I am solving AI Math Hallucinations with Hissab
Posted by prenx4x@reddit | LocalLLaMA | View on Reddit | 9 comments
We all know how bad AI is at Math. Therefore I am building Hissab. So Instead of letting LLMs guess at numerical answers, Hissab turns LLMs into interpreters. Users describe a problem in natural language, and the LLM translates it into precise Hissab expressions. These are then computed by my deterministic calculation engine, guaranteeing reliable and accurate answers.
How Hissab Works:
Natural language prompt → LLM → Hissab expressions → Hissab Engine → Accurate result → LLM → Final response
What do you think of this way of doing calculations with AI? Any feedback is appreciated.
Eden1506@reddit
You can just give the model access to a python environment via tool use...
prenx4x@reddit (OP)
Isn't it much slower/inefficient/expensive to generate new code everytime? Hissab expressions are light weight and almost natural language like. so generating them is easy and faster for an LLM. Plus the code to parse and compute the expressions is also already ready, so the results are computed faster and costs less.
prenx4x@reddit (OP)
I can understand the downvotes, but again not everyone have access to a python environment specially when using the APIs. "Normal" Consumers / non-tech people don't/won't to be doing math like that with LLMs.
lothariusdark@reddit
While I dont mind this post in particular that much, this sub really needs a [promo] or [ad] flair for this.
private_final_static@reddit
The project sound amazing.
And also looks like you are solving a.problem that can be generalized.
That is, turning user requests into annaccurate DSL that can be processed into an interpreter.
Im curious how the step for translating an user free form text into Hissab DSL work.
prenx4x@reddit (OP)
Thanks. I don't think we can have a fully generalized interpreter that can do anything. I mean we can and have i.e. the programming language interpreter. But we can have specialized deterministic engines like a math engine that I have here.
I am infact working on a image generator where the image is rendered by a JSON defined layers array. So the AI will produce this JSON and the renderer (Calling it photomatic) will render the final image. Thus you can have deterministic image generation with pixel perfect accuracy with AI.
This JSON can now be shared, reused, or used in CICD etc.
Lazy-Pattern-5171@reddit
Oh nevermind it has weird pricing structure. 20$/month is definitely a stretch. Is it possible to create a one time sell version with more money with updates that can allow us to hook this up with local LLMs?
prenx4x@reddit (OP)
Yes! I am open to licensing the Hissab engine to make it work with your own LLMs. You can reach me at mufaddal@hissab.io or https://www.linkedin.com/in/prenx4x/
Lazy-Pattern-5171@reddit
Looks amazing! Just the thing I need. Can it work with local LLMs?