I think people send too much time optimizing the notebook instead of optimizing what goes into it. I've seen beautifully organized notebooks that were never opened again and messy scraps of paper that someone referred back to after years.
I wouldn't be surprised if "the best model" becomes a niche concept over the next few years.
For most production systems the winning architecture may end up being an orchestrator that knows when to call a cheap model when to escalate to a stronger one and when to combine multiple outputs.
That's how I was thinking about it too. At first glance it feels very close to reactive programming with dependency tracking. I'm curious whether the real advantage is the API ergonomics or if there are optimizations under the hood that wouldn't be practical with a typical observable implementation.
One thing i have always linked about Jane street projects is that they tend to package ideas that have existed in research or niche system into something developers can actually use. Even if you never adopt the library, the design docs are usually worth reading.
Constraints like these are such a good way to learn programming.
When you have limited memory and a slow CPU, every variable and every line of code has a purpose. It's hard not to come away with a better appreciation for how software actually work.
I think the real value of NotebookLM was never the podcast itself. It was making long documents feel approachable. Once voice becomes interactive instead of one-way, it starts feeling less like a podcast and more like having partner.
This makes me wonder how much of paying attention is really prioritization rather than filtering everything else out. We probably process far more than we're consciously aware of.
This feels like a more interesting direction than chasing ever larger models.
For a lot of use cases having a capable model that runs entirely on-device is a much bigger win than squeezing out a few extra benchmark points with a model that lives the cloud.
I like that this treats as a UX problem not just a technical one.
The challenge is making sure people know the feature exists before they need it. An escape mechanism isn't very useful if you only discover it after you're already in a stressful situation.