Post

Hi, just because a software tells you it’s “reasoning” or “thinking” doesn’t mean it actually is. You won’t believe this but programmers can make software lie about what it does. Maybe programmers need more guardrails.

I never understand these arguments. How does this prove it's not reasoning? I can tell from using LLMs, if they can reason, they can't do it very well. But I never get how explaining the algorithm proves it can't think.

You may not have the mind for coding. For following directions, yes, from a person who you think is better than you. "It" is not reasoning, it's following instructions in the program it was spawned from. AI is just computer programming.

The emergent behavior was definitely not programmed.

this behavior cannot be programmed only trained. Models aren’t programmed as much as they are grown.

Because no part of the algorithm even attempts to mimic understanding or comprehension; which is a necessary part of actual thinking and reasoning.

No Animal brain runs on code. A computer chip has a decent number of connections nowhere near as many as a brain, but what’s happening in a computer chip is driven by software which is nothing like an organic brain. A brain takes in data directly and reacts to it without any set of instructions.

Reasoning requires neuron to neuron feedback connections. Without those, the network is incapable of maintaining internal state, taking different paths based on internal state and selecting which states to progress to the output stage. All of these are fundamental features of reasoning

In other words, an LLM can't reason for the same reason as a car can't fly. It lacks wings. And you can't just add internal feedback connections to an LLM because it relies on back propagation which only works with feedforward networks.