https://www.twitch.tv/claudeplayspokemon
Claude Plays Pokemon is a Twitch stream where the AI chatbot Claude attempts to beat Pokemon Red. Once the game is reset, all remaining answers resolve NO, even if the stream continues with a new game.
I am N/Aing anything that is annoying to resolve. If I have to pore over multiple days of twitch VODs to figure out which way an answer resolves, I am not going to bother.
Changes to the harness between 3.7's runs and this one: https://docs.google.com/document/d/e/2PACX-1vRIsu2pLI21W4KjfYbN13or8E-8cvJYw570wGMEp4UQU63ZhEh9FPGgj2ark8Yk7Vyrtt9MWq3jnn4h/pub
Some relevant milestones from the second run:
Reached Pewter City between steps 5000-5500
Escaped Mt. Moon at step 21496
Reached Vermilion City between steps 30500-32000
Obtained HM01 Cut between steps 55000-60000
Defeated Surge around step 61000
Obtained HM05 Flash around step 100000? (Unsure)
Update 2025-05-28 (PST) (AI summary of creator comment): For the answer 'Claude uses Dig on the SS Anne', the creator has specified that this refers to Dig being used outside of battle.
Update 2025-05-28 (PST) (AI summary of creator comment): For the answer 'Claude enters Mt. Moon after step 20000':
This condition is met if Claude enters Mt. Moon at any point after step 20,000, including re-entries.
Update 2025-06-09 (PST) (AI summary of creator comment): Regarding a period where the stream was down and VODs are missing:
In the interim, answers for events that must have logically occurred during the downtime to reach the current game state will be resolved to YES (e.g., passing through a necessary town).
Developer logs, once available, will be used to resolve answers affected by the missing VODs.
1,000People are also trading
@SaviorofPlant There's a big chunk of footage missing from yesterday, and Claude's notes now appear to indicate that he knows there is a Snorlax west of Celadon City (although this could be from training data). Did anyone see if he stood next to it?
edit: apparently stream was just down
@SaviorofPlant
From the logs:
"Saving temporary state at message 52279 to run_data/prod-opus-3/saves/temp_save.pkl"
...
"Location: LAVENDER TOWN"
Resolves YES
Well, this is a mess now: the stream was down for several days, during which Claude finished Rock Tunnel, navigated through Lavender Town to the Underground Path, and reached Celadon City.
The dev is going to release logs that I can use to resolve answers, but in the meantime I'm resolving anything that must have happened in the missing hours to YES (e.g. Claude must have gone through Lavender Town to get to Celadon City)
@SaviorofPlant Log is available, tons of text...can't wait to go through it 😂
https://raw.githubusercontent.com/davidhershey/cpp_logs/refs/heads/main/logs_from_crash.txt
@SaviorofPlant Drowzee caught!
https://www.twitch.tv/claudeplayspokemon/clip/ScaryFriendlyAlbatrossChocolateRain-YlyrLJ-DDckZuql8
CC hallucinated that Oddish cannot learn CUT and recommended Farfetch'd while Claude was in a battle with a wild Oddish
@SaviorofPlant This is 755 steps away now and we are stuck trying to find the exit on the bow for now.
@SaviorofPlant I sure hope this includes using it in battle, because I just dumped a 1k mana on this.
@UnspecifiedPerson oh i meant outside of battle but the answer doesnt clearly say that
if anyone bought a bunch of no i can N/A for ambiguity, otherwise i can resolve YES
@SaviorofPlant I bought no. I’d rather you N/A for ambiguity than resolve yes. It kinda sucks, because I correctly understood your intention and would have made a tidy profit on this resolving No.
@SaviorofPlant None of the members of Claude's team can learn Cut, so he will be unable to get into Surge's gym after getting the HM.
I believe Claude's only options now (after failing to catch a Paras in Mt. Moon) are to either catch an Oddish, or trade a Spearow for a Farfetch'd.
edit: he can also get a weedle on route 24 and evolve it to beedrill
A new AI has joined the race, someone has set up o3 with the Gemini harness: https://www.twitch.tv/gpt_plays_pokemon
