@sama "We have found a new thing" is a strong teaser the night before DevDay. Whatever it is, I hope the safety and permissions story ships on day one alongside the capabilities
Autonomy should be earned the same way FSD earns it. You start narrow, measure every time it has to step in, and widen access only when the data says it's safe.
Build it like a guardian angel, not like an autopilot you can't override.
Good guardrails stay invisible 99% of the time.
An agent that asks permission for every step gets switched off. An agent that never asks will eventually do something you can't undo. The skill is knowing which 1% of moments call for stepping in.
The scariest failure isn't an agent going rogue. It's an agent being helpful at the wrong time, like replying to an email you never meant to answer or "fixing" something you never asked it to touch.
Being helpful without having scope is the real risk.
Today a Tesla took over for about a second to keep a driver off a curb, and then handed control right back.
That's the best safety design pattern I know of, and it's exactly how AI agents should work π§΅
@aelluswamy A system that steps in only at the moment of risk, then hands control back, is the right way to design safety. It's the same idea behind good AI agent guardrails. It stays out of the way until it really matters
@cb_doge V3 is the first Starlink satellite built for Starship. It's too big for Falcon 9, so today's deploy was the actual start of the V3 network, not just a demo
@Scobleizer@NinjaTechAI Long running agents on your own servers only work if the company owns the logs and has a hard kill switch. Controlling the servers without being able to stop the agent isn't real control