Neuro-sama (the LLM-based AI VTuber) has beaten the game some time ago. As the code isn’t open it’s not possible to confirm whether the StS AI was done with LLMs though. Would definitiely be interesting to see how frontier LLMs perform!
Tuomas Tynkkynen
Karma: 16
This run did have some more things different to previous runs, which I think may be somewhat significant:
By coincidence Claude never got a Pokemon that knew Dig—when he acquired the Dig TM, he already had a full party, of which none could learn Dig at all, and later tossed the TM to make inventory space. I believe in Victory Road digging would reset all progress, thus creating a risk of Claude getting more impatient due to having to solve the puzzles again and deciding to dig out without any actual progress. Though this time Claude was able to make good notes about boulder-switch puzzle solutions and he was able to efficiently re-solve each puzzle multiple times anyway (even though that being unnecessary). Thus Dig probably would not have been a run killer, just a time-waster.
The harness had another change: pressing ‘A’ multiple times in a single reasoning step was now allowed during dialogs and menus. While this did cause new issues sometimes it did help with random encounters in Victory Road as running away can now be done with less steps, thus avoiding context space waste involving random battles. The actual importance of this harness improvement is harder to predict though, but I do feel it had bigger impact than the zooming tool that didn’t really get that much use in the critical parts.