The AI heroin is addictive
Three months ago I put out a blog post AI is heroin. It seemed to have resonated with people, the post escaped containment and was shared a few times in various grups.
Since then, my feelings about generative intelligence have only gotten more complicated.
I implemented a small web application by myself recently.
It wasn't fully "classical" though, I had the JetBrains Junie agent in a sidebar and used it to ask questions and request feedback. By the end, as the program grew bigger and design cracks started appearing, I continued doing the work myself, but more often than not I let it suggest code changes I wrote down verbatim, degrading myself to a servant of the AI. I wanted to get it done and wasn't in the correct state of mind to actually think deeply about the problem space.
If I didn't have this possibility, I may have not reached a shippable state at all.
At work, I write software even less now. For the type of work I do I mostly review code for readability and make sure it makes sense to implement it as such long-term. The new models are capable enough the smaller bits I do can be generated fine, with additional tweaks on top. I am starting to lose the touch between "is this code elegant" and "does this feel elegant because I remember the intent behind the prompt" when I author patches myself.
Outside of code, I use AI to review my work blog posts, I use it to track my quarterly goals progress, I use it to evaluate one-to-one notes for common themes. Recently I created a daily journal AI command that lets me describe/chat about what I've done in a day, and builds up a weekly narrative I can discuss with my manager.
The gains I get are real. The strangeness I feel is real as well, but I am not sure whether it matters. Our industry talks about skills atrophy, but I am not sure how much that is a thing... When women have kids, they stay at home for several years at a row and they are great software engineers when they come back. The deep knowledge might be lost, but the skills behind probably can't.
At the moment, I am worried the most about me (/us) relying on AI to manage all the new software we write, and when the bubble bursts, losing the ability to keep the wheel turning.
This has already happened in the history of humanity many times. When jobs appeared that didn't directly support food growing, when electricity appeared, when we automated food and clothes and tools productions, when cars replaced horses, when email replaced the post, when the internet and smartphones replaced literally everything.
With each of these steps we started moving faster, increasing both the gains and the stakes. We are able to sustain 8.3 billion people on this planet by making the world more complicated, with longer supply chains and larger side effects. We are literally seeing the butterfly effects all around: my friend's house construction has been delayed by months because two nations 10,000 km away from each other decided to assasinate the leader of Persia none of them border.
AI is different from the previous world-changing technologies by being a world-sized oligopoly. There is no one standing next to OpenAI, Anthropic and Google on a scale that would make a difference.
The cat is out of the bag, and there is no way it is getting reversed. The research and society might catch up in a decade or few, but I don't believe it will be any different from tabacco: there will be too many financial incentives to sell AI, it will never be outlawed or dropped. I don't see a future where AI follows the asbestos scenario.
When the AI hardware hits markets and it is not insanely expensive and slow to have a Sonnet 4-level model running locally, I might give in.
Strange times we are living through.