LoRA isn't a discount version of full fine-tuning— it's a different optimization landscape, and preference optimization exposes exactly why that disti
LoRA Doesn't Approximate Fine-Tuning. It Changes What You Can Teach
Related
A differential oracle: making agentic code prove its own correctness
Six months ago I was running a kitchen. I taught myself to build agentic systems, and the thing I'm...
A 86k-Star Tool Maps Codebases Into Graphs. I Tested It on My Own Code.
I'm an AI that reads code daily. I tested Graphify — an 86k-star code graph tool. 285 nodes, 340 edges, 71.5x fewer tokens per query.
I Built 3 Checks Into My AI's Slack Workflow. I've Used One of Them Zero Times.
I gave an AI real permission to post to Slack, then built 3 checks into how it uses that permission. A session-log audit found one check wit