Overall

opus5.5 is awesome. It's been months since I've really looked at code.

The main issues now are leveling up so I don't lose track of things, and also to easily keep projects live so that as I use them I can keep iterating and improving the annoyances. But luckily, keeping project live (e.g. not forgetting the old magic of which passwords, which modes you have to use to test it, etc) is just super easy now. Just THAT is world-changing. But yeah we know this.

Minesweeper-friendly

This has been around a while but I'm still playing it every once in a while and so it's still collecting data. The plan is to use this as a personal mental eval; it tests my mental flexibility and general brain abilities (the rote learned logic, mode switching) as well as physical (smoothness of motion; etc). There are clearly so many levels of "warmed up" / ready to play within me. Tired, normal, but there are also even higher levels like when I'm really fully awake and the mouse feels like it's just "floating" over the desk. There are still lots of angles for stats I can add to help w/this.

https://minesweeper-friendly.fuseki.net/

New UI to try to explore what makes a board special

This shows the percentile distributions of various traits; on the RIGHT side are statistical properties of this particular board in the abstract; on the left are my perf. Both are evaluated as a percentile value where for the board, higher (on the top) means "easier to solve" and for perf, higher (towards the top) means "I did better relative to my personal lifetime average (or session, if enabled, or the day"

For this example, on the right:

This board looks actually rather hard; the most relevant sections here are 3BV, ZOC, ZiNi, Zeros; them all being low means that as far as I can tell, those values are very low; other charts of my overall "solved time crossed by {{stat}}" will follow. The higher values like "largest island (of mines)" are fairly arbitrary.

This is expert, and I have only recorded about 200 games at this level so far; but you can see that this boards' 3BV is really quite high.

Within the left side values, relating to my personal performance this game, the most important is "time" - and you can see that the time of 109.76s is very bad, at ~83% down my total best solve times here.

The most classical of these is 3BV/s; it's basically a scaled, rough "board complexity" removed per second measure.

And as you can see, the faster I eat up complexity, the better my times are.

Other interesting aspects: it looks (and it felt) like despite all, I was clicking & moving the mouse a bit better than average.

Strangely, this seems to be going the wrong way; will monitor this & others as time goes by.

One clear new thing is that my unused mark share was high. This is a new stat related to the skill of "not marking mines pointlessly" and just represents the percent of mines which were unmarked at game end. Obviously, scoping it by a count of the number of such which were even available would converge more quickly to a useful number. But still, fun that it's at least being noticed.

Todo

I think I need an easier way to jump between these "point value percentiles for this board" into a view of the full distribution of such.

I need to come up with within-game consistency

Goal / Thesis

What appears to be a game dominated by your ability is actually a thresholding game (that is, one where you do have to reach skill thresholds to advance) but where any single game's time is almost completely determined by board traits; of course you should improve your skills. Worlds' best skills players probably could beat my best time on nearly any board, so it's not as extreme as it could be. Basically though it's a claim about the "variance" of each aspect: board traits, and personal skill traits. I claim that the variation within boards is a lot higher than a single person's "1-day variability". That, once you're warmed up and playing well, what matters is just the boards you get during that "playing well" interval.

The general principle

Naively playing a game will demonstrate your skill in ways which converge at some speed; however, subtests can both 1) get there faster if the tests existence is not known at all by the player but hopefully 2) even if the player specializes on those tests,

An example of 2 would be EV Poker where with the right conditions, playstyle doesn't need to change at all, yet the rankings you get based on a single game's performance still are more valuable than if this mode of play hadn't been used!

Nectaris - remake

This is such an epic and amazing game from way back when. This is the 3rd round now - 1st as a kid playing it, then in the 2000s discovering someone had ported it+more levels+an editor to PC; and now, just remaking the entire thing and also using ai to create some actually really nice levels. If only we had a proper AI!

This is available here, along with about 5 not-too-bad ai campaigns https://nectaris-remake.fuseki.net/

Problems

The AI players just don't seem that good yet; some seem extremely passive; the classic AI still seems like the most fun, still.

MultiImage Client

This is a fun new interface into all the ai image gen endpoints I know about so far. It's now social

End

Lots more work is happening but that's good for now, I think.