After a frothy couple of weeks in which people have spent an awful lot of time talking about AI and What It Might Mean For Us without in any way spending time thinking about what, to my mind, is The Big Issue Here (to whit: "why are we building all these post-scarcity toys in an era in which scarcity is still very much a thing, and why are we not spending more time thinking about what this is going to do?") here's another piece of ammo for all the DOOMERS currently convinced that the LLMs are just a training cycle away from plotting our demise.
AutoGPT is – let's be clear – a very emergent thing right now, and it's more theoretical than practical, but it's…impressive in scope. The link here takes you to a Twitter thread talking about the principle and demonstrating how it functions in practice, but, simply put, "Auto-GPT is an open-source application showing the power of LLMs like GPT-4 to autonomously develop and manage different kinds of tasks like completing a code session or suggesting a business idea. You give an agent an identity, role/task, goals, and specifics about what to accomplish and it attempts to achieve it in an 'autonomous' way through a framework designed to allow the model to 'reason and act' to accomplish the task." So what this does is basically give a window into the not-too-distant future in which you will theoretically be able to ask your AI assistant to, for example, research your preferences based on a corpus of material and then, depending on its analysis of said corpus, go off and book you a holiday that you will like (checking your calendar for your availability, obvs).
Is this good? It sounds convenient, of course, but then there are all the questions about the extent to which it's a good idea to allow black boxes to behave with this degree of operational autonomy…BUT DON'T WORRY! As I said, this is (probably) MONTHS away from being reality! If you'd like a practical, hands-on way of exploring how this stuff might work, here's a little toy that demonstrates the sorts of steps AutoGPT would go through to accomplish a broad task – oh, and of course someone has set this up with the goal to 'Destroy Humanity', so if you're interested you can track the progress of ChaosGPT at this Twitter feed (at the time of writing it has decided to attempt to use Twitter to manipulate people to do its bidding, which will make you laugh darkly if you've ever tried to use Twitter to change anyone's opinion about anything). Ought we be worried? NOT ABOUT THIS, BE WORRIED ABOUT THE JOBS AND THE MONEY FFS.
[via Web Curios]