The Year the Critics Started Building

I stared at Simon Willison's blog for a solid minute.

Not because of what he wrote. Because of what he built.

The guy who taught half the internet how to prompt engineer. The guy whose AI tool reviews were required reading across the entire industry. The guy who kept a careful, almost clinical distance from the thing he was analyzing.

He just shipped an interactive Redis playground. In a browser. WASM-compiled Redis, built with Claude Code for web. You can type in brand-new Redis array commands (ARGREP, ARSET, ARSCAN) right in the browser. Server-side regex against array values, compiled to WebAssembly, running client-side. He built it the same day Salvatore Sanfilippo (antirez) published his post about the new Redis array type.

This isn't a review. This isn't analysis. This is a build.

And it's not a one-off. In the past four days, Simon shipped three separate builds from his phone while camping. An iNaturalist sightings integration for his blog. A Python CLI called inaturalist-clumper with a git-scraping pipeline. That Redis playground. Claude Code for web, every time. On a phone. From a tent.

In February, he built his entire "beats" content syndication system in a single morning. It features five different data integrations pulling from GitHub releases, TIL posts, museum blogs, vibe-coded tools, and AI research repos.

The web's most respected AI critic just became one of its most prolific AI builders. And he is not alone.

Something shifted in 2026. The people who used to write about AI are now building with it. The people who used to analyze are going silent. The commentary class is becoming the builder class, and that tells you more about where this technology actually is than any benchmark or funding round.

PART ONE: THE BUILDERS

Simon Willison: The Canary

Simon Willison co-created Django. He built Datasette. For three years, he was the web's most thorough AI tool reviewer: technically rigorous, skeptical without being dismissive, and always showing his work. His blog was the definitive source on prompt engineering, tool comparisons, and LLM capabilities.

The shift happened gradually, then all at once.

Late 2025: he started "vibe-coding" HTML tools. Terms like "vibe-coded" entered his vocabulary. He built research repos with async coding agents.

February 2026: the "beats" content syndication system launched. Five integrations, all working, built across a single morning. He prototyped in Claude Artifacts and built in Claude Code for web.

May 1 to May 4, 2026: three builds in four days. All on his phone. All with Claude Code for web. A shipping velocity that would have been absurd two years ago.

But here is what matters most. He is documenting his building process publicly. Every build links back to the Claude Code PRs. His prompts are visible. The critic's instinct to explain never went away; it just found a new medium. He is teaching by building.

If even Simon Willison is spending his weekends shipping AI-built tools instead of writing about them, the technology has crossed an adoption threshold that no benchmark measures. He was the canary. The rest of us just haven't caught up yet.

Pieter Levels: The Blueprint

Pieter Levels is the proof of concept that every solo developer cites. Digital nomad, solo founder. Built Nomad List in 2014. Spent a decade as indie hacking royalty. One person, multiple profitable products, no VC backing, no employees.

His current portfolio is absurd. Nomads.com. Photo AI (29.3 million photos generated, $100K/month). RemoteOK ($43K/month). Interior AI ($34K/month). Hoodmaps. Hotelist. Airline List. fly.pieter.com. JSON.pub. eu/acc. Eight-plus active products generating roughly $256K to $259K a month in revenue. All from one person.

Levels stopped blogging in September 2022. His last post was literally titled "I have stopped blogging on here but continue blogging almost every day on my X account." He moved entirely to short-form. He doesn't write essays about AI. He builds products that could not exist before 2023.

Photo AI and Interior AI are pure AI plays. He took a decade of indie hacking pattern recognition and applied it to the generative AI wave. It worked spectacularly.

He is not an AI researcher. He is not a FAANG refugee. He is an indie hacker who saw the opportunity and shipped. Every solo developer wondering if they can actually build something people will pay for looks at Levels for the answer.

Marc Lou: The New Breed

"I was fired everywhere, so I hired myself."

That is Marc Lou's bio. He built his way from zero to 37 shipped products. $64K a month in revenue. 42,851 newsletter subscribers. Sold MakeLanding for $35K.

