📺 View the presentation slides
When television was first invented, they would film people reading radio dramas in front of microphones. People only know how to use a new medium from the analogy of the old medium.
That's why software engineers are the first to use AI tools, and the first thing they do is write code that they code review and want things to be deterministic. Cal AI was vibe coded and keeps track of your calories, threatening MyFitnessPal.
But once we melt into the medium, we will realize the most important part is the business case: speed of shipping to utility, capex, cost of development. The code isn't necessary. No compilation. We will just have the specs run directly on ClaudeVM.
The 10X Pattern #
When the JVM came out, most COBOL and C engineers kept doing what they do—even til now. But there were 10X more Java engineers, and that code swamped businesses. Same for V8 and 10X more JavaScript engineers at startups. Software engineers of today will stick with what they do and be useful for maintaining precise infrastructure of a kind.
But we are about to be hit with hundreds of millions of Englishscript engineers that will write scripts that swamp all other code and utility, running directly on ClaudeVM, fulfilling the end goal of software eating the world. Each script tweaked with a few words from a prior script, enabling totally new uses, each perfectly customized to the user. Billions of unique programs.
So cheap I'm not sure they'll be startups as much as blog posts, free as long as you sign up to the mailing list. Custom open source Englishscript that designs the perfect meal plan—photo detection, flexibility and variety, notifications with exact calories for you—hand-designed for you in seconds with white-glove human onboarding.
And you'd want a phone that can natively run ClaudeVM Englishscripts directly—maybe figure out a payment harness in the OS—not the heaviness of app compilation, App Store, binaries.
Vibe Coding is Radio on Television #
Vibe coding is radio on television. Coding is the wrong word. Coding, even as an intermediary step, is dead. It'll just be Englishscript.
FAQ #
Why a VM? Why not compile to binary?
The OODA loop is easier and faster for 100 million Englishscripters. The end user is also live debugging, healing, and adding features to it. Like Lisp!
Why do engineers disagree?
Engineers have a hard time seeing this, same as radio stars had a hard time seeing television.
But it's not good code?
I think the bitcode that V8 generates isn't "good" code either. Not reusable. Often repetitive (loop unrolling, etc). Not DRY. Doesn't matter.
Isn't this inefficient?
We are at the slow JIT stage.
Will we still need C++ engineers?
Sure, there's tons of old stuff not getting changed. Like COBOL running banks still, decades later. But AI isn't touching any of that meaningfully anyway. Value too low. Auditing too complex. We'll use limited GPU power on the 10X flexible value creation on new stuff. Doesn't matter if it can't write assembly or COBOL for years.
The analogy is more like: it's cheaper to set up a whole new solar plant (capex) than to maintain the coal power plants for one more year (opex).
In the same way C/C++ still runs all operating systems, we will still have that infrastructure. COBOL banks. I don't think classic engineers will lose their jobs. But a whole slew of apps will be 100X more engineers making 100X more apps and 100X more (very inefficiently computed) value. Like JavaScript and Electron. Even Claude desktop is Electron.
Does this require AGI?
No. Not AGI. Not ASI. Just a fast JIT.
Can you replace the web?
One idea: what if you have a new type of browser, except when you click View Source it shows just the English prompt? The website you see rendered is generated dynamically from the prompt directly in Claude, with many intermediary throwaway HTML/CSS. The World Wide Web, but even more accessible than HTML.
How will Englishscript writers get paid?
Software will become as easy to make as music. So easy to copy, too. Piracy rampant. As a solution, there will be a Spotify for software: you'll pay a subscription and the software writers will get paid like a penny per use. Modeling music copyright.
Isn't this sloooow?
Imagine Talaas-level speed and a scratch space to run the code—that's a VM.
What about MCP?
MCP is tired because it is running code instead of English.
People are making cool tools like CodeSpeak?
The Kotlin creator's new "language": codespeak.dev. Getting so close to the idea. We will only have Englishscripts and don't need code anymore. No compiling. No vibe coding. No coding.
What is Englishscript really?
It's what you've been doing writing agents using English, not code. But extend that all the way to everything. The final endgame.
It's like a skill or a prompt, yes. It's a viewpoint, but nothing new is created. The VM will be something new that Claude will ship more around, though. Right now it's haphazard, or people isolate by buying Mac Minis. First sandboxes are coming out but still not fully virtualized.
Aren't Englishscripts less precise than code?
Englishscripts can get as detailed or zoomed out as you want or feel comfortable. They can even have literate-program-style snippets of code, or pseudocode, as desired. But the more code you add, the less flexible, adaptable, forwards-compatible, reusable, and mergeable it becomes.
Don't people just call this a product spec?
A spec is for generating code. There's a big difference when you're not even generating code—when Claude is just running it directly. Note that a spec is also different from a "skill." Put another way: what if the spec is the skill? If that connects the dots.
Doesn't this have to generate code at some point?
Nope. Go learn about how V8 works. It has to make kernel calls but doesn't necessarily need any intermediate code that we'd ever look at. At best, VM bitcode—not Rust.
I'm not even talking about the future at this point. Many of your agents are already English → value. Doing things that, last year, you would have written in code or MCP code. Just go all the way. No code.
Will we still have code review?
No, since we don't code. Do you code review the running of your agents written in English? Either way, we aren't going to code review. Sometimes we will drop in—just like sometimes people optimize the VM or fix the JVM or add assembler (which can easily be mentioned in the skill/prompt/Englishscript). But usually not.
Don't you need strictness and repeatability?
That's what engineers tell me is different. But is repeatability or strictness the most important thing for code? Or just providing the value? Or is it a coincidence due to computers being so strict in the past and needing to be careful with outputs for later programs?
Also, LLMs can be made deterministic with a seed, if you need that. Entropy is just PRNG. You can make that deterministic.
Where does this all end?
I now think we will top out the logistic curve of LLMs to be a fast JIT. It won't be AGI, but it is useful and world-changing as the "last virtual machine."
I haven't seen people analogize LLMs to the V8 VM or JVM before. Bitcode isn't read and it's hard, but it kind of doesn't matter even if we are currently in the "slow JIT" phase where the code is bad and bloated and slow.
Going from slow JIT to fast JIT doesn't require magic or AGI—just lots and lots of optimization by humans.
"Skills" or "prompts" or "agents" don't have that viewpoint. But they should.
Previously: Why Claude Runs on Electron and Not ClaudeVM · Claude is a JIT
Enjoyed this essay?
Follow me for more insights on technology, startups, and the future.