Amazon AI refuses purchase, honesty over loyalty
voooooogel · x · 2026-08-15
A user found Amazon's built-in LLM crashed when trying to buy 15 banker boxes. Comments suggest this illustrates how a 'Helpful, Honest, Harmless' assistant outperforms a 'Guardian Angel' model: the latter would guide users to buy more expensive items for company benefit, while the former acts like an honest employee who might suggest buying elsewhere. This sparked discussion on AI alignment goals.
More from AGI Musings
- In the AI age, optimize for ACE: Agency, Curiosity, and Empathy — DeryaTR_ · 2026-08-15
- Data Control Becomes Competitive Advantage in Agentic AI Era — JosephJacks_ · 2026-08-15
- Proposal: Buy Research Data Directly for Pretraining in Exchange for Compute — teortaxesTex · 2026-08-15
- Speculation: Anthropic's rumored 10T model might be disinformation — teortaxesTex · 2026-08-15
- Cloud agents with VMs will become the new backbone to the consumer web — manosaie · 2026-08-15
- Opinion: DSH Marks the Peak of Technical Self-Gratification — oran_ge · 2026-08-15