or How AI Will Really Kill Us
It’s obvious that once someone figures out how to close the loop on a model that can update it’s own priors with confidence based on reality, we are fucked.
Of course, that sort of fatalistic outlook is absent if you think you control the off switch, so why haven’t the doge bros done it? Is it really that hard?
I recall an article about a leaderboard competition for small open source models that was dominated by someone who isolated the layers that seemed to do the math thinking and then tweaked the model to run those layers multiple times. Results showed incredible progress. But spending more time thinking doesn’t change the underlying weights. Why isn’t someone letting the model do that?
Seems like they must have tried. Or more likely, they are still trying. Granted I don’t follow development as much as I could, but I certainly follow it more than the average bloke on the street. I’m just surprised I haven’t passively picked up more about the difficulties in this space.