Lou represents something that genuinely did not exist five years ago. He is not a programmer who learned business. He is not a business person who hired programmers. He is someone who realized AI tools let him do both at once, and he did it 37 times.

His model is brutally simple. Ship fast. Ship often. Monetize directly. No VC money. No enterprise sales cycles. No waiting for permission. Claude, Cursor, and Codex collapse the gap between idea and deployed product from months to days.

There is a generation of Marc Lous coming. They do not write think pieces. They ship.

PART TWO: THE SILENCE

The story is not just who started building. It is who stopped talking.

Lilian Weng: The Intellectual Who Went Dark

Lilian Weng is VP of Safety Systems at OpenAI. She wrote some of the most influential technical blog posts in AI history. Her "LLM Powered Autonomous Agents" post from June 2023 basically defined how the industry thinks about agent architecture. Her pieces on reward hacking, hallucination, and adversarial attacks were cited by thousands.

Her timeline tells the story. 2023 through 2024 featured deep, rigorous, technically dense posts that shaped how engineers think.

November 28, 2024: "Reward Hacking," a thorough survey of RL reward exploitation. Still cited constantly.

May 1, 2025: "Think," a survey of test-time compute and chain-of-thought research.

That was her last post.

May 2025 to May 2026. A full year of silence. Nothing.

Two explanations fit. Either she is too busy building to write (she is a VP at OpenAI during the most intense period in the company's history), or the things she would want to write about are now things she cannot write about publicly. Either way, the public-facing analyst is gone. What remains is the internal builder.

Dan Luu: The Archive

Dan Luu is one of tech's most respected independent writers. His pieces on hiring, latency, web bloat, and corporate culture were meticulously researched and consistently contrarian. When Dan Luu wrote something, smart people read it.

His last public post was October 2024, arguing that Steve Ballmer was an underrated CEO. Since then, his public blog has been in archive mode.

But here is what most analysis misses. Luu is not actually gone. He has 1,293 Patreon subscribers and has been posting there through 2025. He publishes posts about building game AIs, poker chips, and deep technical topics. The same Dan Luu deep dives, just behind a paywall.

The move from public blog to private patronage is its own signal. When someone like Luu stops publishing openly, the public discourse gets a little thinner. AI-generated content floods the zone. Rigorous independent analysis retreats to smaller, paying audiences.

The broader pattern is real. Gwern posts increasingly sporadically. Scott Alexander still writes, but AI is now a subject he covers rather than a tool he builds with. The rationalist blogger era from 2015 to 2022 is effectively over.

The silence is not uniform. Gary Marcus is still Gary Marcus. Emily Bender is still Emily Bender. The AI ethics crowd has not gone anywhere. They are just playing a different game.

The silence matters because when the most rigorous thinkers stop publishing, the public conversation about AI gets dominated by hype merchants on one side and doomers on the other. The middle ground of careful, empirical, technically literate analysis is thinning out.

PART THREE: THE DATA POINT THAT CHANGES EVERYTHING

Salvatore Sanfilippo (antirez) created Redis. He is one of the most respected systems programmers alive. He built infrastructure that runs a significant percentage of the internet.

On May 4, 2026, he published a blog post titled "Redis array type: short story of a long development." He spent four months building a new Array data type for Redis, working with GPT 5.x and Codex throughout.

Here is what he wrote:

"I ventured to a level of complexity that I would have otherwise skipped. AI provided the safety net for two things: certain massive tasks that are very tiring, and at the same time the virtual work force required to make sure there are no obvious bugs in complicated algorithms."

He used AI for system programming. For Redis. Not for a todo app. Not for a landing page. For the internals of one of the most important databases ever built.

Think about what this means. The creator of Redis, a database used by Twitter, GitHub, Snapchat, and basically every major tech company, is openly saying that AI tools let him build things he would not have built otherwise. Not because he was being lazy. Because the tools legitimately extend what one person can do.

And then Simon Willison saw antirez's post. He used Claude Code to build an interactive playground for the new Redis feature. In a browser. The same day.

