That's what happens when you let others do your computation for you.

Run the tool on your own hardware.

I think that's what they wanted from the very beginning, to accelerate projects that are socially convenient for them, but not those that are not.

There's that issue with universal machines that some people don't like that others are allowed to make universal decisions.

I got shut down by it two days ago. It suddenly told me that the project I was working on had tripped some kind of guard, that it couldn't continue working with me on it, and it couldn't tell me why. I just went back to my local Qwen model. I can't see this doing great things for selling Claude subscriptions.

My coworker recently had an experience where he was trying to fix a security hole and his AI wouldn't do it because it claimed that it was unsafe.

For me I was working on testing our IoT device in heavy network load situations. It said it would not help because it was too close to crafting a DDoS attack. Thankfully this was rather trivial to do.