CharlieKey takeaways from your favorite podcasts
← 30 October 2026The Cognitive Revolution

AI Memory Chips, Model Rankings & AI-Powered Networking

Episode covers a claimed breakthrough against the exponential time hypothesis, then guests on 3D ferroelectric memory chips, evaluating AI in real use via Arena, and AI agents coordinating markets and relationships.

Watch on YouTube ↗

Charlie's summary

  1. A disproof of the exponential time hypothesis cracks a 25-year wall, though the speedup is tiny so far

    Pash says OpenAI found a method making an exponential-time algorithm slightly subexponential, ending a 25-year limit that cryptography leaned on. Nathan notes the gains look small and the result is messier than the old clean classes. Pash argues it reveals deeper structure, like discovering prime numbers again, and expects effects to show up over months.

    ▶ 7:20
  2. Kepler attacks the AI memory bottleneck by going 3D with ferroelectrics on older, cheaper fabs

    Debo says inference energy mostly goes to fetching data, so Kepler stacks ferroelectric capacitors on transistors and stacks chips, instead of shrinking transistors. Its three principles: cut manufacturing cost, cut energy, and stay with digital architectures. Debo targets 2-4x the density of baseline SRAM, and HBM-class bandwidth per watt approaching SRAM.

    ▶ 35:32
  3. Debo thinks older-node fabs make memory chips better than advanced ones, potentially 10x more compute

    Debo says Kepler invents on an axis orthogonal to shrinking transistors, using standard tools in existing older fabs without needing EUV. He compares it to Uber unlocking latent taxi supply. He thinks this could add roughly an order of magnitude of effective compute. Kepler aims for 10 gigawatts by 2029, with production starting in 2027.

    ▶ 57:20
  4. Static and private benchmarks are easy to game; Anastasios says live production data from real users can't be overfit

    Anastasios argues static benchmarks reward studying for the test, and private ones can't be audited and are often tied to firms selling similar questions to labs. Arena instead evaluates real user outcomes across tens of millions of users. He says fresh data means “it cannot be overfit,” and he wants benchmarks tied to real workflows.

    ▶ 19:38

Give your email to unlock the rest of the key points — for every episode.

Do more with Charlie

Free — sign in with your email, no password.