mrmemes 🎩

phil

"infects our code bases" → enfeebles our brains 🧠

...also, is inevitable, and I look forward to the glory on the other side of the "first most obvious land grabs" at all of this LLM assisted code production, when good design and mature insights manifest into genuinely strong empowering tools that amplify the human.

It took a couple of hundred years before the Industrial Revolution started getting good again for the artisans.

Completely inevitable. It can also work really well when used with constraints and rigor. I've been leveling up at using LLMs to write maintainable and well-factored code, but it really does require me having an idea of what the architecture ought to look like and being VERY judicious about what I'll accept from it.

co-founder @chaindrop § software engineer § privacy wonk § music head § theoretical memeticist

e.g. "write me a report system that pulls the latest block chain data and let's me filter it" is absolutely going to get you slop. Doing architecture up front and using SOLID principles and using protocol-driven design (basically that stuff you learn doing your 10k hours of SWE) can yield good code quickly.

Llms are definitely garbage for this. Add more than a single chain and beyond the 5-10 most popular protocols and you're screwed, can't filter a thing without first being able to categorize and label the onchain data.

Get Data or Die Tryin' /noves 🔮 | linktr.ee/czro

Yep, the context window size and the specific model make a big difference there… but isolating segments of the system, putting them under test and having good separation of concerns are all things you have to force on the LLM right now. It won’t do itself or consistently use them when present.