Read on Twitter
design AI researcher here: design spaces are too big & full of competing priorities for this to work. human designers don't design by hill-climbing from random noise to an "optimal" point. AI will assist human designers & shift workload but not replace https://twitter.com/wilsonsaccount/status/1328011273909514243
historically, improvements in design AI ~always result in a need for better tools (to tell humans why the AI is doing that, give them a way to tell it *not* to do that, etc) & lead to more ambitious goals for the next project. workload expands to fill available human-time budget
humans "coach" ML-based AI by spending time hand-crafting inputs that lead to the desired kind of output. this is just… design. GPT-3 is remarkable for how *few* hand-crafted examples it needs & there's still a lot of skill involved in "prompt design" https://twitter.com/wilsonsaccount/status/1328065926336569345
GPT-3 output has a detectable "grain" to it that comes from the dataset it was trained on. this "grain" can be thought of like an instrument's timbre. to avoid outputs "sounding AI-produced" we'll need artisanal small-batch datasets, which – guess what? – people gotta make
anyway. i normally try to avoid QTing bad takes, but the misconception highlighted in the OP here is so common among silicon valley types that it really needs to be addressed every time it appears
