Recently, the founder of Stability AI made an ominous prediction: there are no programmers in 5 years. Just 2 years ago, nobody was using AI code assistance, but 2 years after GitHub co-pilot beta was launched, we’re now here—41% of all code on GitHub right now is AI-generated. That’s a crazy claim, and I’d be highly skeptical of it, but I’d believe it next year after looking at these crazy new updates in GitHub co-pilot, the launch of Google’s duet AI, and the launch of JetBrains AI, all of which just happened in the last few days. It’s December 14th, 2023, and you’re watching the Code Report.
Before we get into this video, a disclaimer: the tools we’re about to look at may make you want to drop out of your computer science degree and switch your major to Plumbing, but that would be dumb. Because they’re also coming out with AI-powered smart toilets. But more importantly, pessimism leads to weakness; optimism leads to power. It’s possible programmers won’t exist in 5 years, but it’s also possible that programmers will get paid five times as much in 5 years because everyone got discouraged and quit computer science in 2023.
Okay, so now that we have that optimism out of the way, what we’re looking at today is the next generation of AI tooling. But as of today, that’s all it is— it’s not capable of replacing any human yet. The big announcement yesterday is that duet AI from Google is now generally available. It can be installed into your IDE and provides access to an AI chatbot. You can chat in a dedicated panel or get inline suggestions and intellisense, just like Hub co-pilot. But unlike co-pilot, it’s totally free, at least until next year, before it becomes $19 per month. Its real killer feature, though, is its integration with Google Cloud. It allows you to link a cloud project and then manage all the resources directly from the IDE, as well as access documentation and generate code samples for your project. It’s also worth noting that Google is working on its own Cloud IDE based on VS Code called Project IDX, which is not only integrated with this AI but also has things like IOS and Android emulators built in that run on a cloud VM.
Currently, its underlying AI model is not Gemini Ultra, which is claimed to be the best AI programmer out there. If that’s true, duet has the potential to be a GitHub co-pilot killer in the future. But speaking of co-pilot killers, JetBrains also just announced their new co-pilot killer in December. It is possible to use co-pilot in JetBrains IDEs, but the experience has never been that great. It just feels like it gets in the way more so than it does in VS Code. The AI assistant feels much more natural and well-integrated and can do things like chat, refactoring, write documentation, write unit tests, offer companionship, generate commit messages, and so on. But with a price tag of $10 per month, it costs nearly as much as the IDE itself, and the initial reviews are fairly mixed.
What’s interesting though is that it’s powered by something called JetBrains AI service, and this allows the chatbot to be powered by multiple different models. There’s not a ton of detail around this, but it’s an extremely powerful idea. With co-pilot and duet, you have GPT-4 and Gemini, but with JetBrains, you could hypothetically hook up your own fine-tuned model, like maybe the open-source Code Llama or Mistral Chemin with the billions of lines of code at your company, allowing the AI to provide predictable responses based on your coding conventions. That’s pretty cool.
But at the same time, GitHub co-pilot just leveled up big time. Previously, it was based on GPT-3.5, but now it’s using GPT-4. It now has a dedicated chat window, so you basically have chat GPT in your IDE. But what’s really amazing is that it has this Workspace command that allows you to search all the code files in your workspace to write code with the proper context in your project. So far, I’ve actually found it most useful for explaining code when jumping into an unfamiliar project. In other words, it makes it way easier to figure out what the hell is going on with this guy’s spaghetti code. It also now writes your commit messages and is also available in the terminal to explain and refactor commands. But there’s one big problem with these AI tools: sometimes, they spit out suggestions from code and repos that you’re not allowed to use. Luckily, co-pilot will now search across billions of files in GitHub and will return the licenses on any similar code, and that means you can be relatively confident that you’re not stealing someone’s code.
Overall, these changes are pretty amazing and extremely useful. These tools still can’t build complex projects out of nowhere, but the big question is, where are we going from here? Currently, co-pilot doesn’t run your code or create new files, but I think we’re going there eventually. You might be able to just take your client’s requirements and say, “Hey, build this thing in Django.” Then co-pilot will run the commands, create the files, and run the unit tests required to make that a reality. Then you’ll continue fine-tuning it with multiple shots or prompts until it builds exactly what you’re looking for, without ever touching a line of code. Another missing piece is image data. There are already tools that can convert a screenshot into code, and as these tools get better, we’ll likely see them integrated directly in the IDE.
Now, even if this stuff does get really good, I wouldn’t be discouraged as a programmer. Right now is one hell of a time to be alive. Code is just a means to an end. Even if programming becomes obsolete, there will still be engineers pushing the limits of whatever comes next. We still need to develop robots, brain chips, quantum computers, euthanasia pods, laser guns, space travel, time travel, and all sorts of other sci-fi. And it’s going to take some good old-fashioned problem-solving engineers to do that. So stay optimistic. This has been the Code Report. Thanks for watching, and I will see you in the next one.
