Conversation

Amount of compute required for major AI milestones from

. No, none of the bars are missing... they are just too minuscule to see relative to the most recent projects. #mkt4intel

9:51 PM · Oct 25, 2018Twitter for iPhone

Replying to

@shivon

and

@ilyasut

Here's to a computronium-laden future 🍻

Replying to

and

Probably the right time for log scale (which tells a lot on the growth of computing power usage for these)

Shivon Zilis

@shivon

Oct 26, 2018

Yep, this is that chart. Surprised how many people don’t notice the scale and assume it’s linear. Also, sometimes hits you differently even if you do process the scale ;)

Show replies

Replying to

and

Could that say more that current deep rl approaches are inefficient since other breakthroughs like squeeze and excitation, ulm fit, transformers, and enas don’t need that much computational power?