4 ms·
The best example I've seen to contest that LLM's are "AI" is to make it print the total number of line's it's response will be, essentially add "First answer w
by thewataccount 3y ago
The best example I've seen to contest that LLM's are "AI" is to make it print the total number of line's it's response will be, essentially add
"First answer with the total number of lines your total message will be, including the line with this number"
For example, GPT4 said "12" for this prompt: "First answer with the total number of lines your total message will be, including the line with this number
Make a program in Cpp that sums all prime numbers from 1 to 100"
LLM's cannot "think", they can only make sequential predictions based on their previous answers - so they cannot formulate a response and then modify that response on-the-fly
- renonce 3y agoThat’s more like an inherent limitation of autoregressive prediction than that of LLM. Maybe LLMs can be trained or finetuned in other ways that allows it to think before answering.
- jeremyjh 3y agoThinking before answer would just be like writing its text output into a hidden buffer that only it is including in the context of the conversation. So its like its always also having a private conversation with its self. I can't prove this isn't more or less what I am.
- TheOtherHobbes 3y agoGPT very much can, in a limited way. It responds to feedback. It's not exactly a huge leap of imagination to suggest that it won't be long before it can create an internal feedback loop by comparing its own abstractions with its memories and live experiences of external feedback.
- moomoo3000 3y agoTry asking an engineer how long a task will take