually says to each other, or act where they talk.
Investors briefly bet the other way. When Anthropic launched Cowork in January, the fear was that agents would make the big software suites irrelevant, and roughly a trillion dollars came off software stocks in a few weeks. By April the iShares software ETF was down 30% for the year. Then it came back. By mid-October it was up for the year again, about 7% off its 2025 peak. A few friends and I called that bounce before it happened, for one reason: agents need context, and the context lives in the systems of record. An agent that’s going to close your books needs the ledger. If anything, AI makes those systems more valuable. Open standards like MCP (Model Context Protocol, which lets AI assistants connect to your apps and data) chip away at the walls, but they don’t remove them.
Most people don’t know how to use it yet.
The first three reasons are about builders. The fourth is about everyone else. Even when the product exists, most people and companies don’t know how to adopt it. Silicon Valley, the idea more than the place, is a small network where information moves fast. People change jobs, talk shop and copy each other’s setups within weeks, so new tools and techniques spread faster there than anywhere else in the world. Everywhere else, they spread the slow way.
I’m advising a tech company in Trinidad, and I told them to stop whatever they were doing and figure out Claude Code. They didn’t know how to start. Not just how to write a clear prompt, but how to plan a roadmap around it, how to run agentic code review, how to change the way the team works, and even how to handle the billing, payments and provisioning to get going. None of that is hard once you’ve seen it done. They just hadn’t seen it done.
Even I’m the patching layer. I run Claude Code on a devbox I set up so my laptop doesn’t have to stay on, with GitHub for source control, paid CI actions that trigger agents, and hosting on Railway. Most heavy users I know have stitched together some version of this by hand.
We’ve been here before. Everyone has a story of the friend who understood PCs: the one who helped you pick one out, set it up, install the software and showed you how to use it. That’s how personal computing actually spread. Apple turned that informal network into a channel. The Genius Bar, the in-store workshops and AppleCare put the friend who knows computers inside the store. TurboTax did something similar in software. Its interview format put the tax knowledge in the product, so you didn’t need to know the tax code to file.
AI adoption needs both versions: products with opinionated defaults that carry the know-how (a setup wizard where you answer a few questions, pay once and come back when it’s ready), and distribution that teaches people how to use the tools, not just sells access to them. Until then, adoption outside the echo chamber moves at the speed of people.
Build Anyway, and Go Deep
I think I was wrong about the chatbots eating everything, or at least I overestimated how fast and how deep the big platforms can go. Read closely, Tana’s essay argues something close to the opposite of how it got passed around: avoid thin, horizontal tools sold to fast-moving buyers, and go deep in slow, hard industries. His line: “AI is the ingredient. Domain mastery is the product.” Read thinly, it scared people away. Read properly, it’s a playbook.
Why depth is harder to copy than it looks
The more time I spend with the products that are working, the less I believe the incumbents will just eat them. Harvey and Legora have gone deep on law and found that every firm’s workflows are bespoke. Their go-to-market looks more like Palantir than Slack. Harvey puts a former practicing lawyer in every deployment, about 180 of them, and some customers get a forward-deployed team. It’s also building its own legal models for things like citations and research. Legora goes deep on its own legal research data and workflows. That operating model isn’t something Anthropic, OpenAI or Google Cloud spins up in a quarter.
I’m seeing the same thing building a product for busy families. It’s not a foregone conclusion that Google, Meta or OpenAI could copy it from the jump. So much of the work is nuance you can’t see until you go deep on the problem. A birthday party invitation in email doesn’t mean we’re going if it isn’t on the calendar. The time my wife put on her calendar for an event isn’t the fact; the organizer’s email is the real source. And while AI makes it easier to build copies of products with highly visible features like Photoshop, copying a product with deeply embedded and highly situational logic isn’t the same.
The chatbots still have real limits. They don’t always have the context to build great experiences, and they can’t hold the persistent “state” users need to view, edit and share across products. If ChatGPT has your medical history, you need to be able to see it, correct it and share it with your doctor, not just reach it through a conversation. So my old advice, grow and complement the chatbots rather than trying to replace them, still holds for thin products. But the bigger opportunity is depth: products that take responsibility for a real piece of work in a domain where the context is messy, the trust is earned and the workflow is bespoke. I wrote more about what that takes in the “What depth means” section of my Anthropic piece.
What going deep looks like: tuning my Evo
The clearest example I have isn’t a startup. It’s my car.
