@
Building iroh with the amazing folks at number 0 (n0.computer). Generally striving to increase user agency and excited about commons networks. Only works for Canadian CEOs, apparently. Rust, cryptography, CRDTs & more on my feed
@yep!
@W/o chunk groups, you're filling the buffer up only up to 1024 plus a bunch of layers from the Merkle proof relative to the root. W/ chunk groups you fill it up to 16KiB plus the proof. But on the proof level you could again transmit a set of say 64 chunk group hashes at once and derive parents
@There's another thing I've always thought about but never implemented: in theory we could leave out layers in the Merkle Tree again above the chunk groups: what you actually care about is having an "unverified integrity byte buffer size".
@😂 I wonder now, is the NL one wrong though?
@me: [definitely not a yes/no prompt, likely not even a question at all] GPT6: Yes. [...] Every. time.
@annleckie.comReposted by @01a08e37-2b26-717e-8fc3-9b51938089d8Not gonna RS the thread that made me think about this (you've likely already seen it, or a similar one) BUT Every person I know who has cut off or gone low contact with a parent took YEARS of agonizing and trying to make things work because damn that's their PARENTS.
@rfksbrainworm.bsky.socialReposted by @01a08e37-2b26-717e-8fc3-9b51938089d8If you ever wondered why visible light is those particular wavelengths, it's because life evolved to see in the narrow range where water is transparent
@Err wtf this is big?
@It's always about "what can you train the models at?". It's hard to RL this question. I also think it's hard to RL maintainability, although perhaps a little easier.
@Gufo is pretty impressive. Running Qwen 3.8 Flash Next on my strix halo machine locally is now feels entirely differently:
llama.cpp: ~400 t/s prefill, ~20-24 t/s tg (no mtp, it's consistently slower)
gufo: ~1300 t/s prefill, ~30-40 t/s tg (with mtp-7)
@Ludwigshafen/Mannheim
@There are some interesting ideas for future jobs by modern monetary theory proponents: The Job Guarantee. The idea is to have the government fund community or care work at livable wages.
@rklaehn.bsky.socialReposted by @01a08e37-2b26-717e-8fc3-9b51938089d8We have released a new version of
@iroh.computer.
Nothing major in terms of features, but we did manage to significantly reduce the binary size.
So now my entire embedded device zoo is running on iroh from crates.io. No more patch needed.
@Relatable. For real though, do you think it might get activated because of the word "execute"?
@evadeva.bsky.socialReposted by @01a08e37-2b26-717e-8fc3-9b51938089d8@iroh.computer
Was "Dial Keys, not IPs" inspired by reticulum.network? And what would be the differences between them?
@We just ended up realizing that dialing by key alone is already a super useful abstraction. Maybe more things should be built like this. I'm glad to see that's how reticulum is designed as well. Not sure about exact differences. Someone would have to educate me on reticulum.
@It wasn't! We aren't actually super familiar with reticulum and we weren't when we named it. The concept mostly came out of a need for stable addressing in a world where neither IP nor DNS was possible to address a device adequately and securely.
I swear @iroh.computer is erasing basically every network annoyance I could have possibly encountered when ideating about decentralized systems
@No. But you only need to give it one more piece of information: the relay URL (or the DNS pkarr server to resolve it, but at that point the relay URL is likely easier)
@Relays don't have "a table" and don't use any pkarr mechanisms which is why I can't parse your last question.
