[jnovek]

You may be anthropomorphizing the model, here. Models don’t have “assumptions”; the problem is contrived and most likely there haven’t been many conversations on the internet about what to do when the car wash is really close to you (because it’s obvious to us). The training data for this problem is sparse.

[tsimionescu]

I may be missing something, but this is the exact point I thought I was making as well. The training data for questions about walking or driving to car washes is very sparse; and training data for questions about walking or driving based on distance is overwhelmingly larger. So, the stat model has its output dominated by the length-of-trip analysis, while the fact that the destination is "car wash" only affects smaller parts of the answer.

[cowboylowrez]

I got your point because it seemed that you were precisely avoiding the anthropomorphizing and in fact seemed to be honing in on whats happening with the weights. The only way I can imagine these models are going to work with trick questions lies beyond word prediction or reinforcement training UNLESS reinforcement training is from a complete (as possible) world simulation including as much mechanics as possible and let these neural networks train on that.

Like for instance, think chess engines with AI, they can train themselves simply by playing many many games, the "world simulation" with those is the classic chess engine architecture but it uses the positional weights produced by the neural network, so says gemini anyways:

"ai chess engine architecture"

"Modern AI chess engines (e.g., Lc0, Stockfish) use a hybrid architecture combining deep neural networks for positional evaluation with advanced search algorithms like Monte-Carlo Tree Search (MCTS) or alpha-beta pruning. They feature three core components: a neural network (often CNN-based) that analyzes board patterns (matrices) to evaluate positions, a search engine that explores move possibilities, and a Universal Chess Interface (UCI) for communication."

So with no model of the world to play with, I'm thinking the chatbot llms can only go with probabilities or what matches the prompt best in the crazy dimensional thing that goes on inside the neural networks. If it had access to a simple world of cars and car washes, it could run a simulation and rank it appropriately, and also could possibly infer through either simulation or training from those simulations that if you are washing a car, the operation will fail if the car is not present. I really like this car wash trick question lol

[wongarsu]

Reasoning automata can make assumptions. Lots of algorithms make "assumptions", often with backtracking if they don't work out. There is nothing human about making assumptions.

What you might be arguing against is that LLMs are not reasoning but merely predicting text. In that case they wouldn't make assumptions. If we were talking about GPT2 I would agree on that point. But I'm skeptical that is still true of the current generation of LLMs

[jabron]

I'd argue that "assumptions", i.e. the statistical models it uses to predict text, is basically what makes LLMs useful. The problem here is that its assumptions are naive. It only takes the distance into account, as that's what usually determines the correct response to such a question.

[jnovek]

I think that’s still anthropomorphization. The point I’m making is that these things aren’t “assumptions” as we characterize them, not from the model’s perspective. We use assumptions as an analogy but the analogy becomes leaky when we get to the edges (like this situation).