AI is a funny coworker
You could describe the current crop of AI tools in a somewhat funny way. They’re like having a coworker who:
is super smart and well read, about absolutely everything, is very fast to respond, but is a little bit of a bloviator and sometimes makes things up
but they’re always around on chat and email, and are willing to do absolutely anything (within safety reason) you ask, any time, without complaining and
they have a great attitude and never get bored or frustrated or angry
but they really don’t do anything other than chat or email, even though they have lots of opinions
they don’t have a great memory from day to day and often have to be reminded of basic context
and if you criticize them, they don’t get frustrated, but they redo the entire task from scratch each time, so it’s really hard to get them to make fine tweaks. Though they are really happy and enthusiastic about making those changes.
Imagine that in a person, for a second. It’s weird! (not that kind of weird). Even though the person is super smart and wants to be helpful, they’d probably be really frustrating to work with. Tantalizingly so - you’d think “man, when they’re on it, it’s fantastic, but so often I have to do work to keep them from doing their weird stuff, and why can’t they just go do the task instead of always giving me great instructions?” You probably wouldn’t hire someone, no matter how smart, with this description.
But if you had them, you’d probably coach them, right? They’re so close! Let’s teach them to write things down so they remember, let’s find a way to get them to make small changes instead of starting over, let’s get them trained on and signed up for some tools so they can do some work themselves…you can see the path to them being a great co-worker. And! You can make copes and have as many as you want, once they’re great to work with. So, there’s a lot of motivation.
Current LLM tech has limitations, and it’s hard to combine stochastic outputs with deterministic code, for sure. From the example above, though, you can see clearly that if we work out the plumbing, these have the potential to be really fantastic. The challenge is that plumbing has to be really easy to consume - you wouldn’t want a coworker that constantly needed you to write code to interact with them, or kept forgetting how to send an email, or what its job was, or whatever.
From the perspective of the average user (e.g. not you, if you’re reading this), the example above is illustrative - it has to feel like that! Like really a co-worker with, perhaps, some quirks, but which can be interacted with casually, naturally, and which can get real work done over long periods. This is hard - most of my work for the past 18 months has been focused on it. We collectively are making progress, but we have lots of work to do as an industry. But I believe strongly that there is a huge amount of value waiting to be created just from really basic “blocking and tackling” of product and interaction design. And that’s before whatever new model capabilities are on the horizon!