I’m curious to get a general pulse on how people here view the rapid evolution of AI, specifically on lemmy. On one side, there’s growing interest in local open-weights models, self-hosted setups, and practical developer/productivity tools. On the other, there seems to be a growing list, I can’t keep track of, that are seen as major concerns. I’m not here to pick sides or be biased in one corner. Im genuinely curious what people have to say about AI in general. And maybe it will help me and other get a better handle on how we perceived it in our ever changing world.
I’ve been doing scientific software development for 20 years, and here’s my personal experience:
Software development as it had been is already over due to AI. There’s no reason to write code yourself anymore in 99% of cases. Even highly technical stuff (GPU algorithms, for example) are no problem for the AI.
Scientific work is very different than it used to be. The AI doesn’t have the initiative to just do it for you yet the way it can with writing software, but it’s an expert in every field so it provides an excellent foundation for your own novel ideas and it can often suggest new ideas from fields you aren’t personally knowledgeable about.
Socializing is going to be very different soon. Right now the AI is already more interesting to talk to than most humans, but talking to it doesn’t feel like talking to a human and doesn’t satisfy the need for interpersonal interaction. I expect that to change. (I suspect that the current absence of a human-like AI is more because developing it is not good business than because it’s technically difficult.)
A lot of AI critics have objectively false ideas of how well AI works. I think they either haven’t tried cutting-edge models recently or they have deliberately tried to make those models fail. You can break most things if you try; that doesn’t mean they don’t work well.
As I type this, an AI is working on a tricky problem in molecular dynamics for me. It has already found and fixed issues with the algorithm that my co-worker spent a month on, but it thinks that a different algorithm it came up with can do even better. After coming up with the idea, it’s independently implementing it, unit testing it, running simulations with it, and doing statistical analysis of those simulation results. If it gets errors, it can debug them on its own. If the results aren’t as good as it expected, it can iterate over different variants on its own. If the algorithm ends up not working, it can reason about why, from a scientific perspective. What the AI does in hours would have taken me weeks if I came up with the idea, which I didn’t.
Right now human experts can still do what AI does, just slower. I expect that in less than 5 years, that will probably no longer be true. Predicting the future after that point is nearly impossible. Will we all die? Maybe.
Similar field, similar experience. I think it’s easy to sleep on how much LLMs can help your work if you’re just using chatbots- because those give incorrect answers so often that most interactions end up being a waste of time.
But if you give the model tools to test its answers, and prompt it in a way that restricts responses to tool outputs instead of generated answers, suddenly the many mistakes are less of a problem. It wrote a function with 5 bugs that did the wrong thing to begin with? and then it caught all of those mistakes in testing and fixed them, and it took 6 seconds, so it doesn’t really matter. Maybe that’s how I write a function in my head too, I just think through the most obvious issues before typing the code and testing it.
The high level planning still needs frequent and detailed instructions, otherwise it tends to go down the wrong rabbit holes.
Setting up the environment took some time, bit since I started using claude code I actually started to go through my tech debt backlog, because fixes that used to mean I have to take a day off, I now take 5 minutes to write down the issue and the scope of the fix, and 10 minutes to watch it being fixed. I am pretty sure I will stop writing code in a year or two maximum. I am slightly worried what happens to the codebase if prices of tokens get out of control, but I guess local models could help as long as you have the hardware.