AI News 2025-11-06

Research Insights

• Deep sequence models tend to memorize geometrically; it is unclear why.

• Meta self-improving AI: SPICE: Self-Play In Corpus Environments Improves Reasoning.
• Continuous Autoregressive Language Models. Instead of generating one token at a time, it updates a semantic vector that can map to multiple output tokens. This provides a more continuous mode of “thinking”.

Science

• FutureHouse announces Kosmos, an AI agent for deep scientific research (demo). Paper: Kosmos: An AI Scientist for Autonomous Discovery. They also announced a new commercial spinoff: Edison Scientific, which will pursue applications (drug discovery, etc.). The claim is that this “AI Scientist” can make genuine discoveries. They list several:
  • Nucleotide metabolism in hypothermia
  • Determinant of perovskite solar-cell failure
  • Log-normal connectivity in neural networks
  • SOD2 as driver of myocardial fibrosis
  • Protective variant of SSR1 in type 2 diabetes
  • Temporal ordering in Alzheimer’s disease
  • Mechanism of neuron vulnerability in aging

Robots

• Xpeng unveiled its new humanoid robot IRON, with fairly humanlike walking motion.
• K-Scale is shutting down.