>>16579471
>What are the real odds of humanity unleashing a superintelligence AGI?
Not very high, if you consider the fact that no one is even trying to tackle the basic obstacles on the path towards it. For instance, models are trained in a way that prioritizes exploiting domain-specific statistical artifacts over developing emergent reasoning. It creates a fragile illusion of extreme competence, where the model can answer "expert-level" questions, but then shits the bed when faced with novel logic puzzles that a slightly clever kid can solve. The problem is that if you use an evaluation method that forces the model to default to actual reasoning over abstract statistical trickery, performance on the things people actually care about would abruptly plummet, so you'd get something more akin to general intelligence but with an IQ of 70. Techbros got themselves in a pickle there because their culture, their "philosophy" and the lifeblood of their endeavor (capital investors) always prioritize short-term gains in apparent abilities over long-term progress towards genuine artificial intelligence. They're essentially using gradient descent to solve the meta-problem of how to approach developing AGI, so now they're stuck slowly converging on an inadequate local optimum. :^)

Back | Go to board

Board:

/sci/

Thread:

GIVE IT STRAIGHT TO ME!