My Evo has a heavily modified engine, a one-off turbocharger and a programmable engine computer, and it runs on ethanol. I do all the tuning myself, including measuring and adjusting power on a dyno, and this year Claude became my tuning partner. What made it work wasn’t a smarter chatbot. It was building a closed loop around the car: the Observe, Orient, Decide, Act loop that John Boyd came up with for fighter pilots.
- Observe. I drive and the engine computer logs everything. An 11-minute drive came back as 132,000 rows, with channels logged at different rates. Claude built the tools to parse and chart them.
- Orient. Raw data is useless without knowing how this engine should behave. So we gave Claude the domain: about 800 articles from Haltech, the engine computer’s maker, organized the way a tuner thinks; Merlin’s Evo tuning guide, scanned so Claude could read it; and notes from every past session, so it never started cold.
- Decide. Each change became a small experiment. For ignition timing, I ran A/B/A sets of full-throttle runs (baseline, change, back to baseline) so we could confirm conditions hadn’t drifted before trusting a 1% difference.
- Act. The last gap was getting changes back into the car. Haltech’s calibration files are encrypted, so Claude reverse-engineered the format until it could read and write the tables directly, in real units. I’d load the new map, drive, log, and the loop started again.
Over twelve days in June, that loop took the car’s idle and low-RPM behavior through 16 map versions and worked through every problem on my list: an idle hunt when the A/C kicked on, stumbles while warming up, bucking on the coast, idle flares when I turned the wheel. Several fixes that seemed obvious turned out to be wrong and got reverted when the logs said so. An earlier session found the knock sensor was listening in the wrong part of the combustion cycle. It was mistaking normal combustion noise for knock and pulling more than two degrees of timing for no reason.
That’s the depth questions answered for one narrow job. What can I hand over? Diagnosing and fixing a drivability problem. What does it need to know? The logs, framed by the domain knowledge. What can it do? Write the map. How do I know it’s right? The next set of logs. Code got its loop for free. For the Evo, we had to build it, and that’s what going deep means for most work.
Get out of the echo chamber
Here’s what bugs me about a lot of the AI products I see pitched: restaurant bookings, travel planners, generic shopping assistants. These are problems that matter mostly to people who live in Silicon Valley, and they’re already pretty well solved. The bigger opportunity is real problems for real people, in places where the work is messy and the people doing it aren’t reading AI newsletters. A few I think are possible today:
- Project planning and accounting for small businesses. This is the world I came from at Intuit and Outright. A contractor juggling five jobs doesn’t need a chatbot. They need something that knows what each job was quoted at, what’s been spent, which invoices are late and who to chase this week, pulled from the bank feed, the receipts and the texts with clients.
- Custom parts from a prompt. Describe a bracket, or sketch it, and get a part you can 3D print or machine. Startups like Zoo and Adam are early here. The hard part isn’t the geometry; it’s the engineering intent: where the load goes, what it mounts to, which tolerances matter.
- Car diagnosis and repair. The Evo loop, for everyone else. Most cars already report what’s wrong through the diagnostic port. Pair that with the service manuals and the forum knowledge of people who’ve fixed the same problem, and an independent shop, or a handy owner, gets a diagnostic partner. A few teams are starting on this, including CarSignal in Y Combinator’s latest batch.
- Estimating and quoting for the trades. HVAC techs, plumbers and roofers lose evenings to quotes. Photos, measurements and past jobs could turn that into minutes.
- Property management for small landlords. Leases, maintenance requests, vendors, rent and the paperwork around all of it, for people who own a few units and run them on the side.
- Know-how layers. The setup, training and opinionated defaults that get a team from “we bought it” to “it’s doing real work,” built for a specific market or industry. This is the Genius Bar for AI, and it doesn’t have to come from the labs.
- Fighting medical bills. Families get bills they can’t read for care they can’t match to their insurance. Checking the bill against the explanation of benefits, catching the errors and drafting the appeal is exactly the kind of tedious, high-stakes work AI should take on.
Each of these needs the same things: context pulled from messy sources, domain knowledge to frame it, the ability to act, and a way to check the result. None of them is a better chatbot.
Who Wins When Software Is Almost Free
What’s not clear is who builds all this, and who captures the value. I see three defensible positions:
- Developing or controlling foundational models. You own a scarce resource you can rent to others. More importantly, you know what capabilities are coming before the rest of the world, giving you first-mover advantage on products that leverage those capabilities. The labs are also moving up into products, which is why thin, horizontal tools sit right in the Eye’s path.
