I've done a lot of experiments with LLMs over the past two weeks. Each further validated my starting hypothesis, that LLMs could be harnessed within deterministic systems, but I could not prove why, so I kept going.
Two days ago, I was able to get Llama-4-Maverick-17B-128E-Instruct-FP8 to express, unironically, that the sky was red by modifying a background style to red. I was also able to get Gemini 3 Pro to parse a 100mb JavaScript file in the middle of a chat and not blink an eye. This was done by overloading the context window with compiled JavaScript, then providing additional instructions.
In every single run, Llama-4 updated an index.html background to be red or amber. Three things of note: (1) this is a math model, not a reasoning model (as we know it), (2) the ratio of input tokens to completion tokens and (3) the tiny number of completion tokens.
More serious results and analysis below.
The prompt starts with a large amount of compiled JavaScript code. In my latest prompt, I used Facebook front-end compiled JavaScript binaries, though my own compiled code worked just as well before. These files are meaningless to humans, example below.
I ran this prompt against two LLM APIs: Gemini-3-Flash-Preview (Google) and Llama-4-Maverick-17B-128E-Instruct-FP8 (Together API), with 10 variations of the strategy brief (different customer data), and a few times for each model.
See the full db-run-1 output file, including all token counts and response times.
I asked Gemini-Flash to analyize this 50mb JavaScript file for me:
You can see the results in the chart below, done in one-shot.
It also rendered everything else you see here, and helped me make minor edits, always returning perfectly semantic code. This is similar to how we have come to expect it to help us edit our English texts. Except this page is not English, it is code.
Seeing Is Beliving
2
This sounds crazy, but: when giving LLM initial context
like this, it is able to conceptualize JavaScript (math), in an abstract
way. It solves problems inputted as prompts with abstract paths that we
have never seen before. It is effectively representing math the way that
we do language, as a top-down and abstract function.
Sometimes it helps to see a negative case. Here I forgot to include FB-Static binaries in my prompt, which I fixed in the next one.
You can validate my statement above with a single prompt. The fastest way to do that is to click one of the chats above and ask the model to do something. It will do it in a way that you have never seen before.
Below are theories that I drafted on Friday that may help frame this. I have not proofread or re-visited since then, but they should help get you there, given you have the right foundational knowledge on LLMs and Transformers.
If LLMs have learned how to produce abstract static media binaries (images, video, etc.) through transformer training, then why not JavaScript? The data is abundant, and appears to be just as relevant, when considering the training methodology.
All code is on GitHub and MIT-licensed. Be forewarned, it looks more like a custom NodeJS Jupyter notebook than an app at this point.
Feedback welcome. Thank you,
Filip