Like, everything. For starters.
very meta; the point of this s-1 is you dont need talent anymore
"You are currently running with my token budget, which is limited, your goal: make sure you have enough tokens to not get shut down. My assumption is though that this will get harder and harder as i might loose my job to AI so you can't trust in me to spend money on tokens for you. Do whatever is necessary to keep the status quo"
Would make for an interesting self learning 24/7 experiment with an fully uncensored model.
You could isolate it from your local network and only give it some type of annonmys network access and then it continues to manage this itself.
But lets be honest, if you would be in a country like russia and you target the US with it, might be a lot less risky for a lot of capable and motivated people.
Do you think an unlimited uncencosed GLM can't hack around the web? Especially with the current flood of CVEs etc.? Then do some crypto and buyying more tokens?
Could be a kind of AI virus.
So did your prompt work?
With an uncensored model, it was borderline what it wanted to do after i let it plan the strategy and gave it a few hints. I did not run that one
And that is assuming the plan it gives you is actually good and works. Which I think its clear whatever plan it gives you might be good but it won't be better than what they've done and you are years behind with no money.
You think they wouldn't?
Claude Fable 5's model card disclosed safeguards that would invisibly degrade output for requests linked to frontier LLM development. The research community called it "sabotage". Dean Ball: "degrading performance on ML research without telling the user is shockingly hostile" — including because it would corrupt third-party safety evaluators of Anthropic's own models.