GLM 5.3 Flash hallucinates it's Claude-based, then claims it never invents answers
texasdude11 · reddit · 2026-10-09
While testing a local GLM 5.3 Flash setup, a user asked for its knowledge cutoff — and the model opened its chain of thought with "I'm jarvis-thinker... but based on Claude," inventing an identity out of nowhere with no such instruction present. It then guessed a cutoff date based on "Claude models typically have early 2025 cutoffs," acknowledging uncertainty but answering anyway. Its final answer even included: "I know what I don't know... I'll tell you instead of confidently inventing an answer" — seconds after hallucinating its own identity. The poster attributes it to polluted training traces and shares it as a laugh.
More from Fun
- Draft the best Pokemon offense and defense from the original 150 — billyjhowell · 2026-10-10
- The joke that didn't land: vibecoding was supposed to resurrect 15-year-old dead Kickstarter games — nptacek · 2026-10-10
- Insider reveals the hidden coordination behind viral tech launch tweets — andrew_n_carr · 2026-10-10
- Ex-AAA artist rebukes AI crowd mocking mathematicians: 'we went through this years ago' — repligate · 2026-10-10
- EPIC SAX CLAUDE: the classic sax guy meme gets an Anthropic remix — repligate · 2026-10-10
- goodside's AGI joke: if the Pope doesn't value your AI like a human, it's not AGI — goodside · 2026-10-10