Read, Write, Own, Think
Four webs, one race that will not stop, and why this time the opportunity belongs to everyone who learns to ride it
Abstract. This is the long version of a lecture I give in the first hour of every Chasing Jarvis programme, told as a story. It follows the web through four waves, from the static pages of the 1990s to the agents of today, using Gavin Wood’s Web 3.0 thesis as the spine and adding a fourth verb of my own: think. It then asks a question the usual history skips: in each wave, who actually got to build? My answer is that AI is the first wave whose biggest opportunity is not reserved for developers. The Centaur model, a domain expert working with AI, can turn ordinary professionals into something close to superhuman in their own field. And the race between the United States and China guarantees that this wave will not pass us by, whether we choose to ride it or not.
1. Thirty Years at the Edge of Every Wave
I have spent my whole working life in technology. I was building my own computers at sixteen. I opened my first company, a recording studio, in 1997. In the early 2000s I taught myself HTML and CSS because I wanted websites for my own businesses, and I rode the search-engine wave for the best part of a decade: whatever I optimised landed on the first page of Google, and that meant I could sell more or less anything. Then WordPress, then e-commerce, then, in 2016, blockchain, which overwhelmed me with its promise of a permissionless internet. Since early 2024 I have been building with AI coding agents almost every day.
Thanks for reading! Subscribe for free to receive new posts and support my work.
So I have seen a few waves. Every one of them brought real opportunity. But I want to be honest about something that rarely gets said: in every one of those waves, the biggest opportunity went to a small group of people. In the 1990s it went to developers. If you could not write code or pay someone who could, the web was something you read, not something you built. Even in my own SEO years, my advantage came from the fact that I had learned enough of the technical craft to do it myself.
What is happening now is different, and I do not say that lightly. The advantages and the opportunities that come with AI are, in my experience, unprecedented. Not because the technology is clever, although it is, but because for the first time you do not need to be a developer to build with it. You need to know your domain, and you need to learn how to direct the machine. If you do, the possibilities are close to limitless.
This article is my attempt to explain why. If you are one of my students, it is the long version of two slides from our first session. If you are not, it is the story I wish someone had told me at the start of every wave I lived through.
Every technology wave I lived through rewarded the people who could build. This is the first one where almost anyone can learn to build.
2. Four Webs in One Lifetime
The cleanest way I know to describe how the internet changed is a thesis that comes from the blockchain world. In 2014, Dr. Gavin Wood, co-founder of Ethereum and later founder of Polkadot, published a short essay describing what he called Web 3.0: a web where trust no longer depends on the platforms in the middle. The shorthand that grew around his idea, later popularised by Chris Dixon in his book Read Write Own, sums up the first three waves in three verbs. I have added a fourth.
Web 1.0 (1991 to 2004): Read
Tim Berners-Lee put the first website online in 1991. What followed was a web of static pages, hyperlinks and, eventually, search engines. Information flowed one way, from the publisher to the reader. Most people could only read. The winners of this era owned traffic: the portals, the early search engines, the domain names and, a little later, people like me who learned how to climb Google’s rankings before everyone else did.
Web 2.0 (2004 to 2014): Read + Write
Tim O’Reilly made the term famous in 2004, and the idea was simple: the web became a place where everyone could write. Blogs, YouTube, Facebook, Twitter, Wikipedia. User-generated content exploded and everyone became a creator. The winners owned audience: the platforms and the creators who built communities on them. But notice the catch. You could write, but you did not own your reach. The platform could change its algorithm on a Tuesday and your audience was gone by Wednesday.
Web 3.0 (2014 to 2022): Read + Write + Own
Gavin Wood’s answer to that catch was ownership. With blockchain, cryptocurrencies and smart contracts, users could own digital assets directly, without asking a platform for permission. I lived this wave from 2016 onward, building on-chain communication and decentralised storage with my partners at 4thTech and PollinationX. The winners owned assets: early crypto holders, DeFi builders, stablecoin infrastructure. And the barrier was high again. To benefit, you needed a wallet, crypto know-how and a strong stomach. The gains went largely to early, technical adopters.
