< backas a means to an end
February 19, 2026Rangefinder is a micro-utility for the future implementation of TurboFire. Being able to produce poker ranges through a TUI was a leap forward in developing the solver. When previously utilizing LLMs, they would hallucinate their own heads-up ranges for a given configuration, which were often wrong. Human intervention and expertise were required to provide the LLM, and thus the solver, with proper ranges so that the solver could be developed and debugged correctly.
Rangefinder was completely AI generated. It existed as a means to an end for providing structure to the ranges TurboFire would consume, allowing me to craft and edit precise ranges (with percentage accuracy) for the solver. LLMs do not understand the dynamics of heads-up big-blind vs button poker, I doubt LLMs ever will understand the fine details of poker. The LLM does not understand why T7o should not be a 100% call from the sb, or that we're supposed to be opening 80% of hands from the button.
Providing precise ranges to the solver had multiple benefits. Using Rangefinder gave structure to the LLM in terms of design, and more concretely, the input format for the solver. In previous LLM-generated implementations, the models would generate and utilize their own input range format. Working backwards to modify these ranges or change the structure of the input resulted in buggy software, compared to implementing a range-plan beforehand and having the LLM implement the pre-existing input structure. Although this detail sounds obvious in hindsight, its importance did not occur to me when writing the "one-shot" solver generated by an LLM.
I have made significant progress in making a TUI solver. Having a background in poker dynamics, the algorithm behind solvers, LLM coding, and low level programming all aided in the breakthrough of an AI-generated version of TurboFire. Many frustrating hours of work are ahead - reviewing the code, abstracting parts of the solver, and adding the features such as automatically solving completed hands in an agent-style format. I anticipate many roadblocks, frustrated blog posts, and overheated Macbooks in the future.