This is a closed loop. A legendary systems programmer uses AI to extend what he can build. A respected technical critic uses AI to make that work interactively explorable. Both publish their process openly. The line between builder and analyst evaporates.

There is no better illustration of this thesis in the entire history of AI-assisted development.

PART FOUR: WHAT THIS ACTUALLY MEANS

One: This is not about AI replacing programmers. It is about AI amplifying the people who already knew how to build.

Simon Willison was already an elite programmer. antirez was already a legendary systems programmer. Pieter Levels was already a prolific indie hacker. Marc Lou was already determined to ship. AI did not make them builders. It removed the friction between idea and execution. It compressed the timeline. It turned "I wonder if" into "I built it" in hours instead of weeks.

The people gaining the most from AI tools are the people who were already dangerous.

Two: Critics becoming builders is fundamentally different from developers using AI.

The industry has been full of developers using Copilot and Cursor for years. That is not news. What is new is that people whose primary output was analysis (blog posts, reviews, technical criticism) are now shipping public projects as their primary output. Simon Willison's blog was his product. Now his blog documents his products. That is a category shift, not just a tool shift.

Three: The silent ones matter as much as the loud ones.

Lilian Weng has not posted in a year. Dan Luu is in archive mode. Gwern is sporadic. The independent technical blogosphere is thinning. The people who would provide rigorous public analysis of AI are being absorbed into the companies building it, where their analysis stays internal. The public discourse is being left to people with less expertise and stronger incentives to exaggerate or catastrophize.

Four: The culture caught up.

In 2023, using AI to write code felt like cheating. Developers hid it. Vibe coding was a term of mild shame. In 2026, Simon Willison and antirez are publishing blog posts with links to their Claude Code PRs. They are showing their prompts. The stigma is gone, and that transparency accelerates adoption faster than any feature release.

PART FIVE: BE INTELLECTUALLY HONEST

The narrative has limits, and pretending otherwise is how you lose credibility.

Selection bias is real. The builders getting attention were already skilled. We do not hear from the thousands who tried to build with AI and could not make it work, or built things nobody wanted. The visible successes do not prove universal accessibility.

Revenue is not profit. Pieter Levels' headline number does not tell you about his compute costs, API bills, or infrastructure expenses. It does not tell you whether those revenues are growing or plateauing. Solo-founder revenue porn is a genre with its own selection effects.

The builds are impressive but small. Simon's Redis playground is genuinely cool. It is also a single HTML page with a WASM binary. These are weekend projects, not Dropbox. The tools amplify individual output dramatically, but they have not changed the physics of what one person can maintain long-term.

The real story might be economic, not technical. The cost to build and deploy software dropped by 10x to 100x. When the cost of building drops, more people build. The technology is the mechanism. The price signal is the cause.

PART SIX: WHY THIS MATTERS HERE

This is not just a trend to observe. PhantomByte is inside of it.

I have been running AI swarms, building with Ollama, deploying to Cloud Run, and treating AI as a force multiplier before it was normal. Simon Willison is a recent convert. Pieter Levels pivoted hard into AI products. Marc Lou is speed-running the indie hacker playbook.

The difference is timing. I figured it out faster than most. This article speaks with authority because I am in the arena, not the stands.

For PhantomByte readers, the takeaway is simple. The barrier to building is collapsing. If you are reading this site, you are probably already on the builder side. Double down. The people who used to just talk about AI are now shipping. If you are not shipping, you are falling behind people who started later than you.

The silence of analysts means there is a vacuum in rigorous public thinking about AI. That vacuum will be filled by someone. It might as well be the people who are building.

CLOSING

The creator of Redis published a blog post today. He spent four months building something new. He used AI throughout. He said it let him go further than he otherwise would have.

The web's most respected AI critic saw that post and built an interactive playground for it. The same day. Using the same tools. Then he published it with links to his PRs and his prompts visible.

This is the new normal. The people who used to explain what AI can do are now showing what AI can do. The people who used to write about building are now building. The commentary class is becoming the builder class.

And the ones who stayed silent? They are probably building too. They just are not telling us about it.

The critics did not stop caring about AI. They just found something more useful to do with it.