Improving the factual accuracy of language models through web browsing

openai.com

54 points·grappler··5 comments

The next version of copilot will literally be a bot that searches Google and reads through stack overflow posts. There are a lot of jokes you can make, but that is honestly what I do half the day and I've met some people who could barely do that.

First they came for the junior dev, but I was not concerned for I was not a junior dev.

I like the note that there is a new risk since their model can access the web at training time. A model in this situation could potentially do something malicious online!

"Don't worry, we'll just make sure it can only do GET commands, which are, per the HTTP spec, side-effect free and idempotent."

[later]

"Oh no."

---

Ha-ha-only-serious joking aside, this is impressive. Early on in the API release, someone showed you could sort of make GPT-3 operate on text representations to put inside a web loop, but it made a lot of errors. This goes well beyond 'sort of'. It may not seem like 'predict the next word' gets you far, but turns out, it's a good starting point. ("Tool AIs want to be agent AIs.")

I was wondering whether it obeys robots.txt? I guess since it's using Bing, it obeys as well as Bing does?

And then CoPilot has been let loose injecting underhanded backdoor code in the guise of utility functions.

Reply on news.ycombinator.com