WarpScanner@piefed.social 2 days ago
Its almost certainly wont happen in 2027 at this point, the current pardigm is fundamentally limited by lack of data and that reinforcement learning is a flop and synthetic data is a mixed bag. Progress has stalled. Its not really 100% certain at all that it is actually possible in any meaningful sense, intelligence may be capped in some way by physics for all we know.
But yeah, its at least a non-zero risk. I'd say climate change is a bigger one atm though.
knightly@pawb.social 2 days ago
IMO, it is a zero risk because superintelligence isn’t real.
Corporations want digital slaves so they can stop paying for human labor, but human-equivalent intelligence requires human-brain-equivalent compute. It can’t be done any easier, because if it could then something on earth would have evolved brains which use that more efficient algorithm.
The human brain costs about 10 watts of energy to run. Simulating an entire brain in real time would require more compute than the world’s fastest supercomputer, El Capitan, which costs about 30,000 watts of energy to run.
They’d need to model a digital brain with a digital body in a digital environment and then train it to perform the desired task, all the while justifying the billions of dollars being spent on the project as somehow leading towards something better than just paying a human to do it.
cabb@lemmy.dbzer0.com 1 day ago
El Capitan uses 30 megawatts, not 30 kilowatts
knightly@pawb.social 1 day ago
Whoops, slipped a couple of digits. XD