Yeah. I mean I started reading that story and was thinking how cool it would be… Until it started going bad. Something like a GPS for whatever task you were doing at work would be cool.

You should check out the short story Manna. It’s maybe a bit dated now but explores what could go wrong with that sort of thing.

Ahh ok that makes sense. I think even with GPT4, it’s still going to be difficult for a non-programmer to use for anything that isn’t fairly trivial. I still have to use my knowledge of stuff to know the right things to ask. In Feb or Mar, you were using GPT3 (4 requires you to pay monthly). 3 is much worse at everything than 4.

I’m curious about this. What model were you using? A few people at my game dev company have said similar things about it not producing good code for unity and unreal. I haven’t seen that at all. I typically use GPT4 and Copilot. Sometimes the code has a logic flaw or something, but most of the time it works on the first try. I do at this point have a ton of experience working with LLMs so maybe it’s just a matter of prompting? When Copilot doesn’t read my mind (which tends to happen quite a bit), I just write a comment with what I want it to do and sometimes I have to start writing the first line of code but it usually catches on and does what I ask. I rarely run into a problem that is too hairy for GPT4, but it does happen.

In the case I mentioned, it was just a poorly aligned LLM. The ones from OpenAI would almost definitely not do that. That’s because they go through a process called RLHF where those sorts of negative responses get trained out of them for the most part. Of course there’s still stuff that will get through, but unless you are really trying to get it to say something bad, it’s unlikely to do something like in that article. That’s not to say they won’t say something accidentally harmful. They are really good at telling you things that sound extremely plausible but are actually false because they don’t really have any way of checking by default. I have to cross check the output of my system all the time for accuracy. I’ve spent a lot of time building in systems to make sure it’s accurate and it generally is on the important stuff. Tonight it did have an inaccuracy, but I sort of don’t blame it because the average person could have made the same mistake. I had it looking up contractors to work on a bathroom remodel (fake test task) and it googled for the phone number of the one I picked from its suggestions. Google proceeded to give a phone number in a big box with tiny text saying a different company’s name. Anyone not paying close attention (including my AI) would call that number instead. It wasn’t an ad or anything, just somehow this company came up in the little info box any time you searched for the other company.

Anyway, as to your question, they’re actually pretty good at knowing what’s harmful when they are trained with RLHF. Figuring out what’s missing to prevent them from saying false things is an open area of research right now, so in effect, nobody knows how to fix that yet.

That last bit already happened. An AI (allegedly) told a guy to commit suicide and he did. A big part of the problem is while GPT4 for instance knows all about all the things you just said and can probably do what you’re suggesting, nobody can guarantee it won’t get something horribly wrong at some point. Sort of like how self driving cars can handle like 95% of things correctly but that 5% of unexpected stuff that maybe takes some extra context that a human has and the car was never trained on is very hard to get past.

That’s possible now. I’ve been working on such a thing for a bit now and it can generally do all that, though I wouldn’t advise it to be used for therapy (or medical advice), but mostly for legal reasons rather than ability. When you create a new agent, you can tell it what type of personality you want. It doesn’t just respond to commands but also figures out what needs to be done and does it independently.