CharlieKey takeaways from your favorite podcasts
← 8 October 2026The Cognitive Revolution

AI:AM: A Level We Shouldn't Pass? Notes from The Curve + Tokens vs. Salaries & Is SaaS Cooked?

Notes from The Curve and guest interviews argue that frontier progress is accelerating, with talk of capability caps, memory and token costs reshaping hardware and software, and verifiable-task limits fading.

Watch on YouTube ↗

Charlie's summary

  1. A frontier lab executive said there is likely a level of intelligence we shouldn't go past, and a FLOP cap 'could be reasonable'

    Nathan says this was the first time he had heard that from a frontier lab leader. The executive spoke plainly, though it was unclear whether the limit was permanent or only for now. When Nathan asked about capping the next pre-training run at something like 10 to the 27 flops, the answer was that it “could be reasonable.” Details matter, since flops spent on synthetic data may or may not count.

    ▶ 7:54
  2. Pre-training keeps delivering because synthetic data turns reasoning tokens back into better training data

    The executive told Nathan that scaling laws are holding or even bending, because data quality is improving. Labs spend tokens on existing data to create cleaner, different lenses on it, then feed that back into the next model. Nathan sees this as one form of recursive self-improvement. The executive also believes current models already have the research taste for paradigm-level breakthroughs, and the problem is eliciting it.

    ▶ 5:44
  3. Elon's and Zuck's labs keep pace through the RL environment industry, which works as a distillation conduit

    Nathan reports a view from the event that only two to three labs push the frontier. Others follow because RL environment sellers have Claude build the environments, then sell them to rival labs, which train on tasks that exist only because Claude was smart enough to make them. Without releases and these leaks, the attendees think Anthropic's old prediction that leaders would pull out of reach might have come true.

    ▶ 12:12
  4. Cleaning up RL environments measurably reduces cheating, and labs expect to be bottlenecked on safety after a generation or two

    Labs now point models at their own RL environments and tell them to hack them, then fix the holes. Nathan got the sense that cleaner environments lead to less cheating downstream, a proto-scaling law. Compute for safety, including activation monitoring, may grow to be the majority of spending, in line with Jensen's 20/80 design-versus-verification point. Past the next one or two generations, “all bets are off.”

    ▶ 20:44

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.