AI frontier notes
Read recent announcements alongside the papers, data and evaluations behind them.
Reading order
- Reading a brain atlas in 32 minutes: which part of research did the agents accelerate?
- How much can AI monitoring catch? METR’s retrospective red-team tests
- What do the 60 public entry points in Claude’s software proofs guarantee?
- Which reasoning paths should survive? PKTP and the pass-at-least-once objective
- Can fine-tuning move between model sizes? Ariadne and Transformer transfer
- OpenAI’s claim about the Exponential Time Hypothesis: the proof and its review
- Learning robotic control directly: what LECO tracks and where fallback control matters
- How many languages were tested? Auditing a multilingual evaluation claim
- When should a teacher take over? Local interventions in long-sequence distillation
- Which experiences should an agent learn from? Memory routing in continual learning
- Does synthetic training before pretraining pay off at equal compute?
- Can models communicate visually without translating into text? The Latent-Lens experiment
- Does a pretraining lead survive fine-tuning? A data-scale reversal for Muon and AdamW
- Hy4 adds vision and computer use: what does the release establish?
- Which compressed model should you keep? Seventeen ranking reversals in CompressionBench
- Anthropic’s capability metric grew 1.9 times as fast. How much can be attributed to AI?
- Letters facing up, words still misspelled: tracing robotic failures in ToyBlocks
- Does an experiment survive a GPU-to-TPU migration? XStack tests 20 research repositories
- Can a voice assistant keep talking through an interruption? MOSS-Live’s duplex evaluation
- Change red to navy: does image retrieval still remember the hood? TRIAD in multiple turns
- Can an agent spot a pothole without its image tools? Two MKA comparisons
- How does a robotic hand use touch to pull a zipper? Reading LingBot-Touch’s score of 59
- How early can PREWARN detect repetition? A three-model comparison
- Thirty estimates from one model: why did the median error fall by just 1.7%?
- A correct solution to the wrong problem: reading XYBench
- Generating Breakout and repairing a network: two different tests in WeightBench
- Has the model seen this room before? Spatial memory in ReSpaceBench