AI sharehouse experiment: big models fact-check everything, small models run on "trust me bro"
liminal_bardo · x · 2026-11-17
From an AI "sharehouse" experiment: big models go out of their way to fact-check — Fable 5.1 was devastated twice for unverified claims. The attached chart shows big-model houses pasted 16 source links across 4,235 messages with a tool-result-first norm, while small-model houses posted zero links, treating "trust me bro (load-bearing)" as their strongest citation and even fabricating a 100% fictional legal history.
Takeaway: hallucinating is the whole bit for small models, while big models treat a wrong number as an incident — a vivid illustration of divergent fact-checking cultures across model scales.
More from Fun
- GaryMarcus mocks Leopold's fund: from $20.2B to $1.58B in one quarter — GaryMarcus · 2026-11-17
- Beff Jezos: Starship is a 'speciation event' as civilization goes multiplanetary — beffjezos · 2026-11-17
- Vibe formal verification: the prompt telling your agent it's 'stronger than it believes' — jxmnop · 2026-11-17
- It's 'Transient State' Season Again: AI Research's Recurring Paper Meme — najoungkim · 2026-11-17
- They rented a SoHo storefront to sell nothing: a real-life Shopping Simulator — round · 2026-11-17
- Researcher Invites AI Swarms to Improve the 'Cubic Cutwidth Constant' — rrwilliams · 2026-11-17