AI Software Development – What Does The Data Say?
A lot of interesting points in this article, but I have a hard time to put weight on future predictions, such as: Truly autonomous and reliable long-horizon agentic software development is so highly improbable using LLMs that it’s essentially science fiction. Maybe it is with LLMs but maybe other technologies in tandem with LLMs will archive this. (Technological) Progress happens really fast these days.
- AI use increases coding activity very significantly, but that isn’t translating into more value being created. By the time we look at not just what gets shipped, but what gets used, we see no discernible gains.
- LLMs cannot distinguish between recent and out-of-date information in the context, and information in the model itself can often “outweigh” information we give it.
- Repo-level .md files tend to make model performance worse, probably because they add noise instead of signal in many specific tasks.
- LLMs struggle with negation.
- LLM inference is more accurate when we give them examples (demonstrations) rather than just describing what we want. They’re pattern-matchers.
- Large/long-scale industry studies show a clear trend – output is up (more code, more commits, bigger diffs), but outcomes don’t reflect that trend. If anything, the average team is taking longer to ship worse software. If ever we needed proof that software development isn’t a production process…
- In case you were wondering why they suck at the “big picture” – probabilistically, it’s a blur.