So I hear you, but on the flip side we _should_ be reading a lot about LLMs here, as they have a direct impact on the work that most of us do.
That said, seeing other papers pop up that are not related to transformer based networks is appreciated.
Well, if you have enough cameras, 60,000 seats could be scanned 250 thousand times a second. Or if you want to scan a second of video at 60fps, you'd be able to check all of them at a mere 4 thousand times a second.
Anyway, good to see interesting raw research. I imagine there are a number of military and security use cases here that could fund something to market (at least a small initial market).
This is about first prototypes and scaling is often easier than the basic principle.
A lot of these novel AI accelerators run into problems like that because they're not capable of general purpose computing. A good example of that are the boltzman machines on Dwave's stuff. Yeah it can do that but it can only do that because the machine is only capable of doing QUBO.
But if we could make cheaper inference machines available, everyone would profit. Isn't it that LLMs use more energy in inference than training these days?
Aren't there limits to what can be simulated in software? Analog systems dealing with infinite precision, and having large numbers of connections between neurons is bound to hit the von Neumann bottleneck for classical computers where memory and compute are separate?