Dario Amodei shared a quite ... ambitious ... timeline / expectation on when AI will be able to build software on its own. I'm not so sure about it...
Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.
Transcript
From the show. Tap a time to play from there.
At this year's World Economic Forum in Davos, Dario Amodei predicted that within the next 12 months or so, AI would be able to write all code, fully automated, on its own, essentially. And it's worth paying attention to what this man has to say and I'll also share my thoughts on it and why I think that you should take a nuanced look at that. Not just because Dario, of course, is the CEO of Anthropic which is one of the most important players in
generative AI space, especially when it comes to models related to coding. But it's also worth paying attention year, in the beginning of 2025, he predicted that AI would be able to write 90% of all code within the next three or six months and depending on how you look at that, he was not entirely wrong. Now, sure, this timeline is probably, or was probably a bit too aggressive. It maybe wasn't three or six months, more like six to nine months, not sure.
And, of course, it totally depends on your field, so in which area you're programming, also in the programming language you're It will depend on your company, the policies you have in your company. And of course, it also depends on, on you, your preferences, your experiences with AI. I can say that for me, I am at a point I would say since maybe December, November, where for some projects, most of my projects, AI does write 80, 90% of all code for me probably.
But this is now where things get interesting. I'm not talking about wipe coding and I'm not talking about AI doing it all on its own and I'm also not talking about AI getting it all right. Instead, things are way more difficult than that. So it does absolutely not do that on its own. It also is not about wipe coding here. I mean, you can probably wipe code but I've shared my view on that before
me, by the way, just to make sure we all got the same really means that you don't look at the code at all, that you don't care about the code, that you don't necessarily even know how So yeah, this is also not about wipe coding here. Instead, I would say we are at a point where you can hand off many tasks to AI with a good plan. So if you have a good plan, AI can implement it. So I think we're at that
point right now. You, you can definitely do that, at least for some tasks. But of course, you wanna review that code and, uh, I know there are people that tell you you don't need to do that anymore. Might work for them. Definitely not working for me and if you're shipping anything, if you're shipping anything to customers, you're responsible for that. You, as a developer, you can't say, "Oh, the AI got that wrong." No, it's your responsibility and I won't take responsibility in code
I haven't reviewed, I don't understand. Also, still, AI makes many mistakes, so I need to fix mistakes or steer AI in the right direction. And that is also very important. That's, of course, important to the planning part or related to the planning part, that you steer AI in the right direction. But it's also related to what you do with code AI gives you.
So very far away from AI generating 90% of the code and that means I have no work to do. Far away from that. It just means that I got a very fast typer who can implement my plans, but those plans need to be good and the output is kind of varying in quality. Still, for me, it's very likely faster than it all from scratch. If I instead try to build good plans,
split that up into smaller chunks, let AI write the code for me and then review and fine-tune that code. And by the way, with planning, I mean really detailed plans where I also break down the exact libraries I wanna use, the patterns I wanna use, the architecture of the software I wanna step-by-step. So it's not a, a rough plan or a general plan; it's a very detailed plan. That works for me and therefore, for me, with, uh, with these restrictions here in mind, I would say, yeah, it can probably write
90%-ish of the code. But that does not mean that it does 90% of the work on its own. Now, that's important context because that's to last year's statement by Dario. So yeah, I would say we are kind of there, but not in the sense of AI doing it all on its own. Now, what about this year's statement? Though, in this year, Dario essentially said that AI will be able to do what software engineers do and write software fully on its own, end-to-end, within six to 12 months.
That's this year's statement. Now, I fully recommend watching this entire talk. It's very interesting, but I have some thoughts about this statement. And obviously, just to make and obvious, I'm in no way smarter or more capable in judging the performance of AI models than Dario. But I'm also not the CEO of a company that needs to sell these AI models and I can share from my own experiences. Now, as I said, I would agree with last year's statement to some degree, but with many caveats, with many
restrictions. The AI is definitely not writing 90% of my code its own. So naturally, I have a very hard time imagining... that this will come true, and I have a very hard time imagining that this will come true, not just within the next 6 or 12 months, but anywhere in the near future. Now, I totally see that AI is capable of building software on its own in a loop, with the Rolf loop that's getting a lot of hype, uh, related to Claude code. I totally see
that. But the full work of a software engineer includes the tasks I outlined here and arguably more than that. It includes building a good plan, defining which architecture, which patterns, which technologies to use, reviewing the code, and also, of course, analyzing the code, fixing problems with the code, taking responsibility for that code. That is something I have a very hard time
within the near future because right now, the AI I can use is too far away from that. It's a talented, fast writer that makes a lot of mistakes along the way and that needs very clear guidance. And that going to models that can do it all on its own, that can plan out entire architectures cleanly on their own, that are capable of using the latest technologies, of writing error-free secure code totally on their own without review or only with review by other
AIs, that is something I have a very hard time seeing when I look at today's models and also when I look at the progress of models over the last years. Because, yeah, that progress, of course, has been remarkable. It has been steady and, and good, and especially the tools I shared that before got way better. So I shared it before with models. I'm not sure if we're the raw model intelligence, we're still on a linear, leave alone exponential trajectory. For the tools, I would say we definitely are on some
linear trajectory. But I have a hard time seeing that being enough to get us to that full automation anytime soon. Now, naturally, these CEOs have other interests than, than me and you could say I have the us developers also because I also sell programming courses. But let's be totally honest, that is not my role and nothing I will be able to do. I'm just sharing here what
my experience with AI has been, and I'm very open-minded, I would say, for AI. I'm using it a lot, as I said. It does write 90% of my code. It's just far away from that full automation. But yeah, please let me know your thoughts on this as well, your experiences with AI, working because you might not be using AI at all or only for small isolated tasks and you might still write most of the code or you're using it for everything, you're not even looking on the c- uh, code, and you're having great results.
I'm interested to learn more about that, so please let me know. And yeah, have a great time
GPT 6 Astra was released. It's a good model, I have to say. Especially for 3D work it seems to be a real game changer. But due to all the hype, you may be disappointed...