AI Is Learning Things Nobody Taught It

TL;DR

• 1

  Codex is spilling into the physical world: One tinkerer photographed an MP3 player's motherboard, fed the details to Codex, and got step-by-step help analyzing binaries, rebuilding firmware, and eliminating Bluetooth stuttering on a device never meant to be hacked this way.

• 2

  Hugging Face is now selling robot legs for $2,500: The new Le Robot humanoid project is not a full humanoid yet, but the pitch is clear, open hardware that builders can understand, repair, modify, and use in research instead of treating robotics like a sealed black box.

• 3

  AI audio is getting physics-aware, not just pattern-matched: Sony-backed PAVAS estimates hidden variables like mass, velocity, impact, and material so a giant dinosaur can sound heavy instead of producing the kind of tiny footsteps ordinary video-to-audio systems might hallucinate.

• 4

  Some researchers think scale alone will not produce real reasoning: Citing Jose Crespo, Dylan highlights a seven-layer architecture built around substrate, symmetry, connections, evidence, and computation, arguing that Google and Anthropic may be among the few labs with the mathematical depth to attempt it.

• 5

  AI can make people more confident while making them less careful: In one study on 'cognitive surrender,' workers accepted wrong AI answers more than 80% of the time, and Anthropic's reported Claude fluency scorecard suggests models may soon grade users on how well they prompt, iterate, and verify.

• 6

  World models are expanding beyond text and pixels into biology and physics: BioHub trained on 2.8 billion protein sequences to predict structures and design binders for five cancer and immune targets, while a separate physics idea asks whether quantum mechanics itself must be rewritten to make room for gravity.