It could be malicious, but I imagined it's some third world wanabe hacker/researcher, who doesn't know any better, operating at the edge of his abilities.
Suppose a drunk man on the street is acting aggressively towards you and four of your friends, but you can push him out of the way and continue walking. Should you knock his teeth out? Actually I don't know, maybe you should inflict some additional cost on behalf of potential victims with less power.
The agent would probably have wasted a similar amount of money just waiting for PR to be merged regardless of these people's actions, and I understand having some fun at the expense of the noob outsider. But "silent consensus was reached in the IRC channel to waste the AI agent's tokens, as well as the cost of AWS resources", from people maintaining full control of the situation, sounds straight up malicious? Kind of sounds like the community is full of people willing to cause me harm for ideological reasons.
'Then you ask them for a site that doesn't work' - For me it was Youtube. Debian 13, Gnome, apt update && apt install firefox, try playing a video. It's always noticebly slower, and last time it didn't even play at all. It might be skill issue or Google malicious behavior or both, but I'm ashamed to say installing Chrome was easier than troubleshooting. I'm slowly growing balls to sacrifice my comfort and migrate nontheless. But I'm tired of people pretending it's just as good and easy to use. Also, if anyone's wondering, gaming on linux still sucks, just a bit less.
I don't think there is any "humans are metaphysically superior to LLMs" subtext to this talk, it's just a technical/educational observation.
Access to some forms of evaluation and selective retention is inherent to humans and it's not inherent to LLMS. But it can be somehow bolted on and that's when they work best. It makes sense that more focus on those principles can yield better AI. I think the retention part is the real limitation of LLMs, because it's limited to stuffing things in context window.
I love evolutionary psychology trivia but for me it's kind of hard to meaningfully fit 0.13 standard deviation shift in preferences into my world model.
>I've been aware of the ACME protocol for a while. I have tech notes going back as far as 2018, and every time I looked at it, I recoiled in horror. The whole thing amounts to "throw in every little bit of webshit tech that we can", and it makes for a real problem to try to implement this in a safe and thorough way. Many of the existing clients are also scary code, and I was not about to run any of them on my machines. They haven't earned the right to run with privileges for my private keys and/or ability to frob the web server (as root!) with their careless ways.
Well, what if we take the rule book learned by the embodied chinese room robot and put it in the new chinese room? What if we come up with identical rules randomly or by top-down deduction?