84

Source document

[2025-12-09 18:02] ajaxdavis: yeah (random image), im thinking there is always a background process or at least some hypothetical tree of most likely tool calls that will be called in the plan, but you are only loading the most likely for each next step (to reduce big context) (files: AGfvwn8stbleAAAAAElFTkSuQmCC.png) [2025-12-09 18:02] traves_theberge: i like it. [2025-12-09 18:02] ajaxdavis: small context + likely to choose the correct tool and load the correct context [2025-12-09 18:03] ajaxdavis: hierarchical cascading tree tool and context loader [2025-12-09 18:06] ajaxdavis: whats kind of cool, is if you had a debugger, and you let the tree grow out to 100 levels of potential/possible outcomes, you could almost entirely predict the entire body of work [2025-12-09 18:06] traves_theberge: how do you constrain this for a deterministic business outcome is one question i have. assuming there is a finite tasks for use case X. and we have come full circle to the kernel / Controller Area Network bus (CAN BUS [2025-12-09 18:07] traves_theberge: "This broadcast-based, message-oriented protocol ensures data integrity and prioritization through a process called arbitration, allowing the highest priority device to continue transmitting if multiple devices attempt to send data simultaneously" [2025-12-09 18:07] ajaxdavis: i suppose you save it when you are happy with one path (call it a skill template lol) [2025-12-09 18:09] ajaxdavis: well actually [2025-12-09 18:09] ajaxdavis: so in my day job use cases, i wouldn't actually want to save it all that much, make it deterministic [2025-12-09 18:09] ajaxdavis: because its user requests, although they are doing the same thing most of the time [2025-12-09 18:09] ajaxdavis: they might deviate by one tool call here and there [2025-12-09 18:10] ajaxdavis: so instead of saving the "Deterministic" [2025-12-09 18:10] ajaxdavis: i store information about what is the most used pathways, so anytime its generating new trees of pathways, the main one could have a K factor that uses the learned pathways more often [2025-12-09 18:10] ajaxdavis: if the learned pathway is used 99.99% of the time, it essentially approches "deterministic" [2025-12-09 18:11] traves_theberge: right. i like the infinite possibilities, with finite repeatability for X use case. but how do you weight the tool calls in this case of probability to injecting in the tool plan based on the users intent. it seems like one is asking for a fragile execution. [2025-12-09 18:13] traves_theberge: "learned" how. amount of use? [2025-12-09 18:15] traves_theberge: Occam's Razor - when multiple explanations exist, the one requiring the fewest assumptions is the most likely to be true [2025-12-09 18:19] traves_theberge: https://en.wikipedia.org/wiki/Conjunction_fallacy [2025-12-09 18:19] traves_theberge: "more is not always better" [2025-12-09 18:22] ajaxdavis: yeah lots of different ways to skin that cat