I'm not sure the authors realize this is way more than "just a cool demo": if this holds up, it's going to be huge for game QA work.
Instrument your game to output properties of entities near the player and the output is the various control inputs - moment to moment gameplay gets solved. Maybe augment with a tick-by-tick controlled stepping mode if particularly twitchy - an LLM can take care of the higher level reasoning then.
You can "solve" this by throwing your PDF statements at llms and having them extract the transactions. Has privacy implications if you use hosted llms of course.
As for self hosted accounting/budgeting software, IMO Beancount is the best by far: nothing to host, just text files.
This is partially a "how much does the model follow instructions" thing. I use Pi with a vibed LSP extension and Claude (4.6 or thereabouts) almost never followed instructions to use LSP renaming tools - despite it being strongly emphasized in system prompt and agents.md. However I found Codex 5.3 would use them sometimes, and GPT 5.4/5.5 would prefer them.
It opens the list of filenames in a given directory (or set of files passed on cmdline) in an editor of your choice, and then you use your editor to rename them; the changes get applied when your editor is closed.
There was/is a patch available that lifted the unit limit significantly. I played with it constantly. Even played a match at a lan party with friends last year.
Well language detection isn't all that hard in the scheme of things (especially now), but maybe having only training on English makes models less effective programmers. It would be interesting to see that as an experiment.
I mean I am pretty close to leaving the internet as I know it entirely. Seems like just in the last year there has been a lot more LLM slop articles posted directly to HN. It is getting to a "what is even the point" point for me pretty fast. My use at this point is essentially habitual and akin to quitting cigarettes than me still squeezing any actual value out of the internet as I might have in years past.
This is also what I've done for my multiplayer falling sand game: it's very much not vibe coded (too performance-sensitive), but coding agents can launch the game on my steamdeck and run benchmarks, take captures & verify rendering is bit-for-bit identical on a given machine, etc.
Agent can't _play_ the game yet, but that's on my list to experiment with.
[1]: unless you use multiple glyphs for a char. Or maybe it'd be possible to interpret if the glyph-to-char mapping pattern was very carefully thought out..
Initially I read it because just it was interesting but it has ended up being the harness I have stuck with - pi is well designed, nicely extensible and supports many model provider APIs. Though sadly gemini and claude's subscriptions can't really be used with it anymore thanks to openclaw.
Instrument your game to output properties of entities near the player and the output is the various control inputs - moment to moment gameplay gets solved. Maybe augment with a tick-by-tick controlled stepping mode if particularly twitchy - an LLM can take care of the higher level reasoning then.
reply