Qwen models can't stop talking about sunk cost: 31/52 prompts on 27B vs 44/52 on 9B
adonis_singh · x · 2026-10-30
Developer adonissingh documents a bizarre behavioral quirk in Qwen models: asked about Napoleon, breakfast, six-packs, or engagement advice, the model keeps steering answers into "sunk cost fallacy" territory. He also tested 4-bit qwen-3.8-27b, which — built from repeated layers of the 9B — shows the same bias but weaker: the sunk-cost theme appeared in 31/52 prompts on 27B versus 44/52 on 9B. The same layer (L9 N3522 at -0.8 dose) reproduces it in 9B, while 27B's L25 N3522 needs a higher dose with milder effect.
Related event: Qwen3.5-9B appears to have a 'sunk cost neuron' that derails every answer(2 posts)→
More from Fun
- Spam bot calls with the recipient's own number spoofed as caller ID — msg · 2026-10-30
- Ajay Mehta jokes his lab's real frontier is Cute Little Guys as character multiverse expands — ajaymehta · 2026-10-30
- Geek designs custom Revolut card featuring a full Transformer architecture diagram — MrIvanAM · 2026-10-30
- 'The person you're talking to thinks you're a bot': a new awkward AI-era moment — geoffreyirving · 2026-10-30
- 1,656 commits in a month mocked as slop — CodeSandbox CEO jokes his own image is ruined — zeeg · 2026-10-30
- The TNG episode where Data argues AI can be human shaped later philosophy — Afinetheorem · 2026-10-30