How Does Claude 4 Think? – Sholto Douglas and Trenton Bricken

First, the claim that one of the main things holding LLMs back is a lack of expert feedback. To me, that just means the models are guessing—because they don’t have knowledge like humans do, they rely on pattern-matching, not understanding. If the user doesn’t know the answer, the LLM can’t help. That’s not just a minor limitation—it’s foundational. Framing it as a feedback issue is a way of sidestepping the deeper problem.

Second, the speculation about Claude winning a Pulitzer or a Nobel Prize. I get the underlying point—they're wondering whether LLMs are better at creative or scientific work. But couching it in terms of prestigious awards just adds to the hype. Why not just say “creative vs. scientific tasks”? Framing it as “what will it win first?” cheapens what those prizes represent and makes the model seem far more capable than it actually is.

Third, one of them claims a friend at a drug company says they’re about to release a drug discovered by AI. But when pressed for details, it turns out to be pure hearsay. It’s a textbook example of the kind of vague hype that surrounds LLMs—bold claims with no real substance when you dig.

That said, I appreciate that the host, despite being quite friendly with the guests, actually pushes back and holds them to their claims. That's not very common in AI discussions, and I respect it.

Learning that an earlier Opus was highly committed to animal welfare and they’re not sure why; other versions and models have not been

Hearing that their mechanistic interpretability groups recently beat an internal red team quickly, using Claude’s help

Speculation on de-risking AI growth for nation states that aren’t building frontier models

Real real profound comment near the end noting that even if we get no more algorithmic improvements ever, current tech is sufficient to replace most white collar jobs, given enough data.

That last one is a super profound and interesting point to me. We’re at the inflection point where white collar jobs will at the very least be implemented by this tech and overseen by humans, and the current tech and economics make it desirable to do so for companies that pay for a lot of white collar work.

Feels like it’s time to buckle up. Not AI2027 buckle up, but still time to buckle up.

Interview with the same two people from a year ago: https://www.dwarkesh.com/p/sholto-douglas-trenton-bricken

If you are at all interested in the current challenges being grappled on in this space, this does a great job of illuminating some of them. Many many interesting passages in here and the text transcript has links to relevant papers when their topics are brought up. Really like that aspect and would love to see that done a lot more often.

"the AI labs spent a few years quietly scaling up supervised learning, where the best-case outcome was obvious: an excellent simulator of human text

now they are scaling up reinforcement learning, which is something fundamentally different. and no one knows what happens next"

I tend to believe this. AlphaGo and AlphaZero, which were both trained with RL at scale, led to strategies that have never been seen before. They were also highly specialized neural networks for a very specific task, which is quite different from LLMs, which are quite general in their capabilities. Scaling RL on LLMs could lead to models that have very unpredictable behaviors and properties on a variety of tasks.

This is all going to sound rather hyperbolic - but I think we're living in quite unprecedented times, and I am starting to believe Kurzweil's vision of the Singularity. The next 10-20 years are going to be very unpredictable. I don't quite know what the answer will be, but I believe scaling mechanistic interpretability will probably yield some breakthroughs into how these models approach problems.

Trenton please if you are listening give us an explanation for this!