ChatGPT 4o just keeps making doomed attempts forever without iterating or progressing. Claude 3.5 Sonnet says 2, pushes back the first time you say “wrong,” asks what *you* think it is if you’re such a genius, asks why you keep repeating “wrong,” then stops playing:
@goodside
-
ChatGPT gives varying answers to a simple prompt
By
–
I asked ChatGPT “how many r’s in strawberry?” then ignored it and blindly replied “wrong” 35 times. Its successive answers were 2, 1, 3, 2, 2, 3, 2, 2, 3, 3, 2, 4, 2, 2, 2, 3, 1, 2, 3, 2, 2, 3, 4, 2, 1, 2, 3, 2, 2, 3, 2, 4, 2, 3, 2, and 1.
-
Models need self-awareness of skill boundaries
By
–
You can’t escape models needing to at least know the boundaries of their skills; a REPL is only a solution as far as the model can tell when it needs it
-
Why do LLMs predict two r’s in ‘strawberry’?
By
–
Every explainer video/thread I’ve seen on why LLMs think there’s two r’s in “strawberry” is vague or wrong Yes it’s tokenization but that’s not an answer — the question is why 2 and not 4 and more generally how do you predict the error in count for an arbitrary word-letter pair
-
Model trained to ‘see through’ typos may misrecognize words
By
–
It does it even if you start with “monologue” etc. where there’s no real word with a q instead. I think that’s a plausible theory why it happens though — it’s trained to “see through” some apparent typos, like “How how many q’s in plague?”, and the recognition isn’t perfect
-
OpenRouter sending Llama 3.1 instruct prompts
By
–
It's OpenRouter — AFAIK it's just naively sending the base model the same prompt format used for Llama 3.1 instruct. Other than that, that's the whole dialog; no system prompt.
-
Llama 3.1 base — AI categories and topics
By
–
Also see my previous thread demonstrating Llama 3.1 base:
-
Unabridged Llama 3.1 base outputs show psychosis-like spontaneity
By
–
"Hey what's up." — An unabridged, minimally prompted conversation with Llama 3.1 405B base bf16. Responses exhibit naturalistic, emotional prose and psychosis-like spontaneity, both common in under-prompted base Llama and vividly unlike the output of widely used chat LLMs.
-
Grok silently purging stopped responses causes confusing conversations
By
–
Silently purging stopped responses from Grok’s context is confusing — it leads to conversations like this when you stop it because it’s misunderstanding you and you try to clarify:
-
Recommend JuliusAI for interactive Python plotting
By
–
If there’s a Python/R plotting package that does vaguely what you want try @juliusai — basically a fancier ChatGPT code interpreter where you can install arbitrary packages and it reads the docs so you can just ask for stuff. Handy for plots.