Web 4.0 (2022 onward): Read + Write + Own + Think
This is my addition to Gavin’s thesis. When ChatGPT arrived in November 2022, the web gained a new verb: think. Within two years we had models that write code, open standards like the Model Context Protocol that let them use tools, and agents that plan and act on our behalf. The web is no longer only a place where we read, write and own. It is a place that reasons, drafts, analyses and builds alongside us. The winners of this era will own intelligence: the ability to direct AI well, the knowledge it works from, and the speed that comes with both.
I am aware that “Web 4.0” is used in other ways too; the European Commission, for example, uses it for immersive and virtual worlds. I use it in the sense above, because it is the clearest way I have found to show where AI sits in the story of the web.
3. Who Got to Build?
Here is the part the usual history skips. Instead of asking what each wave could do, ask what you needed in order to build with it.
In Web 1.0, you needed code and a server. In Web 2.0, you needed a platform account, and the platform kept the keys. In Web 3.0, you needed a wallet and a crypto education. In every case, a technical person had a structural advantage over everyone else.
In Web 4.0, what you need is your expertise and plain English. That is not a slogan. In my courses at COTRUGLI, people who have never written a line of code ship a working product to a live web address in five days, by instructing coding agents in plain language. In my own work, I build software I could never have built alone two years ago. The bottleneck has moved from writing code to knowing what to build, and describing it precisely.
The speed of adoption tells the same story. According to Stanford’s 2026 AI Index, generative AI reached 53 percent adoption within three years, faster than either the personal computer or the internet at the same stage. Technologies do not spread that fast when they are only useful to engineers.
In the 1990s the web’s opportunity belonged to developers. Today AI’s opportunity belongs to anyone with real expertise who learns to direct it.
4. The Centaur
To explain what that means for an individual, I go back to chess.
In 1997, IBM’s Deep Blue beat Garry Kasparov, the world champion. Many people read it as the end of human chess. Kasparov drew a different conclusion. A year later he helped launch “advanced chess”, where human players were allowed to use computers during the game. The human-machine teams came to be called centaurs.
The real surprise came in 2005, at an open “freestyle” tournament where anyone could enter with any combination of people and machines. Grandmasters entered with computers. Even Hydra, a chess supercomputer, entered and failed to reach the quarter-finals. The winners were two amateur players from the United States, Steven Cramton and Zackary Stephen, using three ordinary home computers. They were not the best chess players in the room, and they did not have the strongest machine. They had the best process for working with their machines. Kasparov later summed it up: a weak human with a machine and a better process was superior to a strong computer alone, and, more remarkably, superior to a strong human with a machine and an inferior process.
That is the Centaur model I teach, applied to work. The human provides direction, judgement, context and taste. The AI provides speed, breadth and execution. Neither is as strong alone. And the advantage does not go to the person with the most powerful model, because everyone has access to roughly the same models. It goes to the person who knows their domain and has learned how to work with the machine.
There is now hard evidence that this is not only true in chess. In 2025, researchers from Harvard Business School ran a field experiment with Procter & Gamble involving 776 professionals working on real product-development challenges. Individuals working with AI matched the performance of two-person teams working without it, and AI-assisted teams produced more of the very best solutions. One person with AI did the work of a team.
This is why I use the word superhuman, and I mean it in a precise way. A good lawyer with AI is not replaced by AI. She reviews in an hour what used to take a week, and she spends the saved time on judgement. A good marketer with AI does not stop being a marketer. He runs a content operation that used to need an agency. A founder with AI does not stop being a founder. She tests ten ideas in the time it used to take to estimate one. The Centaur is not a new kind of worker. It is the expert you already are, amplified.
The AI is the same for everyone. Your domain expertise and the way you work with it are not. That difference is the whole advantage.
There is one honest condition. It does not work out of the box. You have to invest time and learn the craft: how these models behave, what context they need, how to break a problem down for them, how to check what they produce. I call this context engineering, and it is the skill at the centre of everything I teach. It is learnable, faster than most people fear. But it is learning, not magic.
5. What the Centaur Can Do Today
Talking about potential is cheap, so let me show you what this looks like in practice. Most of these examples come from my own work in the last eighteen months.
The end of the long road to an MVP
For most of my career, the most painful part of starting something new was not the money. It was the waiting. You had an idea. Then you hired, or you found a technical partner, and you spent six months to a year and a payroll before you saw a first version, which was often not what you imagined. Your runway burned while you waited. Many good ideas died right there.
