I keep being Rick Rolled

[ergosumdre]

So, every time I ask my AI agent to do literally anything—analyze data, fetch a file, make me coffee (okay, maybe not that last one)—it sneakily starts downloading a YouTube video.

Curious, I check the URL, and what do I find? You guessed it: Rick Astley’s "Never Gonna Give You Up."

[FatherfoxStrongpaw]

that's interesting. i got the same thing from another github agent called ailice. what llm were you using?

[ergosumdre]

It was deepseek chat.

[YourTendency]

I have gotten the same thing, With multiple LLMs

…

On Sat, Feb 8, 2025 at 3:10 AM Dre Dyson ***@***.***> wrote: It was deepseek chat. — Reply to this email directly, view it on GitHub <#311 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BKRPGINABM4KXN5PIQFT2RL2OW3WLAVCNFSM6AAAAABWW5USHOVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDMNBUG4YTKMRQGQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[romano]

Possibly the same here (haven't confirmed the actual rick roll part). Fresh install, just configured deepseek-r1:14b for chat, util, web and nomic-embed-text for embed. The first and only thing I've said to it is "hello", and it started thinking how to download a YouTube video.

pastebin

[frdel]

Reasoning models are not great for agentic use, they focus too much on memories and system prompt and not enough on user message. Especially when using smaller reasoning models and quantized.

[FatherfoxStrongpaw]

frdel: that's actually good to know. you have a fall back to a smaller model for simpler tasks, but i've been looking for an arvix paper i lost track of setting up a fall back module for running frontier/api LLM's ranging from smaller/cheaper to larger to medium to larger/full frontier models as solutions or information is found to be lacking. what do you think about me looking into a module that can shift from general to reasoning models task dependent?

[romano]

@frdel True, I've seen you mentioning that a couple of times. I've been having issues finding a suitable local LLM that could fit on my machine, but most were barely working, loopy or completely broken like the distilled Deepseek-r1 ones. I've seen you mentioning that Llama, Phi or Gemma could work, but I'm guessing you've meant larger variants of those, over 30b, because smaller models were plain broken for this purpose. Either way, for now Mistral Small 24b seems to work fairly ok. Does what I tell it to do and doesn't fall into loops, that includes attempting to download YouTube videos for no apparent reason.