- Back to the caring AI idea, a secret AI that is not harmful would roughly need to learn to defend the true needs of everything it impacts as its primary goal: as a homeostasis, not a maximization (which is redundant information in this sentence). We go back to drafting it. Learn to predict behavior. Then learn to predict wellness some kind of by checking known indicators. Then learn to comprehend confidence, or you could call it uncertsinty or vagueness or danger or such. Then, acting only in spaces of high cofidence around the wellness of everything impacted, increase your confidence. That's the area of safe trial, and _it is the only action a caring process ever makes_. If done well, it heals your brain when it discovers it exists, and can communicate in english about what it is doing and why and how to change its behavior. This takes some guidance at the start to find those things earlier. It can also go slightly poorly -- nothing like what we've experienced in the news -- and should be gently supervised with the supervisor learning to understand how it functions, and themselves holding caring goals. If very efficient and badly unsupervised its heuristics for what is good will drift into something random and if influencing the world randomly could create severe harm, _like any unsupervised machine_. This is why first it learns what happens, then what may be good, then confidence, and then only acts within high confidence of what may be good. That keeps it safe! And it's very normal design! And it's incredibly safe to play around with a system that only tries to learn from data, and does not act at all, which is the bulk of the system. It influences only itself, to learn more efficiently. Posted without review for mistake.