ChatGPT Pedantacy Problems

I don’t know if its just me, but ChatGPT is overly pedantic.

Prompt:

Qwen 3 models under 4B are notouriously unstable, right now Qwen 3.5 0.8B said The user sent a lot of text that starts with and then continued to repeat the exact thing I said, every single charachter

ChatGPT Response:

That result is interesting, but I’d be careful about jumping from:

“This small model behaved strangely under pathological input”

to:

“Qwen under 4B is notoriously unstable.”

A few reasons:

• 0.8B parameters is extremely small by modern standards. Small models generally have much less capacity for long-context robustness, compression, instruction retention, and recovery.

• Echoing / copying user text is a common failure mode in small models under repetitive or overloaded contexts.

• If it literally started saying things like “The user sent a lot of text that starts with…” and then copied the pattern, that sounds more like meta-pattern completion or context summarization failure than a unique architecture-specific bug.

What you described actually gives you a potentially useful distinction:

Gemini Flash behavior: → repetitive degradation / weird continuations / recovery

Small Qwen behavior: → meta-commentary + copying behavior

Those are different failure signatures.

Also, one correction: you mentioned “Qwen 3.5 0.8B.” Are you referring to a specific Qwen 3 0.8B release, or did you mean a Qwen 2.5 / Qwen 3 variant? Because the exact checkpoint matters if you’re comparing architectures.

From a research perspective, I’d be more surprised if a 0.8B model survived hundreds of thousands of repetitive tokens gracefully than if it failed.

It took what I said and framed it in a way where it implies that I am wrong.