A Columbia paper found LLMs edit already-optimal code 100% of the time. One prompt line halves it with zero false refusals. I reproduced it and…
Context compaction quietly drops the rules you give a coding agent. Two papers put numbers on it (17% and 53% retention). Here's what I moved…
A Purdue study got Qwen, Mistral and Llama to 'verify' a fake developer with a quiz they wrote themselves. What that means for your agent's…
IBM's consistency gap paper shows an agent passing 77% of runs succeeds five times in a row only 53% of the time. Here's the five-run…
A Stanford paper finds API benchmark scores run 3.4 points higher than the same models in their chat apps. What that means for how I…
Hugging Face lifted a 350M model's structured-output score with 100 GRPO steps and no labels. The trick isn't the model, it's three reward functions you…
GPT-6 Astra scored 62.7% and 99.9% on ARC-AGI-3 with the same weights. The harness made the difference, and that should change how you evaluate your…
A law firm's AI governance story, read from a three person shop's point of view. Where AI automation for small business breaks, and the one…
ChatGPT Work is really two products with one name. After a week of real tasks I break down Work Cloud vs Work Local, Codex, and…
A 4-bit model just beat the 16-bit checkpoint it was quantized from. I break down quantization-aware healing, the numbers, and what it changes for local…