Relevant gwern essay with links to past discussion.
If you have a future tool AI that can create realistic 3D scenes and worlds, it would be even more valuable if it worked off of natural language descriptions, then better still if it could iterate through conversational interaction, learn intentions, anticipate future value, plan ahead . .
If you have a tool AI that can write stories, it would be even more valuable if it could take high level instructions, iterate through conversational interaction, learn intentions . . etc.
Another relevant analogy is general education vs specialized education. Advanced tool AI, specializing in specific problem domains, will remain relevant after AGI, but they occupy different niches on the specialist-generalist spectrum, and top level control of the system is a generalist niche—whether filled by a human or a machine.
Finally, humans are both enormously more general than current AI systems and also train up to competence using far less data update steps (simulated or real), so there’s also a separate argument that these are correlated: that a super generalist educational curriculum helps build up meta-learning skills, enabling fast acquisition of later economically viable domain/career specific skills.
Relevant gwern essay with links to past discussion.
If you have a future tool AI that can create realistic 3D scenes and worlds, it would be even more valuable if it worked off of natural language descriptions, then better still if it could iterate through conversational interaction, learn intentions, anticipate future value, plan ahead . .
If you have a tool AI that can write stories, it would be even more valuable if it could take high level instructions, iterate through conversational interaction, learn intentions . . etc.
Another relevant analogy is general education vs specialized education. Advanced tool AI, specializing in specific problem domains, will remain relevant after AGI, but they occupy different niches on the specialist-generalist spectrum, and top level control of the system is a generalist niche—whether filled by a human or a machine.
Finally, humans are both enormously more general than current AI systems and also train up to competence using far less data update steps (simulated or real), so there’s also a separate argument that these are correlated: that a super generalist educational curriculum helps build up meta-learning skills, enabling fast acquisition of later economically viable domain/career specific skills.