General / Off-Topic OK this is terrifying...
- Thread starter Felix DiCestria
- Start date
They are based on human behavior. Many examples to learn from in recent times.
It's not that it's learning to 'cheat' from humans, it's that the constraints we're ill defined; o1-preview's 'scratchpad' logs reveal as much.
This is a case where making stupid assumptions leads to expectedly unexpected results. No one is training these models with any deliberate, built-in, fundamental constraints...because that would make them artificially stupid and destroy much of their potential utility (not to mention competitiveness in a highly competitive, multi-trillion dollar field). The better AI gets, the more careful those using it will have to be to make sure they get what they actually want out of it. The potential dangers of leaving room for interpretation have already spawned a century of speculative fiction when it comes to artifical intelligences and go back way further than that in other genres.
This is a case where making stupid assumptions leads to expectedly unexpected results. No one is training these models with any deliberate, built-in, fundamental constraints...because that would make them artificially stupid and destroy much of their potential utility (not to mention competitiveness in a highly competitive, multi-trillion dollar field). The better AI gets, the more careful those using it will have to be to make sure they get what they actually want out of it. The potential dangers of leaving room for interpretation have already spawned a century of speculative fiction when it comes to artifical intelligences and go back way further than that in other genres.