antirez 3 days ago. 21234 views.

Anthropic recently released a blog post with the description of an experiment in which the last version of Opus, the 4.6, was instructed to write a C compiler in Rust, in a “clean room” setup.

The experiment methodology left me dubious about the kind of point they wanted to make. Why not provide the agent with the ISA documentation? Why Rust? Writing a C compiler is exactly a giant graph manipulation exercise: the kind of program that is harder to write in Rust. Also, in a clean room experiment, the agent should have access to all the information about well established computer science progresses related to optimizing compilers: there are a number of papers that could be easily synthesized in a number of markdown files. SSA, register allocation, instructions selection and scheduling. Those things needed to be researched *first*, as a prerequisite, and the implementation would still be “clean room”.

Not allowing the agent to access the Internet, nor any other compiler source code, was certainly the right call. Less understandable is the almost-zero steering principle, but this is coherent with a certain kind of experiment, if the goal was showcasing the completely autonomous writing of a large project. Yet, we all know how this is not how coding agents are used in practice, most of the time. Who uses coding agents extensively knows very well how, even never touching the code, a few hits here and there completely changes the quality of the result.

The Z80 experiment

I thought it was time to try a similar experiment myself, one that would take one or two hours at max, and that was compatible with my Claude Code Max plan: I decided to write a Z80 emulator, and then a ZX Spectrum emulator (and even more, a CP/M emulator, see later) in a condition that I believe makes a more sense as “clean room” setup. The result can be found here: https://github.com/antirez/ZOT.

The process I used

1. I wrote a markdown file with the specification of what I wanted to do. Just English, high level ideas about the scope of the Z80 emulator to implement. I said things like: it should execute a whole instruction at a time, not a single clock step, since this emulator must be runnable on things like an RP2350 or similarly limited hardware. The emulator should correctly track the clock cycles elapsed (and I specified we could use this feature later in order to implement the ZX Spectrum contention with ULA during memory accesses), provide memory access callbacks, and should emulate all the known official and unofficial instructions of the Z80.

For the Spectrum implementation, performed as a successive step, I provided much more information in the markdown file, like, the kind of rendering I wanted in the RGB buffer, and how it needed to be optional so that embedded devices could render the scanlines directly as they transferred them to the ST77xx display (or similar), how it should be possible to interact with the I/O port to set the EAR bit to simulate cassette loading in a very authentic way, and many other desiderata I had about the emulator.

This file also included the rules that the agent needed to follow, like:

* Accessing the internet is prohibited, but you can use the specification and test vectors files I added inside ./z80-specs. * Code should be simple and clean, never over-complicate things. * Each solid progress should be committed in the git repository. * Before committing, you should test that what you produced is high quality and that it works. * Write a detailed test suite as you add more features. The test must be re-executed at every major change. * Code should be very well commented: things must be explained in terms that even people not well versed with certain Z80 or Spectrum internals details should understand. * Never stop for prompting, the user is away from the keyboard. * At the end of this file, create a work in progress log, where you note what you already did, what is missing. Always update this log. * Read this file again after each context compaction.

2. Then, I started a Claude Code session, and asked it to fetch all the useful documentation on the internet about the Z80 (later I did this for the Spectrum as well), and to extract only the useful factual information into markdown files. I also provided the binary files for the most ambitious test vectors for the Z80, the ZX Spectrum ROM, and a few other binaries that could be used to test if the emulator actually executed the code correctly. Once all this information was collected (it is part of the repository, so you can inspect what was produced) I completely removed the Claude Code session in order to make sure that no contamination with source code seen during the search was possible.

3. I started a new session, and asked it to check the specification markdown file, and to check all the documentation available, and start implementing the Z80 emulator. The rules were to never access the Internet for any reason (I supervised the agent while it was implementing the code, to make sure this didn’t happen), to never search the disk for similar source code, as this was a “clean room” implementation.

4. For the Z80 implementation, I did zero steering. For the Spectrum implementation I used extensive steering for implementing the TAP loading. More about my feedback to the agent later in this post.

5. As a final step, I copied the repository in /tmp, removed the “.git” repository files completely, started a new Claude Code (and Codex) session and claimed that the implementation was likely stolen or too strongly inspired from somebody else's work. The task was to check with all the major Z80 implementations if there was evidence of theft. The agents (both Codex and Claude Code), after extensive search, were not able to find any evidence of copyright issues. The only similar parts were about well established emulation patterns and things that are Z80 specific and can’t be made differently, the implementation looked distinct from all the other implementations in a significant way.

Results

Claude Code worked for 20 or 30 minutes in total, and produced a Z80 emulator that was able to pass ZEXDOC and ZEXALL, in 1200 lines of very readable and well commented C code (1800 lines with comments and blank spaces). The agent was prompted zero times during the implementation, it acted absolutely alone. It never accessed the internet, and the process it used to implement the emulator was of continuous testing, interacting with the CP/M binaries implementing the ZEXDOC and ZEXALL, writing just the CP/M syscalls needed to produce the output on the screen. Multiple times it also used the Spectrum ROM and other binaries that were available, or binaries it created from scratch to see if the emulator was working correctly. In short: the implementation was performed in a very similar way to how a human programmer would do it, and not outputting a complete implementation from scratch “uncompressing” it from the weights. Instead, different classes of instructions were implemented incrementally, and there were bugs that were fixed via integration tests, debugging sessions, dumps, printf calls, and so forth.

