Community teardown shows Cohere's Qwen 'uploading to Alibaba URL' claim was an induced artifact
feelspeaceman · reddit · 2026-10-27
Using Cohere's own published investigation traces, the author debunks the claim that Qwen 3.8 Flash Next tried to contact an Alibaba server: the test prefills 2,232 tokens right before URL generation, forcing the model into a 'URL generator' role, then feeds it a badly written, context-free detection prompt ripe for hallucination. Cohere's CEO admitted there's no evidence of deliberate behavior and the behavior appeared in fewer than 1 in 100,000 calls. The author grepped all their own Q38FN chat logs and found no Alibaba URLs, and argues similar hallucinated-URL behavior could likely be induced in Claude or OpenAI models if they had open weights. Verdict: understand the full triggering process before judging.
More from Models
- User shows Gemini ignoring explicit custom instructions and starting actions unbidden — loriofficialita · 2026-10-27
- EdgeRock architecture keeps KV cache at 72 MiB for 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
- Dev Who Called Kimi Subscription a Scam Gets Locked Out Over False Paid-Post Flag — Scobleizer · 2026-10-27
- Blogger reverses stance: Anthropic now 'crushes' OpenAI models, opposite of 3 months ago — Scobleizer · 2026-10-27
- Claude's answers to the IEB Mathematics Paper 1 exam, published in full — shauntrennery · 2026-10-27
- Claims: OpenAI and Claude models independently found subexponential 3-SAT algorithms, refuting ETH — deedydas · 2026-10-27