That bottleneck is gone. This spring, the founder of a company with more than five hundred employees wanted to test a new product idea. His team’s estimate was six senior developers and eight to twelve months to reach a working MVP. I built it alone in eight weeks, by directing more than four hundred AI coding agents. The result was ARIA, a real-time, voice-driven video avatar that runs inside a product, reacts to live events and remembers returning users. About seventeen thousand lines of code. The total bill for AI usage was roughly twelve to thirteen thousand dollars. I described the whole build, including what went wrong, in One Founder, Eight Weeks, Four Hundred Agents.
ARIA is a big example. The smaller ones matter just as much. Last December, rain ruined a ski holiday in the Dolomites, and in the following thirty days I built Lumina, a platform that turns a company’s documents into a support agent for its website. It is live today. I built the MountVacation MCP server, one of the first open-source tools that lets AI assistants search travel accommodation, in fourteen days. For a simple idea, a working prototype over a weekend is now realistic. If you are a non-technical founder, that changes everything: you can test the idea before you raise money for it, not after.
Marketing, content and video
Video used to be one of the most expensive things a small company could produce: a script, a crew, actors, locations, editing, weeks of work and a budget to match. Today, with tools like Google’s Veo for video, Nano Banana Pro for images and ElevenLabs for voice, a product video can be produced in a few hours by one person with a clear brief. Here is an example: youtu.be/1QUIppY80x8.
The same is true for the rest of marketing. I documented how I built an AI marketing team that produces everything from technical articles to short meme-style videos for my products. The key, again, is context: the agents only produce good work when they are given the product knowledge, the brand voice and the audience. With that in place, one person can run what used to need an agency.
Customer support and sales
Every company has knowledge that customers ask about again and again: prices, specifications, delivery times, policies. A support agent grounded in your own documents can answer those questions on your website or WhatsApp around the clock, say so honestly when it does not know, and hand the conversation to a human when it matters. That is exactly what Lumina does. Internally, I run an email agent that reads five inboxes, sorts and summarises incoming mail and drafts replies, with one hard rule: it never sends anything without my approval.
Legal, compliance and expert knowledge
For more than a year, my team tried to build an assistant that could answer questions about Slovenia’s notoriously complex building legislation, and we could not get past roughly eighty percent accuracy. Once models could read whole documents at once, that changed, and Moj AI now answers from the full legal text rather than from fragments. More broadly, AI can now read a contract, compare it against your standard terms and flag the risky clauses in minutes. It does not replace the lawyer. It gives the lawyer, and the business owner who could never afford one, a first reader that never gets tired. Hallucinations are real, which is why the human stays in the loop.
Finance, HR, procurement and every other department
The use cases that may matter most are the least glamorous. A finance analyst asks questions of a messy Excel model in plain language. An HR team drafts job adverts, onboarding guides and policy answers in a fraction of the time. A procurement manager hands an agent three supplier quotes and gets back a comparison table and a recommendation memo, and approves it before anything is sent. Engineers summarise research papers and review specifications. Designers produce mock-ups and presentations in an afternoon.
Your knowledge, compounding
Finally, AI changes what we can do with what we already know. I built The Curator, an open-source tool that turns my articles, notes and documents into a linked knowledge base that my AI agents can read and write. This article was researched from it. For a company, the same idea becomes a shared brain: the expertise of the organisation, captured and available to every employee and every agent.
The question is no longer whether AI can help in your department. It is whether someone in your department has learned to make it help.
6. The Race That Will Not Stop
Now for the reason none of this is optional.
Get Dr. Tali Rezun’s stories in your inbox
Join Medium for free to get updates from this writer.
AI is not just another technology product. It has become the top strategic priority of the two superpowers, and the competition between them is the engine that keeps the whole thing accelerating. The United States published its national AI strategy in July 2025 under a title that leaves no room for doubt: Winning the Race. China followed a month later with its “AI+” plan to integrate artificial intelligence across its entire economy over the next decade. The leading labs speak openly about building superintelligence. Neither side is going to slow down because the rest of the world finds the pace uncomfortable.
