Every week there are hundreds of AI papers. Most are incremental. But every now and then, something drops that genuinely changes the game for working engineers.

Here are 3 patterns I'm tracking right now:

1. Context window management is becoming a first-class engineering problem. Models are getting longer context, but that doesn't mean you should dump everything in. The engineers getting the best results are treating context like memory management — deliberate allocation, not brute force.

2. Code generation is shifting from 'write this function' to 'refactor this system.' The interesting frontier isn't autocomplete anymore. It's agents that understand your codebase architecture and can make cross-file changes. If you're not experimenting with this, you're behind.

3. Fine-tuning on your own codebase is more accessible than ever. You don't need a research team. A weekend, a decent GPU, and your git history can produce a model that understands your patterns better than any general-purpose tool.

I break down stuff like this every week in Mainframe — the AI newsletter for engineers who want signal, not hype. If that sounds useful, come check it out.