Useful until we're useless
People want AI to be perfect right now, and I think that's missing the point
Everywhere you go these days you hear people talking about AI. ChatGPT has been the topic of conversation on Joe Rogan and famous podcasts abound. Ryan Gosling made a commercial about it. Tesla’s (“soon to be”) self driving cars are, for better or for worse, constantly making headlines. And more recently, Microsoft announced their Bing search would even be releasing a chatbot. At first glance things are looking a little disastrous. People are literally clamoring for it to be shut down. But I think this is a little short sighted, here’s why:
Extremely Confident Bullshit
The main gripe seems to be that thing isn’t perfect. ChatGPT makes up sources. Tesla’s can barely handle intersections or bad weather. Bing can get extremely out of line with it’s users. And all these complaints are true and valid.
So we should simply stop expecting it to be perfect and stop acting outraged when it acts imperfectly. Instead we should focus on the value that an imperfect system can provide (which is immense), and take notice of the trends of AI’s improvement (also immense). We should be starting to ask: what might be enabled if it were 10 or 50 or 500% better?
I received this almost comically praising text about ChatGPT today. A few weeks ago I was thinking of what use cases might be good for this sort of thing right now, and guess what? There’s about a million. Content creation seemed well suited. I immediately thought of my friend who has a popular insta blogging account that has become her full time job. I tried to give her a few examples how to use it and today I got this. So even if ChatGPT never becomes a super intelligent doctor able to prescribe you medicine better than a human ever could or whatever, it’s already changing peoples lives. Text summarization, generation, improved natural language search, intelligent chat bots, video games, etc.. etc… are all areas that can be improved wildly just by adding some engineering and great user interface on top of the product today.
In driving’s case, my Tesla model 3 has completely revolutionized my driving experience, and I only have the version that keeps you in the lanes, no fancy intersections or anything. Why? Because you make thousands of micro decisions all the time when you’re driving and the decision fatigue wears on you. With autopilot I pretty much just have to keep track of what lane I want to be in and how fast I want to go. On a road trip or in stop and go traffic, I don’t get stressed or worn down. I come out of a drive feeling like I was more of a passenger. So if it never becomes a perfect self driving vehicle that can drive me to Vegas while I sleep, I’d still be happy to pay for the autopilot ( you don’t have to anymore it’s included).
Up, Up and Away
As for trends, it’s truly rare in this ol’ 21st century to see such explosive progress as we’ve been seeing in the AI world. More important scientific papers are coming out than even leading researchers can pay attention too. All the time the models are getting bigger, more efficient, more clever. Someone gave ChatGPT an IQ test and it scored an 83, which is definitely not a genius, but a lot of it’s mistakes were due to it being bad at math and numbers. And then like a month later OpenAI ships a release that makes it significantly better at math. Don’t blink.
And then we get to GPT-4, which is supposed to be a step change improvement over the ChatGPT model, rumored to be coming out somewhere around Q2 of this year. So what then will the people say if these coming models actually are perfect in many domains? Will a human even be needed for all this crazy thinking we do?
I guess what I’m saying is, when it comes to demanding perfection from these models: be careful what you wish for. Maybe they’ll just be useful until we’re useless.