CharlieKey takeaways from your favorite podcasts
← 7 October 2026The Cognitive Revolution

Formal AI Safety & Risk Strategy + LoRA-Powered Work Agents

AI is now composing music with real instruments and solving landmark math, while the interest-rate effects of the data center boom, formal-methods safety work, and benchmarks for real professional work show where it still falls short.

Watch on YouTube ↗

Charlie's summary

  1. AI can compose with real instruments, not just generate audio, so verifiable-reward domains are no longer the last wall

    Pash gave an agent clips, a Reaper API and open sample sites. It chose the clips, the cuts and the narrative, and played the instruments without hearing them. Pash only added a tempo speed-up whenever Terence Tao says it's going too fast. Nathan takes this as evidence that generalization beyond verifiable tasks is strong: he doubts there is much music feedback loop in training, yet results improve anyway.

    ▶ 2:43
  2. OpenAI's batch of 722 solved hard math problems is a landmark, and humans are already milking its techniques

    Pash says a pre-release model running ChatGPT Pro, about three hours per problem, solved roughly 20% of about 4,000 hard problems. Some of the results stood since the 1800s, and no single mathematician can judge them all. Integer multiplication beat its old limit by a tiny margin, and humans improved on it within hours using the revealed techniques. Pash expects low-hanging fruit in physics and bio too.

    ▶ 17:07
  3. Soaring long-bond yields signal that data center returns are crowding out other capital

    Nathan points to the US 30-year bond approaching 6%, up half a point in a month. He cites the Patel/Dylan argument that frontier labs can recoup a new data center in about a year, then earn for four more. Treasuries and everyone else must pay up to compete. Nathan sees this lending credence to the permanent-underclass worry, since cash now earns near-6% risk-free. Pash expects restructurings, and Nathan thinks Congress may act after the midterms.

    ▶ 24:31
  4. Many people are tactical when they think they're being strategic, so Atlas backsolves from the world it wants

    Evan restructured Atlas Computing into Atlas Ignoda after three times finding someone who needed to do a neglected task. He says people name a desired world, then pick the obvious project in front of them. Put out an RFP and everyone wraps their passion project in its language. Atlas instead finds neglected, timely problems, tests interventions, and recruits owners. Evan's premise: as intelligence gets cheap, coordination gets expensive.

    ▶ 37:29

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.