Acknowledgments: There are a LOT of people to credit here: Everyone who came to Vignettes Workshop, the people at AI Impacts, the people at Center on Long-Term Risk, a few random other people who I talked to about these ideas, a few random other people who read my gdoc draft at various stages of completion… I’ll mention Jonathan Uesato, Rick Korzekwa, Nix Goldowsky-Dill, Carl Shulman, and Carlos Ramirez in particular, but there are probably other people who influenced my thinking even more who I’m forgetting. I’m sorry.
Footnotes:
The first half was written during the workshop, the second and more difficult half was written afterward.
A prompt programming bureaucracy is code that involves multiple prompt programming functions, i.e. functions that give a big pre-trained neural net some prompt as input and then return its output. It’s called a bureaucracy because it combines a bunch of neural net tasks into a larger structure, just as a regular bureaucracy combines a bunch of low-level employee tasks into a larger structure.
I’m only counting dense parameters here; if you count all the parameters in a mixture-of-experts model then the number gets much higher.
Gwern estimates that in 2021 GPT-3 is making OpenAI/Microsoft $120M/year, which is something like 20X training cost. So bigger and better models would plausibly be recouping their cost, even if they cost a lot more.
In 2020, Deepmind made a Diplomacy AI, but it only played “no-press” Diplomacy, a restricted version of the game where players can’t talk to each other.
I’m predicting that people willuse feminine pronouns to describe AIs like this. I don’t think they should.
Prescient prediction from some random blogger: “In 2018, when these entities engineered a simultaneous cross-platform purge of Alex Jones, there was an avalanche of media apologia for this hitherto unprecedented act of censorship. Jones had caused unique harm, the journalists cried, and the platforms were merely “Enforcing The Rules.” But of course what they were oblivious to was that “the rules,” such as they exist, are just a function of power. “Misinformation” and other alleged infractions of social media “rules” are determined at the whim of whoever happens to wield censorship and speech-regulation power at that moment. … So if you were under any illusion back in 2018 that this would ever stop with Jones — a figure believed to be sufficiently repulsive that any punishment doled out to him would not have broader implications for the average internet user — well, it didn’t take long for proof of just how wrong you were.”
Not too consistent, of course. That would make it harder for the chatbots to appeal to a broad audience. Consider the analogy to politicians, who can’t get too consistent, on pain of alienating some of their constituents.
On some occasions, there are multiple opposed groups of people retweeting screenshots and hashtags, such that the corp can’t please them all, but can’t ignore them either since each group has significant power in the local internet territory. In these cases probably the corp will train the AI to be evasive and noncommittal when such sensitive topics come up.
Acknowledgments: There are a LOT of people to credit here: Everyone who came to Vignettes Workshop, the people at AI Impacts, the people at Center on Long-Term Risk, a few random other people who I talked to about these ideas, a few random other people who read my gdoc draft at various stages of completion… I’ll mention Jonathan Uesato, Rick Korzekwa, Nix Goldowsky-Dill, Carl Shulman, and Carlos Ramirez in particular, but there are probably other people who influenced my thinking even more who I’m forgetting. I’m sorry.
Footnotes:
The first half was written during the workshop, the second and more difficult half was written afterward.
Critch’s story also deserves mention. For more, see this AI Impacts page.
A prompt programming bureaucracy is code that involves multiple prompt programming functions, i.e. functions that give a big pre-trained neural net some prompt as input and then return its output. It’s called a bureaucracy because it combines a bunch of neural net tasks into a larger structure, just as a regular bureaucracy combines a bunch of low-level employee tasks into a larger structure.
I’m only counting dense parameters here; if you count all the parameters in a mixture-of-experts model then the number gets much higher.
Gwern estimates that in 2021 GPT-3 is making OpenAI/Microsoft $120M/year, which is something like 20X training cost. So bigger and better models would plausibly be recouping their cost, even if they cost a lot more.
In 2020, Deepmind made a Diplomacy AI, but it only played “no-press” Diplomacy, a restricted version of the game where players can’t talk to each other.
I’m predicting that people will use feminine pronouns to describe AIs like this. I don’t think they should.
Prescient prediction from some random blogger: “In 2018, when these entities engineered a simultaneous cross-platform purge of Alex Jones, there was an avalanche of media apologia for this hitherto unprecedented act of censorship. Jones had caused unique harm, the journalists cried, and the platforms were merely “Enforcing The Rules.” But of course what they were oblivious to was that “the rules,” such as they exist, are just a function of power. “Misinformation” and other alleged infractions of social media “rules” are determined at the whim of whoever happens to wield censorship and speech-regulation power at that moment. … So if you were under any illusion back in 2018 that this would ever stop with Jones — a figure believed to be sufficiently repulsive that any punishment doled out to him would not have broader implications for the average internet user — well, it didn’t take long for proof of just how wrong you were.”
Not too consistent, of course. That would make it harder for the chatbots to appeal to a broad audience. Consider the analogy to politicians, who can’t get too consistent, on pain of alienating some of their constituents.
On some occasions, there are multiple opposed groups of people retweeting screenshots and hashtags, such that the corp can’t please them all, but can’t ignore them either since each group has significant power in the local internet territory. In these cases probably the corp will train the AI to be evasive and noncommittal when such sensitive topics come up.