François CholletVerified account

Somewhere connected. http://tonsky.me/blog/disenchantment/ …

We've only had a bit of time w/ GPU's. Perhaps lifting the compute constraint for convolutions has more to offer yet. The good prediction accuracy w/ GPU's has also increased hype/interest, which will get a lot more minds working on AI down the line.

Creativity comes from constraint

Progress on NN was stalled before GPU were powerful enough and enough data were available

Start posting about your ideas for new things to try!!

Some folks would wonder: "who moved my cheese?". Life is a string of events which make no sense and some folks would attempt to find one and build a new story around the new "challenge". Wether the story is either "fast" or "progress", is in the eye of the beholder. :)

Exactly, I had the same issue when I was doing research on detecting foreground objects in UHD videos. 99% of the state-of-the-art was still based on 320*240 videos. Needless to say that comparing to methods which were not even able to handle a 10s video was not easy.

Too much time is spent trying simple, minutely incremental ideas because "why not? I have access to a GPU". I agree with your statement. I have noticed this in my work too.

It's a very French way of thinking: enjoying creativity that comes from (often self-imposed) limitations. Love it! Inference and learning on edge devices is another natural constraint. I think great things are gonna come from the limit of power that devices consume.

Hear me out.. what if we remove all the computers?