Scientific Software Development and Vibe Coding

Scientific Software Development and Vibe Coding

Foundation models are everywhere I look these days. Claude, Copilot, ChatGPT, Gemini, Kimi, Qwen, Mistral, MiniMax, Llama — and that’s just what I can name from casual memory. Everyone has a proprietary harness that puts these models to work. And truly it’s everywhere. GitHub is exploding with AI PRs and commits and vibe coded projects.

I’ve got some conflicted feelings about LLMs; and I’m worried about where things are going.

Hidden Costs of New Technologies

I’m worried — not because I think the kids won’t turn out alright; they always do — but because we’re at the point where some skillsets are going to vanish. And they’re going to vanish at a pace we’re unfamiliar with.

This concept isn’t new. Societal change alongside or in reaction to new technology is as old as humans. But the rate of technological development in agentic LLMs is outpacing the rate at which humans can be bothered to learn the dying skills. And that’s being compunded with the addictive / compulsive nature of AI usage.

For example, the invention of writing fundamentally changed the need for oral memorization and languages based on these skills are dying. And storytelling itself is a skill that has drastically changed. Today, writing lets us record the things we know and fictions separately. It’s no longer necessary to embellish with parable and excitement and intrigue when recording the things we know, and now this is mostly done either for the joy of it or for recording fictions.

In that same way I’m worried that kids — and adults, for that matter — with regular access to AI will lose their problem solving acquity and critical thinking skills, their ability to write proficiently, and some of the creativity that drives personal and professional innovation.

From what I’ve gleaned, early studies are showing high rates of skill loss and that AI is inducing behaviors similar to addiction — a not so suprising twist based on how social media and product design have gone over the last 20 years.

All of this leads me to believe that academia and technical R&D, especially the parts of R&D that ride the edge of academia and industry, will see a major decline in areas where AI use is prominent.

Manifestations of Concern

Universities are partnering with AI programs, my alma mater has now paired with gemini and offers gemini to all students seemingly without bound on use cases. Some classes and some professors seem to be putting restrictions on AI when it matters more, but I’m not convinced just saying “don’t for your own benefit” will really work all that well. When I was that age I did all kinds of stuff that was not to my own benefit and was logically aware of it the whole time.

People aren’t being taught how to use AI, they’re just using it blindly.

Over budget project scientist used AI and made a fundamental mistake that could have been avoided with some critical thinking skills.