I had to start Strange like three times. It helps that I'm generally a fan of that historical period but the sheer DENSITY of the thing put me off until I really set myself.

And you know … it was okay. Not great, not bad. Worth having read. Probably won't reread.

I enjoyed the book VERY much, but I was let down by I'd guess the last quarter or so. I wanted the story to end more on the wonder end of the wonder-mundanity spectrum.

Have you read Jonathan Strange? Or I should say, like many people, have you started Jonathan Strange?

*may have meant to say orthographic there instead of orthogonal; not 100% sure

I use UMAP to visualize a growing set of vectors sometimes, so I always get a different result, but for example, I can always find the cluster about <whatever topic> because they're always off by themselves in a little arc or whatever. So yes, I think there'd be convergence. That's my guess! 🤷‍♂️

"People are statistically more likely to move house around retirement age" is a statement I'd feel pretty confident about making without evidence to back me up.

Hm. I didn't think about that. My use case does not involve file uploading, so my "I do it this way" MCP suggestion was perhaps out of place. Skill makes infinitely more sense, doesn't it, since you have a whole VM to play with. You could create a document and then … yeah. Hit a forbidden API. 😠

The examples I've seen of really good uses of graphs databases have all been contrived and hard for me to generalize in my mind. Like who killed who on Game of Thrones. Combinatorial stuff. Maybe there are other really good uses but 🤷‍♂️.

I can think of a hard solution. Remote MCP server running in the cloud, maybe on fastmcp.cloud (not an endorsement just something I've heard of) or on like a Raspberry Pi. Should be able to do anything an API client can do, plus works with claude.ai. I have a similar setup for something right now.

If you find a solution and you feel like posting a follow-up, that would be great. I've had this idea in the back of my mind of finding a way to integrate Claude into my Gmail more tightly (because I hate my Gmail), but if they're blocking API endpoints, oh bother.

Is it mitigating or exacerbating that I had not at that time ever heard of Dr. Sutton or the Bitter Lesson? I'm just a dilettante who picks up little bits as I go. At the time I thought the more structured the data was, the better the AI would do with it. I didn't understand the middle L in LLM yet.

I think we're going to see this more and more. I know I've started thinking of tasks in terms of one window or more than one window. If it's more than one, I try to break it down into sub-tasks.

Gosh, if Anthropic rolled out 1M for Claude Max users…

Sorry if this is unhelpful, but have you checked your Claude capabilities settings? There's a setting in there that I believe says Claude only has limited Internet access. You can change it to unlimited. That might be what's blocking you.

"Oh James"

Since the model's last hidden state has to POINT AT the token cloud to generate a coherent output, it really makes me think about how little of the sky this model uses for distinguishability. Especially in bfloat16, cosines only have so many digits.

Soon I hope to get back to mapping the sky. 😁

Cheers for that! I hope to get back to them soon. Meantime, here's a view from the origin of coordinates of the Qwen 3 4B model I've been using, showing the medium-dense token cloud and then about 26° south of it, the tiny overdensity that's been so mysterious to me.

Also, just to clear the record a bit, different movies are not literally connected to each other with internal continuity. That's an illusion created by the filmmakers. In reality, you're watching photographs flicker by 24 times a second. The point is to smile and laugh. Everything's gonna be okay.

That would explain the endless series of "companions."