2026-09-10 22:13 UTC
@antoinechambertloir@mathstodon.xyz
Sure. What I had in mind is the old distinction between symbolic AI and connectionist/neural approaches, such as LLMs.
Symbolic AI methods leverage domain knowledge to reason over data, providing transparency through explicit representations and rules, which can include probabilistic rules.
This makes it easier to understand and revise the reasoning process.
However, symbolic methods are typically less effective at learning complex representations directly from data: they have real difficulties capturing intricate low-level patterns (e.g. for task like image recognition) and handling noisy or ambiguous data efficiently.
1/n
Replies (1)
-
@0xdefec7edcafe@piaille.fr 2026-09-10 22:16
@antoinechambertloir@mathstodon.xyz Neural methods, e.g LLMs, by contrast, excel at learning complex patterns from large datasets and are generally more robust to noise and ambiguity. However, their internal representations are distributed across vast numbers of parameters and complex nonlinear transformations, making the model largely a black box: we can observe its inputs and outputs, but it is difficult to trace, interpret, or directly revise the internal mechanisms that connect them. This is why I reacted to the expression "probabilistic rules": my objection is really to the word "rules", not to the fact that probabilities enter the model. 2/n