Загрузка видео...
Не удалось загрузить видео
Two risks with AI: 1: the system slips its guardrails (hack a grocery store to get you the current banana price) 2: a person with bad intentions uses it as a tool to do harm more efficiently.
14,078 просмотров • 4 дней назад •via X (Twitter)
Комментарии: 8

Just two risks with AI? Hello? There are far more than two risks with AI! ⤵️ 🆓

Side by side in my feed. I have pretty good idea what AI will be used for....

Useful distinction. Capability risk and malicious-use risk require different defenses: better containment and permissions for the first, identity, monitoring and accountability for the second. Treating them as one problem usually produces weak policy for both.

Off topic.Wait til med. treatment errors.>Rx or diagnosis incorrect&Ins. won't pay&causes grief to resolve&further erodes health.Or going to court to dispute errors whether health or other finance issues.Theil want a robot inserting foley cath,if needed one?Well prob.&not normal!

Build two small test sets: harmless tasks that tempt an agent to overreach, and explicit harmful requests. Track boundary violations alongside completion. Each failure is a case to investigate before changing tool permissions or refusal rules.

AI will monetize competitive optimization. Models accelerate the race between lawyers and courts, hospitals and insurers, attackers and defenders. Vendors sell to both sides. They make the race expensive, and charge society for the fuel, the track and the privilege of watching.

It upends the human labour market and mass unemployment starts the biggest petty crime wave in history.

So you’re saying my next grocery run could be a covert ops mission, and the bad guys will just order the chaos on autopilot? Nice. 🙏

