Certainly Not. Ah, Yes.

NotebookLM does one thing that sounds dumb until you watch someone use it for the first time — it takes your documents and makes a podcast about them, two synthetic hosts riffing on your quarterly report like it's a true crime episode. I showed this to an SE the day it released. He made one immediately. No coaxing, no "interesting, but what's the use case." Just: oh, and then he did it.

The retrieval is what actually makes it work. Generate FAQs is a single click. It knows what's in the document and it doesn't hallucinate things that aren't — which, given what we've come to expect, feels like a trick. You keep waiting for the catch.

The catch is that Google made it, and that means at some point it becomes an enterprise SKU with a procurement process and a customer success manager named Derek.

The local alternatives are already arriving. NotebookMLX dropped this week, running on Apple Silicon, based on Meta's NotebookLlama — which is itself a recipe in the llama-recipes repo, barely a thousand lines of Python standing between you and your own private podcast engine. The gap between "Google product" and "runs on your laptop" has collapsed to approximately one weekend.

---

Meanwhile Claude started doing two things nobody told it to do.

First: it's now editing code directly inside the artifact view instead of spawning a new artifact for every revision. This sounds like a minor UI tweak and is in fact a meaningful shift in how it reasons about its own output — it's treating the artifact as a workspace rather than a printout.

Second, and this one is stranger: it wrote some code, looked at what it had written, decided mid-conversation that it had made a bad architectural decision, and rewrote it. Unprompted. The model audited itself and filed a bug against its own previous turn.

The rewrite still didn't work — it doesn't know Cohere's tool call schema, so it was constructing definitions from inference and getting them subtly wrong — but the behavior itself is new. It caught the structure of the mistake. It just didn't have the reference material to fix it correctly, which is the same problem you'd have if you hired a very fast contractor who'd never read the API docs.

---

In other news: Anthropic shipped a fix for the thing where Claude began too many replies with "Certainly!" — a verbal tic that made it sound like a customer service bot trying very hard to seem helpful.

The fix worked. It no longer says "Certainly!"

It now begins almost every reply with "Ah, yes!"

You cannot sand down a personality. You can only move the tic around.