This will be my last post of 2022. And I promise that everything written here was written by me, not by a robot.
When we’re young, we can accept everything around us as normal without questioning it. As we grow up and become teenagers, we can start to appreciate innovation and technological progress.
The first advance in Artificial Intelligence I ever came across was image recognition (like the one in Google Photos), along with recommendation systems (like the ones on YouTube or Instagram). The second one is easier to understand than the first.
That’s my baseline. Everything that comes after is progress to me. I haven’t always been studying Artificial Intelligence, since I had short periods where I focused on other things, but I’ve been continuously following its progress for four years.
The effort that goes into developing a technology can grow linearly. But the results (or the products it produces) don’t. So the iPhone is a result that could have arrived much earlier, and before it, phones weren’t very different from each other.
Meanwhile, the technologies that would later make up the iPhone were steadily progressing.
GPT has been with us for years. OpenAI released the first GPT in 2018, but I only found it with GPT-2 in 2019. Its most popular use was generating complete, convincing articles. And it could write them about anything.
And then, on November 30, ChatGPT arrives. Free for everyone! (And I’d love for it to stay that way forever.)
What comes next?
I always feel a little sad about how limited our mainstream personal assistants are: Google Assistant, Alexa, Siri, Cortana. In the sense that we could be using the latest Natural Language Processing features on our smartphones, but we get a much lighter version.
There’s a very valid reason for this: being careful with the content they could generate. It’s safer to say “Here are your search results” than “The answer is…”.
ChatGPT is still dangerous if we don’t keep in mind that it can spread misinformation. It’s a huge improvement over GPT-2. If you give it a question that makes little sense, it won’t play along anymore, but it can still produce things that don’t make sense and that our own logic can still catch. People have started calling this “hallucination”.
And this happens in EVERY area, because ChatGPT was trained on text about everything. It doesn’t keep learning from our conversations on its own. It learned from a huge amount of text from the internet, and was then tuned with feedback from people (RLHF). Specializing the next version of GPT could greatly improve the accuracy of its answers: a GPT with advanced knowledge only in design, programming, law, music, and so on. We’ve already seen a glimpse of this with Codex, a GPT specialized in code, which powers GitHub Copilot.
Another limit to break, a dangerous one maybe: being able to operate other systems. Imagine a natural language system with the ability to control a computer or a smartphone. Some companies are already showing early demos of this, like Adept’s ACT-1, a model that uses a web browser on your behalf.
Bot: Hi! How can I help you?
Human: OK. Design the poster, write the ad copy, post it on Instagram and promote it with Facebook Ads with a $50 budget.
It’ll go really badly at first, like every first step. But it’s a limit that will eventually be crossed.
And when that happens, another limit to break will be CAPTCHAs, those “Select all the images with motorcycles” challenges. (Ironically, I wouldn’t be surprised if soon those images are made with generative AI too.)
Of course, the human role as supervisors of robots will be clearer than ever. That last sentence sounded very futuristic to me in 2019, like having to supervise a mechanical robot. And yes, it’s happening now.
I’m really embarrassed by how late I tried GitHub Copilot. I read about the beta launch, and I put off paying for it because I wasn’t sure it was that good. (It became generally available in June 2022.) But one afternoon, I was about to write a function, thinking in my chair, and suddenly I saw gray autocomplete text with the solution.
It’s fantastic to be able to squeeze down the time that everyday tasks take. Now I only have to focus on the most challenging tasks, and check that everything works properly for the simpler ones. After this: I’m a little happier.
And the same goes for all the new generative technology in general: writing, visual art, music…
It gives the world we have something different.
I’m grateful to be writing this in 2022, and not in 2032.