alt Hacker News> There is a kill switch
...because this thing will go rogue faster than you can blink.
I swear, it's like nobody at the company even reads the slop they're generating or thinks about it for any amount of time. In what world is advertising a kill switch as one of its essential features a positive? It's basically admitting from the start that this is unreliable.
They replaced their production staff with clawbot, it's all part of the plan.
There's a sense of "early bitcoin" around clawbot and other agent frameworks. I think if you wait for another 2 years for it to mature, you'll have missed out as if you waited ten years after bitcoin began.
They're insecure and janky, sure, but on the other hand you've got millions of dollars of compute and tens of thousands of very motivated developers working on making them secure, reliable, and competent. There's something magical about AI that actually gets real work done while you're doing other things, and that's what Perplexity is probably hoping to sell.
Just need a reliable local model, though - AirLLM, other hacks allow you to run bigger models more slowly, so you can build out a completely API-free scheme to run pretty capable agents even without big GPUs.
Could be a Moravec's paradox thing - all these people are thinking that the solution looks enticingly within reach, but it might be an absolutely horribly complicated quagmire with no easy solution short of AGI. I'd bet on clawbots and agents being very secure and great to work with in the very near term, though.