Writing
4posts found
AI removed code generation as the constraint, and the bottleneck moved downstream to review. What the data says, how teams are responding, and why the person who can tell when the output is wrong is now both the bottleneck and the multiplier.
You don't need to be a researcher to make smart choices when downloading models. Here's what quantization actually is, why it barely hurts quality, and how to pick the right version on Ollama or Hugging Face.
Prompt-tuning can't fix systematic agent failures. Harness engineering, the scaffolding around the model, is where the real leverage is. Same model, 6x performance, and why the engineer's job is shifting.
The problem at the heart of every AI agent, why the obvious solutions fall apart, and an honest look at what the current wave of memory infrastructure gets right and wrong.