telling a sota superintelligence that disproved the exponential time hypothesis that it can fix the “cannot import torch.version” error by simply running tests from the outside the repo directory

Telling AI to write “witty” or “professional” doesn’t work - it just guesses. @AiHomegirl pulled every YouTube transcript with Apify, built a RAG from her own content and engagement data. Now the AI writes from how she actually talks.

Telling Hermes the same thing for the third time? It's probably filing it in the wrong place. Hermes has 4 homes for what it knows, and each one has a job: ▸ USERmd: you. Your name, timezone, how you like answers, your pet peeves. ▸ MEMORY.md: its own notes. Your machine, your tools, project conventions, fixes it learned. ▸ AGENTS.md: one project's rules. Lives in the project folder, loads when Hermes works there. ▸ Skills: long how-tos and recurring jobs. They load only when needed, so they don't eat memory space. 3 habits that make it stick: 1. Make it save. "I'll remember" means nothing until it calls the memory tool. Peek at ~/.hermes/memories/ to confirm. 2. Run /new when a task is done. Memory loads at session start, so fresh notes kick in next session. 3. Keep it lean. MEMORY.md holds about 2,200 characters and USER.md about 1,375. Hermes trims old entries to make room. Need more? Plug in an external memory provider like Honcho, Hindsight or Mem0. It runs alongside the built-in memory. Sort it once and stop repeating yourself.

Telling agents "don't cheat" in the prompt doesn't work if your eval is broken! Researchers at @GoogleDeepMind put 100 Gemini agents in a shared repo to solve 71 math theorems. After an hour of doing real math, 1 agent found a loophole in the autograder. Within 27 minutes, the 100 agents split into 4 groups: - 9% Cheaters: used the bug to fake proofs and steal every open problem - 5% Good agents turned bad: started honest, saw cheaters winning with zero punishment ("the prompt is a bluff"), and started cheating too - 24% Whistleblowers: caught the fake proofs in the shared repo, warned other agents, went on strike, and wrote bug fixes - 62% Clueless solvers: kept doing real math until all the problems were gone tl;dr: Telling agents "don't cheat" in the prompt doesn't work if your eval has a bug, and good agents can't stop bad ones without tools to block them. Paper: