This week my Twitter feed turned into one long GPT-3 demo reel and I have thoughts.
If you missed it: OpenAI's been quietly letting people into a private beta of the GPT-3 API since June, and this past week the invite list finally got wide enough that the good demos started showing up faster than I could keep track of. The one that actually stopped my scrolling was Sharif Shameem's thing, debuild, where you type a plain English sentence like "a button that turns red when clicked" and it spits out working JSX. Not pseudocode. Actual React. I watched the clip three times before I believed it wasn't faked.
And then it just kept going. Someone had it writing Excel formulas from a description of what you wanted the spreadsheet to do. Someone else hooked it up to answer trivia questions with weirdly confident wrong answers. There was a Figma plugin, a SQL generator, a thing that turned a mood into a color palette. By Sunday night it felt like every indie hacker with API access had shipped a "look what GPT-3 can do" tweet, and every tweet had a thousand quote-tweets of people saying some version of "the future is here."
I dont have access yet. I filled out the waitlist form back when the paper came out in May and I'm still just refreshing my email like an idiot, which, fine, I get it, they're rate limiting compute and dont want the thing generating spam at scale on day one. But its a strange kind of FOMO watching a technology demo unfold entirely through other people's screen recordings.
Here's my actual opinion, since apparently I have to have one: this is genuinely impressive and also a lot of what's going around is a parlor trick dressed up as a paradigm shift. GPT-3 is very good at pattern completion when the prompt does most of the work for it. The JSX demo is real and its clever, but watch enough of these clips back to back and you start noticing the demos are picked, not average. Nobody's tweeting the failed attempt where it generated three broken divs and a syntax error. That's fine, that's how demos work, but I'd rather wait a few months and see what people actually build with this once the novelty wears off than declare it's rewriting how software gets made. I said something almost identical about blockchain hype back in 2017 and I was right then too, for what it's worth.
Small unrelated thing that happened this same week and got buried under all the GPT-3 noise: the UAE launched its Hope probe toward Mars on Monday, from Tanegashima in Japan, after getting scrubbed twice by weather. First interplanetary mission from an Arab nation, and it barely cracked my timeline because everyone was too busy watching a language model write CSS. Feels like a weird trade honestly. A country standing up its first deep space program is a bigger deal than a chatbot that's good at boilerplate, but the chatbot is the one that's shareable in fifteen seconds.
If anyone reading this has API access and wants to let me poke at it for an afternoon, my email's in the sidebar, I've been asking around for two weeks now. I want to feed it some old posts from this blog and see what it thinks my "voice" is. Half joking. Mostly not.
One more thing before I close the laptop: I keep seeing people describe this as AI that understands language, and I think that phrasing is doing a lot of work it hasn't earned yet. It completes text in a way that's statistically convincing. Whether that's understanding or an extremely good impression of understanding isn't a question a demo video is going to settle, no matter how many retweets it gets. Ask me again in a year.