New York Site views —Page views —
← All series

AI frontier notes

Read recent announcements alongside the papers, data and evaluations behind them.

Reading order

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