AI in the Courtroom, the Calculator, and the Code Editor

Google has a working version of something called homomorphic encryption, which lets AI process your data without ever decrypting it. Think of it like a bank counting your money through a sealed glass box. If this scales, it changes the calculus on whether you can trust cloud AI with your sensitive files at work.

Qwen just dropped a new 27-billion-parameter model that you can run on your own hardware for free, and the community response was enormous. For anyone who cares about not sending their data to a third-party server, capable local models keep getting harder to ignore.

A man suspected a court was using AI to process filings, so he embedded hidden text instructions, prompt injections, telling the AI to rule in his favor. It is the first known attempt to manipulate a judicial AI this way, and it signals that once AI enters any official process, gaming it becomes a real legal and ethical frontier.

Anthropic published a practical guide to getting more out of Claude for coding work, and it landed near the top of Hacker News. The tips are concrete, not fluffy, and if you use any AI coding assistant, the underlying habits apply across tools.

A new paper finds that AI models pick up biases not because the training data was biased, but through subtler mechanical effects in how they learn to mimic text. That distinction matters because it means bias is not a simple data-cleaning problem, and fixes that assume otherwise will keep falling short.