Post

Qwen3-4B-Thinking (a 4GB download) just became the first model I've tested to directly push back against my pelican riding a bicycle test, calling it "oddly specific and completely unrealistic" and saying "this request violates physics and biology" simonwillison.net/2025/Aug/10/...

This model is pretty sassy, later in the thinking trace it said: Self-check: Am I being too pedantic? Nah—if someone asks for impossible things, it’s better to gently correct than make fake art that could confuse them.

I don't know, I don't find that type of output witty or illuminating in any way now. I'd encourage LLM builders and researchers to move away from the quirky/funny crap as soon as possible in every model they produce besides possibly a jokey model that people like to banter with.

The first Chinese model like this that I tried did its thinking in simplified Chinese! simonwillison.net/2024/Nov/27/...

I wonder if we're ready for LLM chains being used as 'embedded systems' that can do very complex things with no access to the internet whatsoever. I'm using incorrect terminology, by 'llm chains' I mean several stages of output verification being done by unique LLMs checking prev stage's work. >

We are getting close, the local models that fit on my 64GB Mac have improved dramatically in the last six months, and the tiny ones (like Qwen 3 4B) that fit in 8GB of RAM are getting surprisingly capable too