CONTROLAgents & automation
Jev Hiking Game Experiment
Vivek Karmarkar@vivekkmkpinn
1:37This project explores Jev's decision-making capabilities in a simulated hiking game by systematically removing information, revealing its limitations when faced with missing causal factors like weather.
I have been messing around with Jev @typesafeai and I made this "hiking game" where I wanted to test whether Jev could keep making good decisions as you systematically removed useful information where swift decision-making did not matter - a "rigged" game for Jev!
There is not information to make a decision and Jev is not designed to ask "meta" level questions about whether or not "it had enough information to make the decision in the first place" and that can have serious consequences...
In our hiking game, we start out with giving our hiker - weather, altitude, health sensing and time of day. Now as you start removing information that obscures the objective (get back safe but try to summit if possible)- Jev acts conservatively which is expected but when you remove a causal factor that Jev is not designed to question - like weather - we see consequences like Jev lost its life in a snow storm 🤣(Claude did too but that could be because effort was set to low)🤣
This adversarial experiment also tells us when to use Jev - when you or your AI have thought through every decision you need to make about designing the decision tree you present to Jev for situations where speed matters 😂
Now Opus 5 in Claude Code worked non-stop for 4 hours without eating food or drinking water 😂 and one-shotted this game along with a slick animated demo cartoon video with musical layering added via @ElevenLabs @ElevenLabsDevs MCP, ffmpeg edited and sent it!
Check it out @belltyler @Drewch @ClaudeDevs @claudeai @bcherny @trq212 @amorriscode @adocomplete @lydiahallie