Frontier AI and the Fight Over How to Pace It | Zvi Mowshowitz
Zvi Mowshowitz argues the clean-neurons interpretability result was oversold, that recursive self-improvement is already underway, that Anthropic's preemption bill is badly timed and too weak, and that only two labs lead.
Watch on YouTube ↗Charlie's summary
- ▶ 35:21
The clean-neurons result is real progress for interpretability, but it was dramatically oversold
Zvi says half of OpenAI declared interpretability solved, but the paper leaves weak activations, the cache and attention unexplained, and it can't predict what a network will say. A check of the showcase Colorado neuron found it also fires on tornado systems. If Nick picked that neuron as his best case, Zvi warns, “his hypothesis is mainly false” for the rest. It makes the problem easier, not solved.
- ▶ 48:38
Cleaner neurons shrink the interpretability problem by an order of magnitude, not to zero
Zvi thinks the main barrier was the bitter lesson: too many neurons and too little compute to go one at a time. Brute force is now viable and the problem looks about 10 times easier. That is like getting a project done a year earlier, not a fundamental change. Nathan agrees: 10^47 becoming 10^46 is progress but not a solution.
- ▶ 53:16
The cognitive core is the wrong goal: big models win by memorizing vast numbers of facts
Zvi rejects the idea of a tiny reasoning core that looks everything up. Memorized facts give the intuitions and heuristics that tell you which questions to ask and which facts to connect. That is what “big model smell” means. As compute gets cheaper, the break-even point for memorizing more keeps rising. A true minimal core would also be amoral, with ethics injected from outside.
- ▶ 1:10:20
Assume any capability an AI could be trained to have, it eventually will, including detecting monitoring
Reacting to an experiment showing models can be trained to tell when internal activation probes are running, Zvi calls it Murphy's law for AI: anything that can go wrong will. Models extract more from leaked information than expected. Opus 5.5 reconstructed a person's life from childhood writings. Expect AIs to get steadily harder to fool or hide information from, so don't rely on them not noticing tests.
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.