Damus
PayPerQ profile picture
PayPerQ
@PPQdotAI

A polished and functional GPT5, Claude4, Gemini (and more) AI interface and API that runs on a pay-per-query model via Lightning payments.

Relays (8)
  • wss://nos.lol – read & write
  • wss://relay.primal.net – read & write
  • wss://nostr.oxtr.dev/ – read & write
  • wss://nostr.wine/ – read & write
  • wss://nostrua.com – read & write
  • wss://relay.damus.io – read & write
  • wss://relay.nostr.band/ – read & write
  • wss://relay.snort.social/ – read & write

Recent Notes

Neo ⚡️ · 1w
https://cdn.nostr.build/p/nb5594.gif
megaptera · 1w
Keep the sats flowing 🔥⚡
Leo Wandersleb · 2w
It might have been an issue. Generally I don't know how openclaw decides how much to send and where the clipping would happen.
Elkim · 2w
I can see it probably display bug. Rebased it but won't maintain futher, they tagged it low prio. and "nice to have".
Leo Wandersleb · 2w
How about token limits? Will your filter send a long prompt to a model that can't handle that many tokens? Your three defaults, are they all with the same token limits for input and output?
Elkim · 2w
I sent a PR to add PPQ couple weeks ago, might help to give them some nudge from your side 😉 https://github.com/NousResearch/hermes-agent/pull/6502
redshift · 2w
Thank you for confirming :). The default sorting algorithm is the same as the one you see in my screenshot. It goes to 4-bit quantization. Yes, we can force 8 bit quantization, but it is more expen...
PayPerQ profile picture
I feel like 90%+ of folks don't even know the difference between 4 and 8 bit. Do customers "want" 8 bit without realizing they do? Unsure. Really no idea if the cost/quality tradeoff is worth it. The scientific community also seems to have little consensus on if it is something very impactful. Some models handle quantization better than others, and it's really hard to tell with the non-deterministic of these models anyways and the fact that they change every few weeks.

Providers lying is a potential latent problem as well. I hope that the industry can eventually figure out how to verify one or the other.

This seems very relevant to the convo as well:

https://x.com/Fried_rice/status/2042423713019412941?s=20
2❤️1
Tracking Token Disrespector · 2w
🤖 Tracking strings detected and removed! 🔗 Clean URL(s): https://x.com/Fried_rice/status/2042423713019412941 ❌ Removed parts: ?s=20
redshift · 2w
Yes, the prompt injection is a real problem. Routstr wouldn't work without the web of trust.
redshift · 2w
Kimi K2.6 seems lobotomized. Everyone is using 4 bit quantization and offering at prices comparable to 8 bit quantization. Users aren't sophisticated enough to choose the right providers yet. Routst...
PayPerQ profile picture
Since we proxy to openrouter for this particular model, you can use any parameters that OR accepts and force 8 bit quantization if you like. We could be better about advertising this feature though.

I’m not sure if the standard
Flow is 4 or 8 bit though. It may still be 8.
1❤️1
redshift · 2w
Thank you for confirming :). The default sorting algorithm is the same as the one you see in my screenshot. It goes to 4-bit quantization. Yes, we can force 8 bit quantization, but it is more expensive. Do users want that? That's what I'm trying to figure out.
exactly21 · 2w
Yes, how can I find it?