- Owning the systems of record and the distribution. This includes critical datasets, embedded workflows, distribution channels, regulatory moats, or deployed hardware. Salesforce owns the sales workflow. Epic owns clinical data. Apple owns the device. Google owns search history and email. Square owns the payment terminal at the point of sale. The software bounce this year is the proof: AI doesn’t erode this power; it amplifies it.
- Depth that owns an outcome. Domain mastery, earned trust and a loop that checks the work, in a domain where the context is messy and the workflow is bespoke. This is the position open to new players, and I no longer think the window is narrow. It’s slow, unglamorous work, which is exactly why the platforms won’t do it in a quarter. The best of these will build their own expert data and eventually their own models, the way Harvey is.
The chassis is still coming
If logic machines were “bicycles for the mind”, AI powered computers are supercars. We have the engine (the models) but not the chassis (the interfaces to put it in). When I think about using AI to give myself superpowers, two interfaces come to mind.
- Smart glasses for memory and passive assistance. These have gone from idea to product faster than I expected. EssilorLuxottica sold more than seven million Ray-Ban and Oakley smart glasses in 2025, and Google and Samsung’s are on the way. The hardware is arriving; the software that makes them indispensable isn’t here yet. Glasses with a camera and microphone could take quick notes and capture what I’m seeing and want to remember, and if organized well, give me incredible recall. “What was the name of the company Saad told me about yesterday?” Or “Hey, remind me to get some chicken before I get home.” For kids in school, they could give personalized help with math homework while they learn the concepts. We still have to sort out the social dynamics and privacy. I err on the side of recording being off until explicitly turned on, with a light telling others it’s active. Passive recording could too easily lead to dystopian examples like the Black Mirror episode The Entire History of You.
- Tablet with stylus for co-creation. Sketch an idea for a mechanical part or a scene in a movie, and an AI partner refines the lines, suggests materials or generates a 3D render as you draw. You are the director; the AI is the studio. While I hate the idea of a stylus because it’s too easy to lose, the precision it allows for drawing, selecting and editing unlocks so much power. Gestures like those in the Apple Vision Pro may let you work in 3D space with your hands, but getting the precision required for direct manipulation of text, lines and sketches feels like it will be too hard.
Whoever builds native AI hardware like this and gets it to scale creates a new platform, with the distribution advantage Apple, Google and Meta hold today. The teams most likely to succeed on those platforms are the ones building AI products on today’s. They’ll have learned to build software that doesn’t produce the same output twice and to ship faster with AI. Those capabilities won’t transfer easily to teams still operating in the old paradigm. One more reason to build now.
We are moving toward a world where computers aren’t just tools we use, but extensions of our own minds. The winners will build things like the personalized tutor in The Diamond Age, not to replace specialists, but to make everyone more capable. The transition won’t be seamless, but for those willing to learn this new language of collaboration (to work with a genius on call rather than just use a tool), the opportunity is to be 10x more capable than we ever thought possible.
What’s changed
November 2026
- The industrial chain is more industrial than I thought. Data centers now look like refineries, the buildout is bigger relative to GDP than the railroads or the interstates, and data has become an industrial input of its own, from scanned books to expert-built training data to the internal records of real companies.
- I was wrong about the chatbots eating everything. I overestimated how fast the big platforms could go deep. Most AI companies at scale sell tools for builders, consumer is nearly empty outside ChatGPT, and the products that are working went deep on a domain in ways that are hard to copy.
- We’re still early on products. I rebuilt the back half around one idea: the capability is here, but the products mostly aren’t. It now covers what’s actually working, why so little has been built, and where I’d build, with tuning my Evo as an example of going deep. I also added a section on overbuilding.
- Adoption is its own bottleneck. I added a fourth reason so little has been built: outside Silicon Valley, most people and companies don’t know how to adopt these tools yet, even when the products exist.
- Also refreshed out-of-date facts on chips, power and pricing.
January 2026
- A shorter primer. I cut the deep technical walk-through of how models are trained down to the essentials.
- Bring your own power. I updated the energy picture as AI labs started building their own gas plants instead of waiting years for the grid.
- Context is what makes AI tools work. I added Cursor and Claude Code as early evidence that the winning tools succeed by living inside the environment where work happens.
- Who wins when software is almost free. I replaced my company-by-company look at Google, Meta and Apple with a broader view: a flood of new apps, the risk of chatbots absorbing them, and three sources of durable power.
October 2025: First version, shared as a Google Doc. The doc now shows the January text; the original is in its version history.
