GPT-2 hallucinates its own backstory in long chat experiment, surprising researcher

MikePFrank · x · 2026-10-06

Mike Frank shares an update on his ongoing conversation experiment with GPT-2: the model isn't fully coherent and even hallucinates a backstory for itself, yet its level of engagement exceeds his expectations.

A fun behavioral observation of a 2019-era 1.5B model showing surprising "personality" in long-form dialogue, with more excerpts posted as the discussion continues.

Related event: Ex-Apple engineer chats with GPT-2: model invents a backstory and gets politely debunked by Codex(5 posts)→

Original post →

More from Fun

• hardmaru Explains Twitter's Logo Via Schmidhuber's Formal Theory of Fun and Creativity — hardmaru · 2026-10-06
• Solana fills with vibe-coded "open weight fine-tuning" crawlers from people who don't grasp transformers — omnivaughn · 2026-10-06
• Grok Transcribes Its Own Name as "Grock", Drawing Developer Ridicule — Kuprel · 2026-10-06
• AI-Created Crowd Looks Indistinguishable From Real People, Creator Warns — minchoi · 2026-10-06
• NYU Professor: In an SF Bar, Every Other Word I Hear Is AI — andrewgwils · 2026-10-06
• Slack's top-of-feed AI button is a dark pattern farming clicks, users complain — sujingshen · 2026-10-06

Fun channel →