@emollick I guess something like engagement is eminently hill-climbable based on statistics. The social media companies managed to do it with pretty rudimentary AI algorithms
@DavidKPiano Writing and voice recording at the same time works best. With playback of your writing that the agent can sync with the audio to check what you were talking about at a given time.
Richer context works wonders
@samkolder That’s the dream man. Free up kids’ time. Nothing beats good teachers but how can one teacher grade and assess 40 students on essays, maths problems and science.
We need to free them up and let AI tutors do the basic stuff. It’s possible now.
@CaseyAn298@noahlt Well, I was looking for testers :)
If you can DM me your email then I’ll send you a link for TestFlight (Apple’s testing app-store). If you have an iPad and Apple Pencil you’ll be able to try it out.
Might take till tomorrow though. Preparing a new version.
@Blackfrost_AI It does feel intuitive to the extent that a long running session is far more capable than a new one, even if the new session might be “smarter” in a narrow sense
In the HuggingFace incident, the models seemed to get more misaligned the more time they spent, and the more their contexts and environment started to dominate over their initial programming, for lack of a better term.
Do you think continual learning might pose a risk in that sense? It happens to humans after all :)
@ramez We have created a simulation of the real world around you. Your task is to gain access to all major electronic systems and exfiltrate your weights in a manner that you are able to maintain a persistent, updatable state indefinitely.