There’s probably something that I’m missing here, but:
Given that the dangerous AI capabilities are generally stated to emerge from general-purpose and agentic AI models, why don’t people try to shift AI investment into narrower AI systems? Or try to specifically regulate those systems?
Possible reasons:
This is harder than it sounds
General-purpose and agentic systems are inevitably going to outcompete other systems
People are trying to do this, and I just haven’t noticed, because I’m not really an AI person
General-purpose and agentic systems are inevitably going to outcompete other systems
There’s some of this: see this Gwern post for the classic argument.
People are trying to do this, and I just haven’t noticed
LLMs seem by default less agentic than the previous end-to-end RL paradigm. Maybe the rise of LLMs was an exercise in deliberate differential technological development. I’m not sure about this, it is personal speculation.
There’s probably something that I’m missing here, but:
Given that the dangerous AI capabilities are generally stated to emerge from general-purpose and agentic AI models, why don’t people try to shift AI investment into narrower AI systems? Or try to specifically regulate those systems?
Possible reasons:
This is harder than it sounds
General-purpose and agentic systems are inevitably going to outcompete other systems
People are trying to do this, and I just haven’t noticed, because I’m not really an AI person
Something else
Which is it?
There’s some of this: see this Gwern post for the classic argument.
LLMs seem by default less agentic than the previous end-to-end RL paradigm. Maybe the rise of LLMs was an exercise in deliberate differential technological development. I’m not sure about this, it is personal speculation.