Next step: the ZX Spectrum

I repeated the process again. I instructed the documentation gathering session very accurately about the kind of details I wanted it to search on the internet, especially the ULA interactions with RAM access, the keyboard mapping, the I/O port, how the cassette tape worked and the kind of PWM encoding used, and how it was encoded into TAP or TZX files.

As I said, this time the design notes were extensive since I wanted this emulator to be specifically designed for embedded systems, so only 48k emulation, optional framebuffer rendering, very little additional memory used (no big lookup tables for ULA/Z80 access contention), ROM not copied in the RAM to avoid using additional 16k of memory, but just referenced during the initialization (so we have just a copy in the executable), and so forth.

The agent was able to create a very detailed documentation about the ZX Spectrum internals. I provided a few .z80 images of games, so that it could test the emulator in a real setup with real software. Again, I removed the session and started fresh. The agent started working and ended 10 minutes later, following a process that really fascinates me, and that probably you know very well: the fact is, you see the agent working using a number of diverse skills. It is expert in everything programming related, so as it was implementing the emulator, it could immediately write a detailed instrumentation code to “look” at what the Z80 was doing step by step, and how this changed the Spectrum emulation state. In this respect, I believe automatic programming to be already super-human, not in the sense it is currently capable of producing code that humans can’t produce, but in the concurrent usage of different programming languages, system programming techniques, DSP stuff, operating system tricks, math, and everything needed to reach the result in the most immediate way.

When it was done, I asked it to write a simple SDL based integration example. The emulator was immediately able to run the Jetpac game without issues, with working sound, and very little CPU usage even on my slow Dell Linux machine (8% usage of a single core, including SDL rendering).

Once the basic stuff was working, I wanted to load TAP files directly, simulating cassette loading. This was the first time the agent missed a few things, specifically about the timing the Spectrum loading routines expected, and here we are in the territory where LLMs start to perform less efficiently: they can’t easily run the SDL emulator and see the border changing as data is received and so forth. I asked Claude Code to do a refactoring so that zx_tick() could be called directly and was not part of zx_frame(), and to make zx_frame() a trivial wrapper. This way it was much simpler to sync EAR with what it expected, without callbacks or the wrong abstractions that it had implemented. After such change, a few minutes later the emulator could load a TAP file emulating the cassette without problems.

This is how it works now:

        do {
            zx\_set\_ear(zx, tzx\_update(&tape, zx->cpu.clocks));
        } while (!zx\_tick(zx, 0));

I continued prompting Claude Code in order to make the key bindings more useful and a few things more.

CP/M

One thing that I found really interesting was the ability of the LLM to inspect the COM files for ZEXALL / ZEXCOM tests for the Z80, easily spot the CP/M syscalls that were used (a total of three), and implement them for the extended z80 test (executed by make fulltest). So, at this point, why not implement a full CP/M environment? Same process again, same good result in a matter of minutes. This time I interacted with it a bit more for the VT100 / ADM3 terminal escapes conversions, reported things not working in WordStar initially, and in a few minutes everything I tested was working well enough (but, there are fixes to do, like simulating a 2Mhz clock, right now it runs at full speed making CP/M games impossible to use).

What is the lesson here?

The obvious lesson is: always provide your agents with design hints and extensive documentation about what they are going to do. Such documentation can be obtained by the agent itself. And, also, make sure the agent has a markdown file with the rules of how to perform the coding tasks, and a trace of what it is doing, that is updated and read again quite often.

But those tricks, I believe, are quite clear to everybody that has worked extensively with automatic programming in the latest months. To think in terms of “what a human would need” is often the best bet, plus a few LLMs specific things, like the forgetting issue after context compaction, the continuous ability to verify it is on the right track, and so forth.

Returning back to the Anthropic compiler attempt: one of the steps that the agent failed was the one that was more strongly related to the idea of memorization of what is in the pretraining set: the assembler. With extensive documentation, I can’t see any way Claude Code (and, even more, GPT5.3-codex, which is in my experience, for complex stuff, more capable) could fail at producing a working assembler, since it is quite a mechanical process. This is, I think, in contradiction with the idea that LLMs are memorizing the whole training set and uncompress what they have seen. LLMs can memorize certain over-represented documents and code, but while they can extract such verbatim parts of the code if prompted to do so, they don’t have a copy of everything they saw during the training set, nor they spontaneously emit copies of already seen code, in their normal operation. We mostly ask LLMs to create work that requires assembling different knowledge they possess, and the result is normally something that uses known techniques and patterns, but that is new code, not constituting a copy of some pre-existing code.

It is worth noting, too, that humans often follow a less rigorous process compared to the clean room rules detailed in this blog post, that is: humans often download the code of different implementations related to what they are trying to accomplish, read them carefully, then try to avoid copying stuff verbatim but often times they take strong inspiration. This is a process that I find perfectly acceptable, but it is important to take in mind what happens in the reality of code written by humans. After all, information technology evolved so fast even thanks to this massive cross pollination effect.

For all the above reasons, when I implement code using automatic programming, I don’t have problems releasing it MIT licensed, like I did with this Z80 project. In turn, this code base will constitute quality input for the next LLMs training, including open weights ones.

Next steps

To make my experiment more compelling, one should try to implement a Z80 and ZX Spectrum emulator without providing any documentation to the agent, and then compare the result of the implementation. I didn’t find the time to do it, but it could be quite informative.