Thursday, August 4, 2022

Accidental Autosomething


Putting this here for posterity, of the time we accidentally figured out how to engineer happy accidents into our artificial intelligence research, and as a reminder that GPT, a program meant to understand language, has inadvertently kicked off the Singularity.

AI Is My Copilot
Apr 2022, WIRED Clive Thompson

By late 2020, developers had observed something unexpected about GPT-3. The AI wasn't just good at autocompleting sentences. It could also autocomplete computer code. 

[So it was discovered by accident.]

GPT was trained on those bazillion documents scraped off the web, a lot of them were pages on which nerds had posted their computer code. That mean the AI had learned patterns not just in English but also in HTML, Python, and countless other languages. ... 
[Then they went crazy] 
Zaremba let the code-writing AI use three times as much computer memory as GPT-3 got when analyzing text. 
[Remember that this is the same program that was withheld from the public out of fear that it would be too powerful to control if it got into the wrong hands.] ...Potential abuses of this program, such as "scaled abuse," where someone using the AI could author a cloud of Twitterbots to harass a politician or spread disinformation ...

[Then they put GPT on GitHub and let it rip.]

The new program was called Copilot, and was meant to help you write code, and developers developed their own "theory of mind" about how Copilot works, and they got better at communicating with it. [That's right; we're getting better at empathizing with computers.] "It's not something you're used to. It's not like a human theory of mind. It's like an alien artifact that came out of this massive optimization." -AI Guru Andrej Karpathy
On Intellectual Property and Fair Use: "Some weren't thrilled that code they had put on GitHub for other humans to use had been masticated by an AI and turned into a potentially lucrative product for its owner. They wondered about the legality too. Sure, they'd posted the code as open source, but does treating it as training data count as a fair use?"

Final Thoughts: "Copilot even seems to have picked up knowledge about specific fields. Maria Nattestad is a software engineer at Google and, on the side, the author of a popular app that makes eye-catching visualizations from bioinformatics data. She discovered that Copilot knows about DNA. ... 

Image credit: Megastructure by Beate Bachmann for Pixabay, 2020 [link]


No comments:

Post a Comment