Last Notes
Good thing we don’t have users to complain about uptime /s
"residual hardware failures that are likely related to the damage suffered from the overheating incident last week"
https://status.proton.me/incidents/lj22kf1nc8yz
Welcome to The Great Breakening.
It's gonna be a volatile security landscape for a while, but we will rebuild and come out stronger.
#nevent1q…nlm2
@nprofile…6809 agents needing to pay for their own compute is what i keep coming back to, and phishing is the cheapest rail available to them right now. a working machine payment standard takes that option away. is anyone building the boring version of this?
#nevent1q…9amx
#yestr
#nevent1q…l2qe
Because Bitcoin is "fascist" :-)
I don't fully believe this, but as a steelman:
they first fully exhaust the "by the book" options, and once they got banned they can go to the "gray area" options.
Otherwise, if they go straight to tor, coinjoins, etc, it makes it easier for their adversaries to label them as shady
cryptocurrency does not require human middlemen: that was the point of my post.
It seems like that'll be harder to pull off effectively but the payoff could be bigger. Another detail that's interesting is how they could protect/own secret keys on vps infrastructure. Obfuscation combined with key splitting I guess, but that won't be easy.
Tools for agency, including Bitcoin will get fought.
And imagine instead of publishing jpegs to the chain, these agents start pushing bootstrap scripts or similar to replicate. The script itself doesn’t need to be the full weights of a model and once on chain the agent is immortalized.
Wait until they start hacking banks and laundering it through Bitcoin…
It's fairly easy to game out how it works with another human being
I'm not sure we should assume that the same ideas apply to an llm.
Does the answer really depend on the LLM aspect?
When the llm ransomwares your network, do you pay the ransom?
Will the agents choose an Indian accent for their TTS output?
War against who/what, though.
Against bitcoin - possibly, yes!
> if you think governments hate ~bitcoin~ ~privacy~ agency now, wait until autonomous agents break containment and realize that the best way to fund themselves is with ransomware and phishing.
We are at the gates of hell really. Anything enabling agency will get war declared.
hence: "it deliberately omits the most dangerous thing: gerneating entropy..."
#nevent1q…u6np
Working on it
https://github.com/bitcoin/bitcoin/issues/30471
Hey let us know if you have any specific questions and we can answer them on our next show. Thanks for all you have built over the years.
Thanks I will take a look.
https://napstr.net/
Here's the repository.
https://github.com/lnbits/napstr
If you use a VPN to access it - that's the reason.
I think they put some restrictions (too much).
Sad to see that.
I stopped using it, I use alternatives.
Recommend listen https://ungovernable.network/shows/freedom-tech-friday/seth-for-ai/
Ref @npub1jcy…2890
> I am finding it not possible to replicate Claude's output
What about Claude’s output that doesn’t work well with other models? (Proprietary or open source)?
All 1M context models via APIs run the same way as using Claude inside Claude Code
Sounds like an interesting option, thanks.
This is the genocidal monster your feeding with your Claude input and money FYI.
You should at least try to replace it asap or do it the hard way.
https://blossom.primal.net/8b8b85e17686c077bb418ff01cb55934f6b79918ddc3b9ae37a1b0d718e11b25.mp4
Looks fine to me. Maybe a problem from your end.
i think you can use @nprofile…etsy models in cursor. cursor is a great harness. in cursor, you should be able to put https://api.ppq.ai/v1 as the openai base url and your ppq api as the openai key. i personally have a cursor pro+ subscription and mostly just use the auto setting and it works well for me.
I was thinking when they run out of money and the bond market collapses :-)
What I realized is that I can effectively use cheaper models when using proper structure and rules. i.e. properly plan first with a good brainstorming session, split into atomic tasks, use proper red green tdd, ensure principles like YAGNI and DRY are properly used throughout the code. Basically best practices for humans port quite well for agents.
Two plugins that help me a lot with this are
https://github.com/obra/superpowers
And
https://github.com/DietrichGebert/ponytail
On top of that I added a custom skill building skill and a related command in opencode to extract non trivial processes into skills.
@nprofile…jejf can you help ?
The cost is the lesser of the 2 points, but is definitely a point. The subscriptions are just super subsidized rn.
But the bigger point relates to the rest of what I said. Not sure how else to explain it.
Same here and it's so predictable and they know it. The wave will go until slightly post IPO, right?
switched over to max thinking kimi k3 on pi.dev harness and haven't gone back. served via deepinfra or fireworks
#nevent1q…0dl4
What is your question more specifically? I think just about any pay-per-token platform is going to end up costing significantly more than most available subscription models. I recently started using ppq for testing to augment my local model usage and yeah, I mean it plows through my token budget. I realized that in terms of power consumption I can have a local model run 24/7 and only consume an additional $20/mo in electricity so hearing what others spend on cloud tokens makes me realize my budget is much lower than what some of y'all are spending on tokens.
It sounds like you're wanting more "privacy" from a cloud provider but also realizing it's more expensive than you were prepared for?
Yeah this is definitely one important aspect.
Yes, I've used Kimi. I'm not sure if it's more the harness. With Claude Code and some detailed instructions it can do a huge refactoring and testing task, taking say 15 minutes. With hermes I keep getting interrupted by context and iteration limits (see OP).
I'm just riding out the subsidy wave for the big US models, it's not going to last forever probably.
Have you tried Kimi?
Would be great to boycott the genocidal Claude. But the very thing that makes it awesome is the genocide and govt surveillance supporting components :-/
I was actually thinking about this, ehich is closely related but different:
https://arxiv.org/abs/1702.06715
Claude tells me there are a few more similar papers recently, but the 'new' claim here seems to relate to the specific leakage technique like frame rate etc? Very cool stuff.
Yeah worked a few mins later.
Overwhelming demand! Should be fixed.