Linear Digressions

Anthropic just announced they're baking invisible watermarks directly into Claude's generated text — and while everyone else was busy having opinions about it, we were busy asking the more interesting question: how does it actually work? Turns out it's not hidden Unicode characters or first-letter secret codes — it's something far more elegant, operating at the level of word choice itself. We dig into Google DeepMind's SynthID text approach, published in *Nature* in 2024, to understand the clever statistical machinery behind watermarking language model outputs without anyone being the wiser.

Humanity's Last Exam was designed with a bold premise: questions that human experts can answer, but AI models can't. Originally dubbed "Humanity's Last Stand," this benchmark is a massive academic collaboration — hundreds of contributors, thousands of fiendishly hard questions spanning a wild range of domains. In this Better Know a Benchmark installment, we unpack what HLE is actually testing, how it was built, and what it means when a model finally starts cracking it.

When a language model tells you it's absolutely certain, is it actually more likely to be right? Kaitlyn Zhou's research says: not necessarily — sometimes confident phrasing correlates with *worse* accuracy, echoing a very human Dunning-Kruger effect. In this conversation, Kaitlyn (soon an assistant professor at Cornell) walks through why LLMs talk this way in the first place — tracing the tendency back through training data and the RLHF annotation process, where it turns out humans don't love confidence so much as they punish uncertainty — and what that does to the person on the other end of the chat window, who turns out to rely on confident (and even flatly-stated) answers far more than they should. We also get into her newer work on voice cloning, and how a cloned voice can sound more "native" and more trustworthy than the real one it's based on.

Reasoning models don't just answer your question — they *think out loud* first. In this episode we dig into the class of AI models that generate intermediate chains of thought before arriving at a final answer, exploring how the internal reasoning process works. Are these models genuinely "thinking," or is something else going on under the hood?

This week we’re covering model distillation: the technique of using a large "teacher" model's outputs to train a smaller, cheaper "student" model that mimics it. They cover the two big reasons labs do this — making lighter, faster, more focused models for specific tasks, and the more contentious use case of effectively copying a rival's flagship model by hammering its API with questions (with a callback to the old Bing/Google search controversy). They also get into why it's so hard to prove distillation happened, why some models occasionally introduce themselves as "Claude," and a surprisingly old idea: a 2015 paper by Geoffrey Hinton, Jeff Dean, and Oriol Vinyals on distilling knowledge using the full probability distribution over a model's outputs — not just its single most likely answer — and what that "soft label" approach captures about how a model relates concepts to each other.