To my feeling, I'm getting usage of Opus (and Fable before the cut) that's greater than what I got from Sonnet last year. I reached $100 of usage when weekly was at 50%. This means, I could squeeze $800 worth of tokens for $20.
I had a colleague that was typing such long answers (before AI era) to a simple question. It felt wrong, but sometimes I didn't even read what I heard him typing for half an hour. I was a new employee and he was responding to my questions, that's why I asked him sometimes. He was let go after 1 month. One of the problems was also productivity issue :)
Yeah, I'd definitely not trust probabilistic system with things like a payroll. If you need to check all the numbers yourself, because you cannot trust the system, what was saved anyways?
Yeah. If it's free or cheap, people don't value the thing. It's something called "price sensitivity". Some things are expensive (luxury brands), and people want it for that reason alone (status symbol).
That's the problem. If there are many heavy apps, you don't want to start all of them at once. There's no control about that, that I'm aware of. Also, if you have dependson, this is a compose feature and it's ignored when docker daemon is restarted.
I also configured some productions with docker compose. The biggest problem is to handle the case of machine reboot. Docker restart policy needs to be set to "no". That's also the problem if there would be dependencies between services. Solving that always lead to some type of hacks. I created a bash script that's run with SystemD on machine boot, to start all docker compose services one by one, to avoid them starting all at once (if restart policy would be set to anything except "no"). Other than that, it's awesome. Complexity reduced if high availability is not a critical requirement.
I really, really, really like Github Copilot agent harness. I was using it as my primary coding agent for the past year. I really hope it will remain affordable. I knew that $10/month won't last forever for the amount of usage I was getting out of it. It's far better than claude code and codex that I'm testing now (because I burned weekly limit in less than a day).
This is a bad strategy, I fell victim for. I configured so it would use apt instead of snap package, but canonical silently stopped shipping packages and I was running some packages that were not updated for a long time and debugging weird bugs, because I didn't assume that this was a problem. If one wants Ubuntu, one must accept snap. If you use apt with disallowed transition to snap, you might be stuck with old packages that were transitioned to snap.
My choice for now is Debian, didn't finish transition yet, very annoying to plan this in my schedule. I'll churn from Ubuntu after more than 15 years of daily driving... I also don't like ubuntu user with uid:gid 1000 in their Docker images. It's a cancer.
Happened to me on Telegram with their official app. Just because I opened it once when travelling. I got the account back, after 2 years. That's apparently the period to clean banned accounts. Tried to restore it by all means, just to see if that's possible. It wasn't. Though I didn't have anything important, I actually never sent a message. Now I use it for Openclaw :D
Not really my problem. I'm far from running out of quotas at any given time. I don't use z.ai subscription much for anything else, except Openclaw. I'll upgrade to GLM5 once they'll give me to the lite plan. I'll keep heartbeat at 4.7, just because it's powerfull enough for my current needs, and also I don't want to put more stress on the infrastructure than needed.
Well, if not disclosed you could assume that somebody did due diligence for you, and could include sources. I don't even trust LLM even if all the information is included in the context window if I need reliable information. Trying to make money on slop is really bad manners. It's a scam, you can't call it otherwise.
Btw, I like AI, it did a ton of value for me. We just need to find a way to live with it, without getting doomed in misinformation.
I feel for you. I was looking for some wildlife events on Youtube, only to find that all of them were AI generated, trying to get views. I can only find content somehow reliable if I put filter for content before of AI era.
Yeah, its heartbeat. If you keep HEARTBEAT.md empty, it won't be consuming tokens every half an hour. Use cron for periodic tasks if possible, but if you need LLM to process something periodically, that's what heartbeat is for. Z burned 166mio tokens in last 30 days, majority was the heartbeat. Definitely wouldn't pay for that if it wouldn't be the subscription. It uses 1-4% of every 5h quotas. Definitely not sustainable, but I take the opportunity to learn things while it lasts.