You know what's a very useful tool for a very small group of people? 3d printers. I own two. Right now one of them is spooling out a cryptex for my daughter's birthday. The other one hasn't been used in a while because it's primarily useful for turning digital designs into wax for casting in metal. The cheez whiz guy was mostly bought for fixturing but as soon as I had it I discovered that I could make entire cable runs and solenoid fixtures and sensor adaptors and stuff with it? Boy howdy I now have opinions about filament. The goop guy was bought entirely for casting. That's not how most people use them, though. Cheez whiz printers are mostly used for making tacky crap your wife could buy off of Amazon for half the price. Goop printers are mostly used for making orcs with tits to use in your Warhammer campaign. I can't think of any of my friends, colleagues or associates to whom I would say "buy a 3d printer."

how much did you need a natural language processor anyway?

I've been researching bringing an AI onto our website. We need it for scheduling. There's this whole iterative goal-seeking mess whereby the clinician has X subset of appointments available and the patient has Y subset of availability for appointments. Finding the X-Y Venn diagram sweet spot is an iterative process whereby our receptionists play Battleship over the phone. AI sits there and does the mindless job of matching schedules, thereby freeing up my receptionists for problems that actually require intelligence.

Do I want AI to do all our scheduling? Hell nah, not even if it could. I want it to do the tedious goal-seeking BS.

I saw a thing on Twitter a couple-three days ago where a programmer wrote a piece of gmail middleware for a developmentally-disabled friend with a landscaping business. The friend types some halting language in response to an email, that email passes through GPT-3, the client sees formal, gramatically-correct business language. The guy with the disability has no problems mowing lawns and pruning trees but his written language skills have been an impediment. Likewise, if English is not your first language and you need to do a lot of conversing, models such as this go a long way towards leveling the playing field.

Look at it this way: this technology allows you to create a context-sensitive instruction manual. "I have six screws left and the door is wonky what do I do." It allows you to half-ass recipes: "I have eggs, milk and flour and want pancakes, what else do I need and in what proportions." It allows you to freeball your layover: "I have eight hours at Logan International what do most people do for fun on a Tuesday in Boston." Will it do any of this stuff perfectly? Hell nah. Will it do it better than asking random strangers? Probably not. But it will do it better than not asking at all and with a barrier to entry this low, that's not nuthin'.

imagine a social media that's 80% propaganda bots powered by AI language models?

I doubt we need to imagine it. It's coming. Thing of it is, though, this isn't an AI problem it's a social media problem. All social media networks aren't profitable enough to pay for moderation, and the only remuneration available is through extremely low-value advertising. As such, social media reflects the incentives.

"Perfect Nazi garbage for everyone" changes the incentives, principally by diluting the value of the network and suppressing the value of advertising. If these models end up choking the arteries of Facebook and Twitter like a platter of bacon cheeseburgers I will dance a goddamn jig.

The content on social media from people you don't know is already valueless. It's already generic garbage. It's already useless. If it can be deprecated to the point where there's no reason to show it to others? We all win.