I used to think this, but the AI labs sure seem neck-and-neck in the model race. Doesn't appear that anyone is developing an enormous lead. So I've become skeptical of the runaway king-of-the-world-maker model scenario.

The open models seeming to be ~6 months behind is very encouraging, too.

▲

mrob 7 hours ago | parent [-]

AI progress can potentially be extremely non-linear because of feedback effects. The first to build an AI smart enough to accelerate building even smarter AIs wins (or loses along with everybody else if it's more successful than they expected).

▲

Analemma_ 6 hours ago | parent [-]

People have said this, but so far if anything the opposite has been empirically true. OpenAI had a huge lead and it just didn't matter, Anthropic and Google both caught them and now they're neck and neck. It seems like compute overhang forecloses the possibility of runaway progress which eliminates all your competitors.

	▲	mrob 6 hours ago \| parent [-]
		Any feedback process has a hard threshold for instability. The PA system doesn't howl until the microphone is close enough to the loudspeaker. The atomic bomb doesn't explode until the fissile material reaches critical mass. If you don't know where the threshold is you can't extrapolate. Compute is a limiting factor now, but there have already been huge improvements in compute efficiency, e.g. mixture of experts. It seems extraordinarily unlikely that there are no more to be found. And compute capacity continues to increase too.