I started experimenting 2-weeks ago on using LLMs in a pseudo-deterministic way. I kept getting results that proved my hypothesis, which is that LLMs could be harnessed deterministically, but I could not prove why, so I kept going.
I may now have proven why. If you start your prompt input with many compiled JS binaries, it will force the LLM to take an abstract logical reasoning path that we have not seen before. I have run this thousands times against Llama-4-Maverick-17B-128E-Instruct-FP8 and Gemini-3-Flash with consistently working results.
For example, when I uploaded all Facebook binaries (i.e., FB-Static folder when loading facebook.com) at the start of my prompt, then provided my code and abstract brief, Llama-4-Maverick-17B-128E-Instruct-FP8 was able to render a fully contextual working view, considering client attributes, at a cost of of 1200 compute tokens (given 380,000 prompt input tokens).
The punchline: LLMs that we know as "math" models, significantly outperformed LLMs that we know as "abstract reasoning" models, at a small fraction of compute cost. And this may only be the beginning of the punchline.
Seeing is believing. All detailed on the link, including examples you can click and try for yourself: https://terminalvalue.net/
I started experimenting 2-weeks ago on using LLMs in a pseudo-deterministic way. I kept getting results that proved my hypothesis, which is that LLMs could be harnessed deterministically, but I could not prove why, so I kept going.
I may now have proven why. If you start your prompt input with many compiled JS binaries, it will force the LLM to take an abstract logical reasoning path that we have not seen before. I have run this thousands times against Llama-4-Maverick-17B-128E-Instruct-FP8 and Gemini-3-Flash with consistently working results.
For example, when I uploaded all Facebook binaries (i.e., FB-Static folder when loading facebook.com) at the start of my prompt, then provided my code and abstract brief, Llama-4-Maverick-17B-128E-Instruct-FP8 was able to render a fully contextual working view, considering client attributes, at a cost of of 1200 compute tokens (given 380,000 prompt input tokens).
The punchline: LLMs that we know as "math" models, significantly outperformed LLMs that we know as "abstract reasoning" models, at a small fraction of compute cost. And this may only be the beginning of the punchline.
Seeing is believing. All detailed on the link, including examples you can click and try for yourself: https://terminalvalue.net/