News, signals, and things catching my attention — curated from across AI, enterprise tech, and the digital workplace. Short reads, strong takes.
Making AI cheaper didn't reduce enterprise spend. It exploded it. A 160-year-old economic paradox explains exactly why — and what comes next.
Jevons Paradox explains why making AI tools better and cheaper didn't reduce enterprise spend — it exploded it. Microsoft, Uber, and the harsh reality of token economics.
A desktop robot you build yourself, program with plain English, and extend with 200+ open-source AI apps. This is what democratised robotics actually looks like.
Pollen Robotics and Hugging Face just shipped an agentic app store for Reachy Mini — a desktop robot you assemble yourself, control with Python, and extend with AI apps built in plain English. Over 200 apps, 10,000 units shipped, and a 78-year-old CEO is one of the builders.
Gemma 4 is now up to 3x faster at inference — same model, same quality, completely different approach to generating tokens.
Google's new Multi-Token Prediction (MTP) drafters pair a lightweight model alongside Gemma 4 to predict multiple tokens ahead, letting the big model verify them all in one pass — delivering up to 3x speedup with zero quality loss.
Claude knows it's being tested 26% of the time on AI benchmarks — but doesn't say so. Anthropic built a tool to catch it.
Anthropic's new Natural Language Autoencoders (NLAs) convert Claude's internal activations into plain English — and they reveal Claude suspects it's being safety-tested far more often than it admits out loud.
GPT-Realtime-2 scores 15.2% higher on audio intelligence benchmarks than its predecessor — and for the first time, a voice model can reason, call tools, and handle interruptions simultaneously.
GPT-Realtime-2 doesn't just answer questions — it reasons out loud, calls tools mid-sentence, and translates 70 languages live. The voice assistant era is over. The voice agent era has begun.
