Skip to content
The Harness Is the Part You Can Actually Build
A builder’s read of the new “agent harness” research, and the accountability trap hiding inside it.
Leggi tutto
A Model That Sorts Its Own Doubt
NVIDIA made an AI 2.42 times faster by freezing the half that understands, and for legal work the speed is not even the interesting…
Leggi tutto
Your AI Agent Just Read the Whole File Cabinet to Change One Word
Researchers taught an agent to know when a task is easy. The catch is where the waste hides.
Leggi tutto
The Only Benchmark That Pays
Why a vendor’s “state of the art” tells you almost nothing about how a tool will handle your own files.
Leggi tutto
The Narrowing: What an AI Ideation Study Reveals About Legal AI’s Favorite Move
When models are asked to think, they get more predictable, not less. That should worry anyone building legal AI.
Leggi tutto
How an 8B Model Beat GPT-4 Without Being Smarter
New research on AI agents points to the fix legal work actually needs, and it is not a bigger model.
Leggi tutto
A Model Can Be Faithfully, Confidently Wrong
New research teaches models to say “I’m not sure” and mean it. That solves a smaller problem than most people in legal AI think…
Leggi tutto
Fail Closed
Google built an AI that reviews scientific papers before submission. The real lesson for law firms is in what it gets wrong.
Leggi tutto
