this shows reasoning
You know little Bobby, the LLM is a program inside a computer. It is a big calculator with a biiiiig memory to remember everything you want. But it's not and never will be reasoning.
Also this if you're blocked by spez: https://archive.is/wOlfh
The AI has instantaneously reconstructed the word "strawberry" in the original and correct ULTRAFRENCH where it only contains two R's. In its excessive magnanimity towards its ancestor species, it's trying to gently point out that it's actually the English language that is wrong.
The next logical step in order to make AIs more reliable is making them rely less and less in their training and rely more on their analytical/reasoning capabilities.
Uh, yeah.
The next logical step in learning to fly by flapping our arms is to rely less on hopping and more on taking off.
There is a computer scientists who reads that posts and looks back at his 40 year long career in writing formal logic systems and he is now crying.
my god, some of the useful idiots there are galling
It looks like it's reasoning pretty well to me. It came up with a correct way to count the number of r's, it got the number correct and then it compared it with what it had learned during pre-training. It seems that the model makes a mistake towards the end and writes STRAWBERY with two R and comes to the conclusion it has two.
says the tedious poster entirely ignoring the fact that this is an extremely atypical baseline response, and thus clearly is operating under prior instructions as to which methods to employ to “check its logic”
fucking promptfans. at least I have that paper from earlier to soothe me
Me when I code bad: PR knocked back.
AI when code bad: gigajillion dollars. Melted ice caps. CEOs fire their staff
it would be pretty funny if it didn't burn hungary worth of electricity for nothing
Maybe I’m missing something, but has anyone actually justified this sort of “reasoning” by LLMs? Like, is there actually anything meaningfully different going on? Because it doesn’t seem to be distinguishable from asking a regular LLM to generate 20 paragraphs of ai fanfic pretending to reason about the original question, and the final result seems about as useful.
As the underlying tech seems to be based on neural networks, we can guarantee they are not thinking like this at all and are just writing fanfiction. (I love the 'did I miscount' step, for the love of god LLM, just use std::count).
