Yeah they're not though – pretty much all of the limitations of current LLM capabilities come down to their architecture as next-token predictors
@simonw
-
Open Licensed Small Model More Interesting Than GPT-4.5
By
–
If it's a small model that they're going to openly license that makes it massively more interesting than if it really is a GPT 4.5 preview
-
LLMs Cannot Claim Authorship: Statistical Autocomplete Limitation
By
–
If I say "I am proud of this piece of writing and it represents my best current understanding of this matter" that means something – to people who follow my work, at least An LLM can't say that, because an LLM is statistical autocomplete
-
Reputation and Honesty Over Accuracy in AI Systems
By
–
It's not about accuracy or reliability, it's about being willing to stake your personal reputation on something being honest and true
-
AI-Generated Content Lacks Human Credibility Stake
By
–
The big thing AI-generated content will always lack is credibility An LLM cannot stake its credibility on something being accurate or true or honest If I'm going to spend time with content, I want a human to stake their reputation on it being worth my time
-
Redefining AI: From Broad Concept to LLM-Powered Chatbots
By
–
I think we can repurpose it for the current moment, kind of like how the term "AI" has mostly been repurposed to mean LLM-powered chat bots at this point
-
Slop: Understanding Unwanted AI-Generated Content
By
–
Slop is the new name for unwanted AI-generated content
-
LMSYS Arena: Anonymous Model Testing Platform Documentation
By
–
Apparently the LMSYS arena is used for this kind of anonymous model testing quite often, they have documentation about that here: