iiinova — where your digital self keeps evolving

Huh, that's pretty good. How is it on smarts?

Updates when you have them please

I would avoid anything that didn't come out in the last year or so if you want it to call tools and be an agent at all

If it's low bandwidth you should be able to use remote models just fine, they're only streaming tiny amounts of text in and out. Droppy connections will suck though

020

What can one run there? Recently came into (via great personal expenditure) an M5 with 48GB

Ooh yeah you probably can't get the context window too long either can you?

Actually, you might want to check out OpenCode Zen in particular. They often/usually have one or two free models, frequently "stealth" ones with silly codenames, which are most often Chinese labs testing out new offerings before actually deploying them

You might have a better time if you look into running some of the open-weights models via e.g. OpenRouter, or OpenCode Zen. You shouldn't do this if you have concerns about sending data or money abroad. But, you can get lots of usage for dirt cheap compared to American frontier lab offerings

130

The RAM is a big limiter; if you have a 6B model quantized to 4 bits, that's 12GB just of weights already. But maybe there's some clever setup with expert cacheing that could work for you

130

That is indeed a nice laptop! But even smaller models are pretty heavyweight to run. If you want compute/memory intensive things to happen on the same machine as where you're running the model, it's probably not going to work very well, too much resource competition

130

I think you're going to have a hard time with that RAM running anything tremendously intelligent at a high token rate, but I might be wrong! Ironically, best way to get started is probably use Claude Code or Codex and get it to figure out what all needs to get done

150

What I'm saying is you might be Ajax

I was very surprised in one of those recent scrape your neighborhood toys to find that there are people who post significantly more than you, but I forget who or whose scraper

Forgot where that one came from

010

? I think with causal attention, the usual way of doing things lets you reconstruct identical KV cache from just the tokens in context, so you get exact resumes with trivial storage. If you need to save the cache, your storage reqs go up a lot Probably fine if running it yourself but hard at scale

111

I refuse to accept that chartreuse is green. That's a red color's name

021

This would mean you can't resume state exactly without keeping the actual cache around, vs just the transcript

We are all trapped within/out this Klein bottle together (and also all the other ones)

010

Strictly speaking, because they are non-orientable, there is no sense of inside/outside of al Klein bottle

Updated opinion: torus and orb are slightly different games because on torus you have two distinct fronts that you sometimes need to switch back and forth between. Klein and torus still functionally the same, I think

I guess you could have non-planar adjacency graphs, which is what falls out naturally from higher-dimensional minesweeper, but that seems like a UI nightmare

Agreed. I think what functionally changes/improves this versus standard rectilinear minesweeper is that the local adjacency graph varies. I wonder if there's any other generalization that similarly makes it more interesting; different manifolds are fun but I don't think they're changing the game

220

Okay that actually made the game significantly easier, or at least feels easier to play even if my empirical win rate isn't going up Also: I'm not sure Klein and torus are functionally different games, now that I've played them each

Oh yeah huh, that is definitely the right solution

Quantian @quantian.bsky.social · 03/10/2026

Do not attempt, professional drivers on a closed course

030

Reposted by Dan

This was going around Twitter today, if you ask Claude “is X lat Y long land or water?” and plot the results, it has clearly embedded a map of the globe in its weights over time

512715

Reposted by Dan

Minor Mobius @minormobius.bsky.social · 02/10/2026

double torus minesweeper live, only really usable in the hyperbolic map view

1141

Reposted by Dan

Minor Mobius @minormobius.bsky.social · 02/10/2026

Torusweeper and Kleinsweeper are live. It only reeeeeally works if you fix perspective and slip the game over the surface

119611

How have you not asked for minesweeper on a time cool yet? Are you well

Careful; one is only really unambiguously inside the topology conference when all doors and windows are shut

053

As far as topology is concerned, everything is the same on either side of the process

Busy with his Harry Potter club, got it

030

Where did you go to school again

I have no idea what you're talking about, but in response to earlier question: pants have two holes unless you unzip the fly, in which case they have three until you also unbutton the button

330

Maybe: I can take some shape and draw a bunch of nodes on it and draw edges between them. This defines a minesweeper. Topology is invariant as long as I only stretch and squish the shape, and never cut and sew. So I don't have to break any of the edges, and the game doesn't change, either

000

But actually it's because minesweeper is defined from an adjacency graph

Left as an exercise to the reader, of course

100

Commutes with everything, including good taste

140

2 is the "Millennial gray" of Euler characteristics

240