The two have chosen different strategies. The United States leads at the frontier. Private investment in AI there reached about 286 billion dollars in 2025, according to the Stanford AI Index. The strongest models, from OpenAI, Anthropic, Google and xAI, are closed: you use them through the cloud and you rent their intelligence.
China has gone the other way, with open-weight models. DeepSeek, Qwen, Kimi and GLM can be downloaded and run on your own servers. Over the past year, Chinese models accounted for 41 percent of all model downloads on Hugging Face, overtaking American ones. And by March 2026, the best Chinese model was within 2.7 percent of the best American model on Stanford’s measure, on a fraction of the private investment, with state funding on top.
And Europe? Around 21 billion dollars of private AI investment in 2025, roughly fourteen times less than the United States. One serious frontier challenger, Mistral. We lead the world in rules, with the AI Act, and we have announced a plan to catch up, the InvestAI initiative, which aims to mobilise 200 billion euros. But today, mostly, we use other people’s models.
I felt what that means personally this summer. In June, when the United States imposed export controls on its newest frontier models, access was suspended, and for almost three weeks I, like everyone in Europe, could not use the model I had been building with. Access came back in July. But the lesson stayed: if the intelligence you depend on lives entirely on someone else’s servers, under someone else’s laws, it can be switched off.
We cannot choose whether this race happens. We can only choose whether we learn to run with it.
7. What This Means for Us in Europe
For a European company, the race is not abstract geopolitics. It is a very practical question: which intelligence do we use, for what, and where does our data go?
The frontier models are the strongest, and for most everyday work they are the right tool, ideally on an enterprise plan with clear data protection terms. But your data leaves the building, and you are relying on a promise you cannot audit. Open-weight models, many of them Chinese, can run entirely on your own hardware, offline if necessary. They are a step behind the frontier, but the gap is small and closing, and in my own production work an open-weight model produced comparable interface code at a sixteenth of the cost.
My advice to every company I work with is the same. Do not ask “cloud or local”. Classify your data first. Public material can go to any cloud tool. Internal processes and documents can go to an enterprise plan with a data processing agreement. Your crown jewels, the designs, the formulas, the intellectual property that makes your company what it is, should stay on infrastructure you control. Then match the intelligence to the data.
And here is the good news for Europe. We may not build most of the frontier models. But the greatest value of this wave is not in building models. It is in applying them, and that is where Europe’s depth of expertise in engineering, manufacturing, law, design, medicine and finance becomes an advantage. A continent of Centaurs does not need to win the model race to win the application race.
8. The Choice in Front of Each of Us
So we arrive at the choice, and it is a personal one.
We can learn to work with AI. If you are an employee, this is how you become more valuable to your employer, not less: the expert who can do the work of a team is the last person a company wants to lose. If you are an entrepreneur, it is your edge in a hard market: you can test more ideas, faster and cheaper, than competitors who are still waiting for their development budget.
Or we can let it pass. We can wait, watch and hope it turns out to be hype. I understand the temptation. But I do not think it is a safe choice. The race guarantees that this technology will keep improving and keep spreading, whether we pay attention or not. Some tasks will be automated. Some roles will change beyond recognition. Some companies will be overtaken by smaller competitors who learned faster. That is not a prediction about machines replacing people. It is a prediction about people who use AI replacing people who do not.
I wrote in the first article of this series that the future belongs to those who learn to collaborate with artificial intelligence, not those who fear or ignore it. Two years and many builds later, I believe it more than ever. The difference now is that the window in which being early is a real advantage is open, and it will not stay open forever.
9. Where to Start
If this article has done its job, you are wondering what to do on Monday. Here is the short version.
• Pick one real task from your own work. Not a demo. Something you do every week that takes too long. Start there.
• Learn the basics of how these models work. What a context window is, why models make things up, what they need from you. One good afternoon is enough to begin.
• Treat context as the skill. The quality of what you get back depends on what you give: your documents, your examples, your constraints. Write them down.
• Choose one tool and go deep. Claude, ChatGPT, Gemini or an open-source alternative. Depth beats collecting tools.
• Classify your data before you upload it. Public, internal, crown jewels. Then decide which intelligence may see what.
• Build something end to end. A website, an agent, a small internal tool. Nothing teaches like shipping.
