Someone did recently "test" it by roleplaying through each step of what it'd take to self replicate - make an Azure account, pay someone on TaskRabbit to solve the captcha for it, setup a prepaid card to use for billing, creature the Azure instance, deploy code, etc.

It didn't do too well. But give it a year or two and I'm sure it'll do it flawlessly.

> How long until someone instruments it to compile code, spin up Azure instances, and run it?

If you just use react and let it run, say, Python, it is quite able to try to do things like import libraries and call external services without any coaching (I extend the simple ReAct implementation for the ChatGPT API that was posted here to add exec() as well as eval() support under a new action, and it made an attempt which failed because (1) it left a placeholder for an API key in, (2) the required library wasn’t actually available to it, and (3) there wouldn’t have been an API key for the service available even if it had recognized the need to remove the placeholder.

Even though there is work on pushing models forward, too, I don’t think the potential of existing models combined with existing tools via ReAct (and possibly other tool integration patterns, ReAct is just the one I’ve seen and tried) has been explored much at all.

Similar: There’s a line in the Wolfram blog post (the screenshot under “Audio”) they tell the AI

“Run it yourself”

And it does