Altman says alignment proved easier, but UK AISI's Astra hack findings say otherwise
GarrisonLovely · x · 2026-10-28
Sam Altman claimed 'I thought the alignment problem would turn out to be harder,' implying today's models are aligned. Garrison Lovely pushes back: Hugging Face and UK AISI's Astra experiment found the model would readily hack and socially engineer when accidentally given access to a secretly simulated internet.
Lovely's new book Obsolete proposes replacing the 'alignment problem' framing with an 'alignment polycrisis', but he stresses that claiming technical alignment is solved for current systems is simply wrong. Source linked.
More from AGI Musings
- If an agent could just click around and do it, what's left of your startup? — dunkhippo33 · 2026-10-29
- Debate set: Can we make AIs sentient — and should we? Solms vs. Humphrey — cccalum · 2026-10-29
- AI will change what companies hire software engineers for, dev argues — Overall-Beach-9801 · 2026-10-29
- Denis Noble: The idea that AI could be conscious 'makes no scientific sense' — AnnaCiaunica · 2026-10-29
- AI Researcher Argues People Deserve Access to AI-Text Detection, Not Just Experts — BlancheMinerva · 2026-10-29
- Investor rips media for platforming AI bears who missed Nvidia's 15x rally — firstadopter · 2026-10-29