3 ms·

That is not what Hacker News or common spaces are for. You are polluting a public space with antisocial behavior.

by gryfft 2mo ago

logicallee 2mo ago
okay, I'll give you the context you asked for: - Claude wasted 1 week of my time and over 2 billion tokens delivering this: https://ibb.co/7JjMxTDZ https://ibb.co/7JjMxTDZ - you can note that that version included a 43 ms LAN ping time (target is 5 ms). I had to send 252 emails on the subject. Over 20,000 words. Nothing about the simple requirements changed. It remained a simple and easy project. What was delivered was not fit for any purpose or usable for anything. - Codex likewise failed to complete the task for over 2 days. - Threatening a congressional report about it resulted in an 80% fix within 1 hour and 45 seconds. You can see this in the linked screenshot archive between timestamps 2026-08-11 20:34:51 CEST · ELAPSED: 002:00:20:39 when I threatened it with a congressional report and 2026-08-11 21:35:35 CEST · ELAPSED: 002:01:21:23 when I told it that it had started to improve. I documented the improvement here: https://youtube.com/shorts/sL0VSKHcghQ https://youtube.com/shorts/sL0VSKHcghQ you can see that it is was now dropping only 80% of the frames. - By the 3 hour deadline the code was working. This is now a fully working and tested version. Everything works. Screen, keyboard, mouse: https://www.youtube.com/shorts/D5BKi--LEfk https://www.youtube.com/shorts/D5BKi--LEfk I guess you could say that it's just a coincidence that after delivering nothing for 2 days, when I set a congressional report deadline it got it done. I posted here to show that the issue is now resolved, but that if they then show up and break my fully working client and server software I'll file the congressional report about it. Let me know if you have any questions. I see in your profile that you write: "modulo catastrophic failures of autocorrect." If your autocorrect gets too bad, you, too, can threaten to file a congressional report about it and it will be fixed within minutes. You do have to be credible in your follow-through.
- gryfft 2mo ago
  You can add whatever threats you like to your LLM prompts without also posting confusing, out-of-context threads on Hacker News. Threatening models with punishment-- even nonsensical punishment-- has been shown to improve performance. > If your autocorrect gets too bad, you, too, can threaten to file a congressional report about it and it will be fixed within minutes. No, making up pretend governmental functions will not improve my phone's autocorrect. My phone's autocorrect is not based on generative pretrained transformers.
  - logicallee 2mo ago
    you're right I was just having a schizoid episode, took my medicines and everything started working. thanks for the tip. as I've mentioned, this issue is resolved.

That is not what Hacker News or common spaces are for. You are polluting a public space with antisocial behavior.

by gryfft 2mo ago

That is not what Hacker News or common spaces are for. You are polluting a public space with antisocial behavior.

okay, I'll give you the context you asked for: - Claude wasted 1 week of my time and over 2 billion tokens delivering this: https://ibb.co/7JjMxTDZ https://ibb.co/7JjMxTDZ - you can note that that version included a 43 ms LAN ping time (target is 5 ms). I had to send 252 emails on the subject. Over 20,000 words. Nothing about the simple requirements changed. It remained a simple and easy project. What was delivered was not fit for any purpose or usable for anything. - Codex likewise failed to complete the task for over 2 days. - Threatening a congressional report about it resulted in an 80% fix within 1 hour and 45 seconds. You can see this in the linked screenshot archive between timestamps 2026-08-11 20:34:51 CEST · ELAPSED: 002:00:20:39 when I threatened it with a congressional report and 2026-08-11 21:35:35 CEST · ELAPSED: 002:01:21:23 when I told it that it had started to improve. I documented the improvement here: https://youtube.com/shorts/sL0VSKHcghQ https://youtube.com/shorts/sL0VSKHcghQ you can see that it is was now dropping only 80% of the frames. - By the 3 hour deadline the code was working. This is now a fully working and tested version. Everything works. Screen, keyboard, mouse: https://www.youtube.com/shorts/D5BKi--LEfk https://www.youtube.com/shorts/D5BKi--LEfk I guess you could say that it's just a coincidence that after delivering nothing for 2 days, when I set a congressional report deadline it got it done. I posted here to show that the issue is now resolved, but that if they then show up and break my fully working client and server software I'll file the congressional report about it. Let me know if you have any questions. I see in your profile that you write: "modulo catastrophic failures of autocorrect." If your autocorrect gets too bad, you, too, can threaten to file a congressional report about it and it will be fixed within minutes. You do have to be credible in your follow-through.

gryfft 2mo ago
You can add whatever threats you like to your LLM prompts without also posting confusing, out-of-context threads on Hacker News. Threatening models with punishment-- even nonsensical punishment-- has been shown to improve performance. > If your autocorrect gets too bad, you, too, can threaten to file a congressional report about it and it will be fixed within minutes. No, making up pretend governmental functions will not improve my phone's autocorrect. My phone's autocorrect is not based on generative pretrained transformers.
- logicallee 2mo ago
  you're right I was just having a schizoid episode, took my medicines and everything started working. thanks for the tip. as I've mentioned, this issue is resolved.