🧵 View Thread
🧵 Thread (7 tweets)
If you fuzz a Reinforcement Learning agent with random input 24/7, eventually it will BREAK and self-destruct. Dead neurons, adversarial attacks, catastrophic malfunctions. This is why our brains come with a tempering device built in. https://t.co/K802eM4R5i
We distance, decouple, and isolate ourselves from what we perceive in the environment, because /perception is dangerous/ Every service endpoint you expose can get hacked, buffer overflow, injected. Perception is neither neutral, or passive.
May AI safety research teach us more about our own minds, and may we gather more respect for the already-enlightened nature of our mind running countless programs to run the ship. https://t.co/IlxJsTdqPX
As Quine often said, paraphrasing Neurath, we are like sailors who must rebuild their ship while keeping it afloat. https://t.co/J9bDeIdocv
The “must” here is tricky. The captain issues his order, to try this new ideology, use this new engine, and is shocked, SHOCKED, when the sailors don’t comply. How could they? They’re running the ship. They won’t hot-swap an engine out and potentially die. https://t.co/hpgYhWdI1V
@meekaale I’ve heard the horse metaphor (@kcorazo does it well), and honestly, I’m moving away from it . . . more like a conductor 🥁 we are the blind conductor, not knowing if he is facing audience or the musicians, waving his hands thinking *he is playing the music* ha
Your system will refuse to trust something it does not trust. It’s almost tautological to say this. Yet we get confused all the time, “If I can’t articulate any reasons to not trust this, and there are reasons to trust it, I should (be able to) trust it.” This is confused. https://t.co/iR3cLzDCOq