jilliterate

Javascript engine made by LLM interpreting the ECMA specification

Status: alpha.

What?

ECMA specification [https://tc39.es/ecma262/] consists of "algorithms" in informal but pedantic structured English language. As an experiment, I tried to feed it to LLM. The result surpassed my expectations.

Which one?

I get the best results with Mistral (namely codestral) [https://mistral.ai/].

I also tried Qwen run locally through Jan.ai [https://jan.ai/]. Qwen has a tendency to chat, that means it generates human-readable output instead of Python code.

Gemma through Groq [https://groq.com/groqcloud/] is the fastest model but makes syntax errors and can't understand some instructions.

How?

1. Download

First, check out the repository. Then run the download script, that will download ECMA spec and convert it to XML.

./download.py

Then run the specification script that will convert it to a Python pickle format.

./specification.py

2. Configure

Configure your favourite LLM platform. Provide the URL to OpenAI-compatible endpoint and export the API key, then test it.

export LLM_API_URL=https://api.mistral.ai/v1
export LLM_API_KEY=<your API key>
export LLM_MODEL=codestral-latest
export LLM_CONFIG_EXTRA='{}'

./codegen.py

The script will generate some algorithms. Check if they are right. You may put the environment variables to the file my_platform.sh and source it. Always make sure the file is in .gitignore. Do not commit your API keys.

Some example AI platform configs are in the file llm_platforms.sh.

3. Generate

Finally, generate the interpreter.

./generate.py

This script will send parts of the spec to the configured AI platform. Note that this will take several hours and will likely deplete your token limit.

Now go to gencode directory and run each file through Python interpreter to check for syntax errors.

cd gencode
python3 ./ao_library.py
python3 ./early_errors.py
python3 ./evaluate.py
python3 ./sdo_library.py

4. Implement

TODO: implement the backend, EBNF parser and all internal structures and you have the Javascript engine. Try doing it by feeding chapters of ECMA grammar to AI, with an appropriate system prompt.

Now what?

With a little help from human devs, we could have a working Javascript interpreter that will beat Chrome and end Google's monopoly.

LLMs could be a whole new paradigm in programming, a level above programming languages, just like high-level languages are a layer above assembler.

I see similar projects implementing applications that have a formal specification, especially interpreters and web standards.

I'm in

If you wish to contribute, try the following:

1. Find more LLM platforms and test if they work. Add them to llm_platforms.sh.
2. Modify the prompt, especially to fix syntax errors in the generated code ([prompt.xml]).
3. Try fine-tuning the model.
4. Implement EBNF grammar parser with all necessary features needed by Javascript. Try doing it by constructing a prompt for AI.
5. Implement all the backend structures referenced by the ECMA grammar by feeding pieces of the spec into AI.
6. Make a better logo (AI-generated images preferred.)

Also look at my other project, an SVG-first web browser: [https://github.com/haael/guixmpp].

Help

Sponsor me, as I need to pay for Mistral!

[https://github.com/sponsors/haael]

Spread the word

Share this project on social media.

Send me crypto

• Bitcoin: bc1q0neyed0cug40jce9t8xxswsnr68w8rs788u5ws
• Ethereum: 0x4D7914F809D7C101dCe71258FA5498475AED18c9
• Solana: 7KGJWfcoopQRbgcLGfrKyLPykhdzf5eE88tGth1jjuEj

Bitcoin	Ethereum	Solana