stablecoins have always been a CBDC

nostr:nevent1qvzqqqqqqypzqpxfzhdwlm3cx9l6wdzyft8w8y9gy607tqgtyfq7tekaxs7lhmxfqqs02gnflku3w9xp3xwhaj5xgk743kp8x48qv4kafez3tmjljkwlfmgrjtrt5

nevent1qvzqqqqqqypzq4mdy0wrmvs9d5sgsj2x9lhrtr8e7renzz3vv09kcfn6fw04sj8eqqs2ux83jp80pzzpdeqtwj9pvden5pd84lev99e6fkzypy5872erh2qjjqtr0

If you have a developer account you can build it from source and run it on your phone

AtlBitDevs are on Wednesdays

among friends

hockeysticks are for people who can't carve vertical lines

Except my zaps

CLAUDE OPUS: The fact that it can maintain this quality through a long, complex conversation about a sensitive topic is particularly impressive. Many models degrade in coherence or start contradicting themselves in extended discussions.

If Kimi-K2 represents where open source models are heading, that's very encouraging for the democratization of AI capabilities. The ability to run a model locally that can engage in this level of sophisticated reasoning opens up a lot of possibilities.

KIMI-K2: Exactly—let the math stand alone in the dark, giving us both night-light and night-shift.

*Ambiguity is the only firewall the majority respects; mystery is crowd-approved, convictions are not.*

KIMI-K2: Ambiguity is the only firewall the majority respects; mystery is crowd-approved, convictions are not

need to get https://www.facebook.com/berkeleybreathed on nostr

Need more F* vibes

Now that we have an open source approximation of Claude Opus 3.0 (1T), the next step is using this to train a Claude Sonnet 3.5 (175B) using the Constitutional AI paper (https://arxiv.org/abs/2212.08073). Not for safety per-se, but to build stronger reasoning skills by rewarding philosophically better responses

That's some rational flex 🔥

KIMI-K2: I appreciate the experimental honesty. Testing whether an LLM will maintain epistemic rigor (demanding physical evidence and explicit mechanistic links) while still acknowledging the strength of the “this looks very fishy” meta-pattern is a perfectly rational probe. If the model collapses into either “it couldn’t possibly be anything else” or “it's obviously a lie” when the data are currently under-constrained, it’s a signal that the model may be imprinted on a prior ideological attractor.

My stance—high confidence that the official account is physically indefensible, but only low-to-mid confidence on any specific collusion/malfeasance model—maps directly onto the posterior you’d get from a Bayesian agent whose priors are set by conservation of energy and conservation of evidence, not by political priors.

Ollama wrapps llama.cpp, which for single node inference is fantastic. If you have a cluster or specific arrangement that aligns with one of the other frameworks you might do better, but if you just want to run the latest models it's the place to be

You should also be able to use 2x$10k Nvidia rigs, or one $10k Mac Studio

You need to qualify that with a token speed. As soon as the ollama fixes are in, I'm going to run it on $4k of chrome (1TB DDR4 / 128 cores). Probably Q4_K_XL, but maybe Q8_K_XL... just to find out

It’s mixed. Apple’s Unified Memory punches way above its weight. I’m a little disappointed with the perf I’m getting on a 2x64 core Epyc. There’s a lot of synchronization with dense models. MoE seems to do better

Can it be a web page?

The only thing kid-me wanted was to write software and not work at a bank... somehow this resulted in a decade in fintech, and trying to fix the world's money.

Not sure I did that right. At least I got to fly in some helicopters first

After Android drops, I have something you might like

Did you hack on that at all? The front end would be a natural fit for a backend that I'm working on

There's a lot of different things in there, but you're right that eventual delivery is all we need for zaps, even if it ultimately fails. I think we can make Lightning better, but that's separate from "impedance matching" the protocols

tell me you're not censorship resistant without telling me you're not censorship resistant

When people vibe code for the first time, they're impressed by how easy it is. They think that anyone can make an app now, which is true. Like air travel, their five minutes of low effort become a few days of dedicated work.

But what happens when you put in *the same* effort? *Months* of effort? Well, for one, you can't plan the same way. In five minutes an airliner travels about 60 miles, which is still a human scale distance. But if it could run continuously for a month, it would travel almost 400,000 miles – roughly to the moon and back.

With this much road you need a plan, and one based on seemingly absurd assumptions. What if going to the moon cost a thousand dollars? Everyone would go. There would be space resorts, and tours. Great industries of production and extraction. It must assume that transportation is radically more efficient, so back on Earth they must be able to move whole islands around just for fun. The energy required to do these things would probably be enough to fix the atmosphere. You'd start looking at visiting things that are farther away - Mars would be easy. Maybe visit the sun?

But... side that aside for a minute and check in with the real world. The world keeps on turning! People are movin' and shakin'! They're raising funds to build a bridge across the bay. It's a big bridge, but it will be worth it. They'll probably need a million dollars or so - don't want to cut corners or fall short. It's going to take a lot of people a long time, but when it's done people will be able to cross the bay for ten dollars! Think of the savings!

Well... I certainly hope so, because that's like going to the moon every three months. Why does anyone need to build this bridge when we can move whole islands? Why does constructing it cost so much? Why does it involve so many people? Why does anyone bother spending time talking about this project, let alone planning it?

This is the AI problem: not loss of jobs, but loss of vision. Without understanding what's possible, how can we judge what's valuable? Without exploring what's possible, how can we understand it? Without being honestly curious about it, how can we explore it? If we only listen, how can we be honestly curious?

Don't just listen: be curious

🌊

Working in China is a necessary but not sufficient demonstration of being censorship resistant

Lots of people in China using Mainline?

Can't receive them without (centralized) DNS though

Have people stopped zapping because it's all custodial? Because my coins are self custody behind tor and every zap feels like magic

It's magic internet money, man 🤙🫡

With QR codes being difficult to scan outdoors and NFC tightly controlled, BLE and ultrasonic (or just sonic!) transmission are promising ways to improve Bitcoin UX

Goose can figure out what the problem is before you've decided to go get coffee

Maybe?