"But more recently, I have become a believer in this misalignment problem... and that is one of the things that has me super concerned."
Takeaway:"Including a lot of information just in general about your task is often very helpful."
Takeaway:"Ensembling techniques will take a problem, and then you'll have like multiple different prompts that go and solve the exact same problem... and then you take the answer that comes back most commonly."
Takeaway:"It is not a solvable problem... It has to be the AI research labs... It has to be innovations in model architectures."
Takeaway:"The most common technique by far that is used to try to prevent prompt injection is improving your prompt... This does not work, this does not work at all."
Takeaway:"Getting AI's to do or say bad things... It is not a solvable problem. That's one of the things that makes it so different from classical security."
Takeaway:"Generally not so useful anymore, because as you just said, there's these reasoning models that have come out, and they by default, do that reasoning."
Takeaway:"But from my perspective, roles do not help with any accuracy-based tasks whatsoever."
Takeaway:"If there were one technique that I could recommend people, it is few shot prompting, which is just giving the AI examples of what you want it to do."
Takeaway:"People will kind of always be saying it's dead or it's going to be dead with the next model version, but then it comes out and it's not."
Takeaway: