When Claude Code Skates on Thin Ice
I love Claude Code and Sonnet 4.5, but when training data is thin, the reasoning failures are hilarious
Frogs and Scientists
There’s a very old joke about scientists and frogs …
Vision Model
I’ve been on a deep-dive working with vision models over the past few weeks; specifically Qwen3 VL, which just launched the last week of October. We’re running inference on the model via OpenRouter providers. Claude Code has been my go-to AI Coding tool for this project.
Two consequences of working with an extremely new model like Qwen3 VL:
- The coding model (Sonnet 4.5 for me) will have literally no training data to help ground its reasoning
- There will be lots of complex problems to work through
”Frog with no legs is deaf”
The old frog joke came to mind multiple times over the weekend, when I had asked Claude to assist in diagnosing a problem. It would step through a valid chain of reasoning, and then at the very end, out would pop a ridiculous conclusion, like the joke’s “frog with no legs is deaf.”
The takeaway: When your model skating on thin ice—little or no context to ground its reasoning—be suspicious of its conclusions.
