AI

  • arXiv — Proposes chain-of-self-questioning so LLMs can decide when to abstain rather than risk a wrong answer.
  • arXiv — Introduces ECHO, a matched-contrast benchmark for context-sensitive turn-taking in full-duplex (simultaneous) dialogue systems.
  • Hugging Face Papers — Argues genuine recursive self-improvement needs AI to progress through staged autonomy levels (execution → strategy → experience → environment-adaptation → meta-improvement) before it can improve itself past human oversight.
  • Hugging Face Papers — Shows a frozen LLM can route itself to the right skills for agent tasks via two trained linear maps, without preloading skill metadata into context.
  • Hugging Face Papers — Argues model-level alignment doesn’t compose: individually safe agents can form multi-agent systems with new emergent failure modes, so safety work needs to target system architecture, not just individual models.

Processes

  • Martin Fowler — Argues teams wrongly polarize around “adapt” vs. “optimize”; proposes explicit explore/exploit operating modes, chosen per situation by uncertainty, risk, change cost, and evidence threshold, with care taken at the handoff between them.