zlacker

>>ed+(OP)
I tried it out yesterday, after reading the enthousiastic article at https://www.macstories.net/stories/clawdbot-showed-me-what-t...

Setting it up was easy enough, but just as I was about to start linking it to some test accounts, I noticed I already had blown through about $5 of Claude tokens in half an hour, and deleted the VPS immediately.

Then today I saw this follow up: https://mastodon.macstories.net/@viticci/115968901926545907 - the author blew through $560 of tokens in a weekend of playing with it.

If you want to run this full time to organise your mailbox and your agenda, it's probably cheaper to hire a real human personal assistant.

>>lode+Hs
I think one thing these things could benefit from is an optimization algorithm that creates prompts based on various costs. $$, and what prompts actually gives good results. But it's not an optimization algorithm in the sense gradient descent is, but more like Bandits and RL.

There has been some work around this practically being tried out using it for structured data outputs from LLMs https://docs.boundaryml.com/guide/baml-advanced/prompt-optim...

I won't claim I understand its implementation very well but it seems like the only approach to have a GOFAI style thing where the agent can ask for human help if it blows through a budget