{
"$type": "site.standard.document",
"bskyPostRef": {
"cid": "bafyreiem2a7ar4cdxsrim4s2nilni4wqmxhzwfjmhndybe2fujxyk2veae",
"uri": "at://did:plc:lk3jfj3zq4k4wxnk474axylu/app.bsky.feed.post/3mn3vijlovba2"
},
"path": "/t/chatgpt-pedantacy-problems/1382101#post_1",
"publishedAt": "2026-05-30T20:28:11.000Z",
"site": "https://community.openai.com",
"textContent": "I don’t know if its just me, but ChatGPT is overly pedantic.\n\nPrompt:\n\nQwen 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\n\nChatGPT Response:\n\nThat result is interesting, but I’d be careful about jumping from:\n\n> “This small model behaved strangely under pathological input”\n\nto:\n\n> “Qwen under 4B is notoriously unstable.”\n\nA few reasons:\n\n * **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.\n\n * **Echoing / copying user text** is a common failure mode in small models under repetitive or overloaded contexts.\n\n * 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.\n\n\n\n\nWhat you described actually gives you a potentially useful distinction:\n\n**Gemini Flash behavior:**\n→ repetitive degradation / weird continuations / recovery\n\n**Small Qwen behavior:**\n→ meta-commentary + copying behavior\n\nThose are different failure signatures.\n\nAlso, 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.\n\nFrom 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.\n\nIt took what I said and framed it in a way where it implies that I am wrong.",
"title": "ChatGPT Pedantacy Problems"
}