🤝 The study group that beat the answer key
Two models with no answer key grade each other's work and out-learn the one trained on real labels. Plus a safety report card and a DNA test for weights.
The Editorial Team
The desk behind AI Reach-out Frontier. We read the whole paper so you don't have to pretend you did.
🔓 The AI's private thoughts weren't private
Providers hand you an encrypted copy of a model's reasoning. A paper feeds it to a weaker sibling and reads it aloud. Plus the week's research news.
📐 The refactor no human reviewed
One engineer let an AI rewrite 189 files of a 717k-line app with zero code review — the trick was auditing the spec instead. Plus the week's research news.
🐉 The AI that reasons without words
A 150M-parameter model scores 29.5% on ARC-AGI for $0.0007 a task — by thinking in silence. Latent reasoning in plain words, plus the week's research news.
🎧 The grading rubric that rewrites itself
AudioRubrics, plainly: teach an audio model with a rubric that evolves mid-training. Plus Qwen3.8-Max, DeepSeek V4-Flash, and Anthropic on open weights.
🧠 The AI that actually remembers you
A new paper builds memory directly into the model instead of bolting notes on the side. We explain Metis in plain words, plus the week's research news.
News
Welcome to AI Reach-out Frontier
The research behind the headlines, in plain words. One paper explained and the news that matters most.