Analog optical computing for sustainable AI and beyond

[@JohnVance]

If I were starting over today, I would get into analog computing and not look back.

[@ergo6450]

I'm a comp eng undergrad and really interested in this forgotten art of analog computers;

[@johnbolt2686]

Really appreciate this. I was just discussing optical computing for inference at a conference a few days ago.

[@evanpettijohn4049]

Great presentation!

[@Chaitanya315]

I wonder if they do some amount of redundant calculations, given that small vibrations might knock things off the optical path..

[@dwinantosaputra6679]

Thank you Microsoft.

[@emwave100]

"Make physics do the work." Physics always does the work.

[@riothecat]

Great presentation! Here’s a follow up question: would this optical approach also lend itself to LLM model training? Or is it more suited to inference?

[@GNARGNARHEAD]

I don't believe so, but, they seem to be much better suited to efficiently scaling the more traditional neural net designs.. Jürgen Schmidhuber talked about these over on Machine Learning Street Talk earlier in the week, if you're looking for a frame of reference.. exciting times 😁

[@matteofrattini9133]

I was asking myself the same question. If "training" means adjusting the polarity of filters, then I'm not sure it would be very efficient to physically manipulate millions and millions of components at each epoch. But then again maybe there is a way to do it electronically? I don't know enough on the engineering part

[@GNARGNARHEAD]

awesome!

[@Paroissien]

In a video with so much potential, beats me why you guys would be so neglectful with the subtitles.

[@user-jm6gp2qc8x]

maybe with a person with huge racks could attract more audience

[@magnetsec]

Those audiences are not needed here. Also big brain is hotter than big racks.

[@user-jm6gp2qc8x]

@@magnetsec "Turnaround"

[@user-jm6gp2qc8x]

@@magnetsec wrong. you must attract every kind of audience

[@matek1324]

@@user-jm6gp2qc8x false statement 'you must attract every kind of audience' - super wrong.

[@asaejapan7143]

@@magnetsec The non-native audience with lower knowledge in the field would have benefited from decent subtitles, not to mention the audience with a slight hearing impairment.