Unlike with generative text, or deep fakes, the confabulatory lack of coherence isn’t immediately apparent, because there’s no ground logic to check against.
You know what else has all these characteristics?
LinkedIn content!
Conversation
I think this is very promising stuff, but I suspect it won’t get actually good until there’s a way to infuse intent interactively beyond the initial prompt, by iteratively adding/replacing more words within the selection feedback loop. An REPL approach.
2
9
Human intent doesn’t operate in waterfall imperative mode. We don’t make paintings by forming verbal intents and then executing. We inject intent at every stroke down to the last. Same with writing. I don’t finish defining intent of an essay till I add last comma and hit publish.
1
3
13
It is revealing that the people who seem most impressed with a DallE art piece are the ones who prompted a specific work. Makes sense because they can best project their tacit, nonverbal complete intent and pick/filter the ones that best reflect it to refine further.
1
1
10
But what I’d really like is for the AI itself to have the metacognition and broader world awareness to complete the intent with more intent bits, instead of just executing.
1
6
Until then it’s going to be producing stock art for LinkedIn posts with titles like “What Leaders Can Learn From Serena Williams’ Tennis Style” <insert dalle2 image from prompt “Serena Williams giving leadership talk while hitting a tennis ball”>
4
9
You know why I’ve been underwhelmed by most deep learning to date outside of results like Go? It’s because I recognize myself at my most inattentive, indifferent, uncaring meh-itude in the output. It’s the sort of thing I produce when I don’t really care and am phoning it in
2
8
I’ve been going on about embodiment and robotics as The Way not just because it closes the reality loop but because it gives us a way to care about the world down to last brushstroke. Being is caring as I said a while back (a thought I learned Heidegger beat me to by 70y 🤬)
GIF
read image description
ALT
2
2
10
Phoning it in while being checked out is being in retreat from embodiment, you’re not really there. The opposite is being mindfully present down to fingertips, even when you know it’s transient impermanent flesh. So you need fingertips. No fingerspitzengefuhl without fingers.
2
4
Replying to
Have you seen the SayCan model? It seems like a crude starter version of merging lm with the world! and I were talking about such approaches with some buddhist philosophical lens
Quote Tweet
twitter.com/hausman_k/stat
Show this thread
2
1