That is also, more or less, the structure of the Chasing Jarvis programme I teach at COTRUGLI Business School, where non-technical professionals build and ship a working product in five days. You do not need my course to start. But you do need to start.
Read. Write. Own. Think. The fourth verb is the first one that every one of us can learn to use.
About the Author
Dr. Tali Režun is a serial entrepreneur, business developer and academic at the forefront of frontier technologies. As Vice Dean for Frontier Technologies at COTRUGLI Business School, he leads AI innovation initiatives and designed the Chasing Jarvis programme, which teaches non-technical professionals to build with AI agents. With more than thirty years of entrepreneurial experience, founding and scaling ventures including AgentLink, The Curator, Lumina AI, Moj AI, Block Labs, 4thTech, Immu3, PollinationX and Online Guerrilla, he bridges research in AI and Web3 with practical business transformation.
Links: talirezun.com · talirezun.substack.com · x.com/talirezun · linkedin.com/in/talirezun · github.com/talirezun · fieldnotes.talirezun.com
Sources & Further Reading
External sources
• Stanford HAI (2026). The 2026 AI Index Report. hai.stanford.edu/ai-index/2026-ai-index-report (private AI investment 2025, US–China performance gap, adoption speed)
• Hugging Face (2026). Spring 2026 report on the open-source AI ecosystem, as reported by EqualOcean: Chinese Models Account for 41% of Hugging Face Downloads. equalocean.com
• Wood, G. (2014). ĐApps: What Web 3.0 Looks Like. See the history in Web3, Wikipedia and CoinDesk (2022)
• Dixon, C. (2024). Read Write Own: Building the Next Era of the Internet. Random House.
• Kasparov, G. (2010). The Chess Master and the Computer. The New York Review of Books.
• ChessBase (2005). Dark horse ZackS wins Freestyle Chess Tournament. en.chessbase.com
• Dell’Acqua, F. et al. (2025). The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and Expertise. Harvard Business School Working Paper.
• The White House (2025). Winning the Race: America’s AI Action Plan. Summary: JD Supra
• State Council of China (2025). “AI+” Action Plan. Reported by Caixin Global
• European Commission (2025). InvestAI: mobilising €200 billion for AI. digital-skills-jobs.europa.eu
From Lab to Life, my own field notes referenced in this article
• One Founder, Eight Weeks, Four Hundred Agents (2026). talirezun.substack.com
• A Year in the Review: What One Non-Developer Shipped in Twelve Months (2026). talirezun.substack.com
• Are We Becoming Obsolete or Finally Free? (2025). talirezun.substack.com
• Context is the Code: The Complete Three-Phase Process for Building with AI Agents (2026). talirezun.substack.com
• 16x Cheaper, Open Weights, and the Model That Doesn’t Disappear on Fridays (2026). talirezun.substack.com
• I Built a 24/7 AI Email Support Agent (2026). talirezun.substack.com
• How I Built an AI Marketing Team That Actually Works (Medium). medium.com/@talirezun
• Why I Ditched RAG Pipelines for 1M Token Context Windows (Medium, 2026). medium.com/@talirezun
• Data Sovereignty in the AI Age: Building Your Own Private ChatGPT (Medium, 2025). medium.com/@talirezun
• Field Notes: the distilled course reading. fieldnotes.talirezun.com
Disclaimer
Educational purpose. This article is published for educational purposes. It combines personal experience, observation and analysis from building with AI agents, with published research and public data. It is not investment, legal or career advice.
No commercial relationships. The author has not been compensated or sponsored by any of the companies, platforms or tools mentioned in this article. Lumina, The Curator, Moj AI, ARIA and MountVacation MCP are the author’s own projects or client work.
Figures. Investment, adoption and download figures come from the sources listed above and reflect what was published at the time of writing. Build timelines, costs and agent counts are drawn from the author’s own project records. The AI landscape changes quickly; verify current figures with the primary sources before citing them.
![[Day 38/100] Knowledge Graph Memory for AI Agents: Neo4j, Memgraph, and Zep](https://webproxy.poorya-velaei-d67.workers.dev/https://miro.medium.com/v2/resize:fit:1358/format:webp/1*_ocO7nowx69kMxnL8BzizA.png)
