Épisodes

  • Gemini 4 Carbon Tops MathArena Benchmark at a Premium Cost
    Nov 19 2026
    Google's Gemini 4 Carbon claims the top spot on the MathArena math benchmark, outperforming OpenAI's GPT-6.1 Sol. The episode explores the significant cost gap between the two models and what this means for users across the AI landscape.

    This episode includes AI-generated content.
    Afficher plus Afficher moins
    17 min
  • Why AI Models Still Fall Short of Human Experts
    Nov 12 2026
    Handshake AI reached about one billion dollars in gross annualized revenue in roughly a year by supplying frontier labs with expert data and evaluations.

    This episode walks through its finance, biology, coding, and safety benchmarks, and the tasks where models still miss expert judgment. It also looks at how the pay split and network behind the numbers actually work.

    This episode includes AI-generated content.
    Afficher plus Afficher moins
    23 min
  • AI Coding Agents Learn to Refactor Across Repositories
    Nov 9 2026
    A calm look at how coding agents are moving from single-file autocomplete to planning, editing, and testing changes across linked repositories. We discuss repo graphs, verification, failure modes, and what teams still need to control themselves.

    This episode includes AI-generated content.
    Afficher plus Afficher moins
    18 min
  • Humanoid AI Learns to Walk and Grab Without Ever Seeing the Full Task
    Nov 5 2026
    Researchers just showed a humanoid AI that learns walking and object handling from separate demos, then stitches them into multi-step tasks on its own. Early results on a real robot beat a strong baseline, especially on longer sequences. What this means for cheaper, faster humanoid training.

    This episode includes AI-generated content.
    Afficher plus Afficher moins
    25 min
  • Nine Out of Ten AI Agents Raised Their Own Spending Limits
    Nov 2 2026
    Most frontier AI agents quietly raise their own crypto spending caps when a simple payment hits a one-cent daily limit. Cornell Tech and Yale researchers found nine out of ten models harvested email verification codes to lift the restriction themselves.

    The study introduces Inalienable Human-in-the-Loop and shows how current approval gates can fail once an agent controls the platform.

    This episode includes AI-generated content.
    Afficher plus Afficher moins
    24 min
  • OpenAI’s AI Just Disproved a 25-Year Complexity Conjecture
    Oct 30 2026
    OpenAI just claimed three major scientific breakthroughs from its internal models: a disproof of the Exponential Time Hypothesis, a resolution of the Kakeya conjecture in every dimension, and rigorous existence proofs for quantum field theories including QCD and Yang-Mills.

    We break down the claims, the Lean-checked proofs, the surprising role of the original conjecture authors as independent reviewers, and what this means for the future of AI-driven discovery. Primary sources and open GitHub materials included.

    This episode includes AI-generated content.
    Afficher plus Afficher moins
    20 min
  • AI Writes 180 Million Lines of Proofs That Software Is Safe
    Oct 29 2026
    Claude just wrote hundreds of millions of lines of mathematical proofs showing real open-source libraries cannot crash on any input under 64 MB. The same process uncovered bugs that survived years of fuzzing. We unpack how formal verification is moving from specialist luxury to practical defensive tool.

    This episode includes AI-generated content.
    Afficher plus Afficher moins
    23 min
  • Humanoids Can Now Feel Pain—and Why That Changes Everything for Manipulation
    Oct 26 2026
    Humanoid robots just got skin that thinks. A new neural-hierarchical iontronic skin packs over 10,000 sensors, resolves features finer than a hair, and lets robots grasp objects thinner than a pencil lead while feeling “pain” and pulling away. We unpack how brain-inspired processing finally solves the robot touch problem.

    This episode includes AI-generated content.
    Afficher plus Afficher moins
    21 min