@yhancik yeah, I get that. It's complicated. IMO this is one of the issues with centralized spaces. IMO it'd be nice if the onboarding was more organic based on attaching onto social connections that are established off platform rather than getting exposed to large volumes of low context info.
@yhancik Oh? tell me more.
@roknrol that's a fun idea. Social extortion. I think "be blocked by B until you block A" would be less friction. I'll make sure to add it the next time I make a new social media platform. :P Maybe only notify users following A to keep the spam down.
This is kind of how fedi blocking works already (which I like), but that breaks down more with large instances that lose social context between users. IMO >1000 active users is when the context collapse muddies enough that individuals can't expect to trust everyone else.
@vbatts No, the WIN 4. It's more like a PSP. it's honestly perfect as a self contained device compared to everything else. Pretty easy to take out and use and you can get them to be pretty powerful. Driver support is decent. AMD ROCm GPU accelerated stuff doesn't work with integrated graqhics though which is inconvenient. Having a physical keyboard is what really makes it. I'm not a fan of how the WIN 5 removed it. :/ It's less portable than my current setup tho!
local llm harness tinkering
I added "summary" fields to my edit/writ/shell tools in my harness and now I have short human readable messages as it's doing stuff which has been a nice middle ground between full auto and manual review. If the command is within the strict allow list, I just see the summary since I can trust that it's just find or sed or something.
@neoluddite I'm absolutely Golgolas
hopefully there's noöne amongst my followers who needs to be told this (for reasons, ofc, of all of you being honest, law-abiding people who understand and respect the norms of their respective societies), but people who brag about how much they love breaking the law will rat you out the first chance they get, and people who try to incite you to leftistly commit crimes work for your country's domestic intelligence service
me to my wayfarers: "hey cyber shades, what's the top news on Facebook?"
my sunglasses talking like Knight Rider: "a child has built the Statue of Liberty out of trash and he requires 1,000 faves on his post. also he is a war veteran and has no legs. also it is his birthday today."
me: "this child, is he also Jesus for some reason?"
sunglasses: "yes jackdaw, the child is also Jesus."
me: "cyber shades please send a sad react and reply that this is so true and inspirational and important."
gripes about matrix protocol
@cmdrmoto It sucks because I don't have a functional alternative client ecosystem to go to since libpurple isn't very usable anymore. I can't escape the trap. Either I use shitty matrix clients, go insane making my own, or make a dozen clients for all the puppet bridges from scratch. 💀
@vbatts I use waybar for the systray with some minimal configs. https://github.com/RangerMauve/my-gpd-setup/tree/default/.config/waybar I think the generic GNOME systray utils for wifi and bt? Gestures on niri are fun. Used it with my magic trackpad before I ditched it for the twiddler. It's not perfect but it's the best in it's form factor and the easiest to learn. took a few months of a few mins a day going through their trainer web app.
@vbatts Yeah. I found the battery situation to be the most annoying part. followed by a lack of usb-c displayport support. the Khadas Edge 2 is the least annoying out of the box one I found, but the lack of storage expansion has been a PITA. https://www.khadas.com/edge2 It handles usb-c docks well though. MNT is neat but still a bit more DIY than I have time for ATM :P
Was re-reading Melanie Mitchell's 2019 book, "Artificial Intelligence: A Guide for Thinking Humans", https://melaniemitchell.me/aibook/ and was relieved to (re)read how she refers to "units" instead of "neurons" in artificial neural networks. Even adds a note about doing so to avoid the dismay of neuroscientists. Indeed.
Got reminded again today of how powerful biological neurons are by Mickey London's lab latest preprint:
"What can a neuron compute", Aizenbud et al. 2026 (Segev, London)
https://www.biorxiv.org/content/10.64898/2026.06.08.730984v1.full
The secret sauce of biological neurons? Mostly dendritic morphology:
"Bounds on the computational complexity of neurons due to dendritic morphology"
Anamika Agrawal, Michael A. Buice 2025
https://www.biorxiv.org/content/10.1101/2025.07.11.664448v1
A point spectacularly well made by Jones and Kording a while back:
"Can single neurons solve MNIST? the computational power of biological dendritic trees", Jones and Kording 2020
https://arxiv.org/abs/2009.01269
... in line with plenty of prior work, even by the same lab, like:
"On Functions Computed on Trees", Farhoodi et al. 2019
https://direct.mit.edu/neco/article/31/11/2075/95669/On-Functions-Computed-on-Trees
The key ingredient? That a presynaptic neuron delivers multiple inputs across the postsynaptic neuron's dendritic tree. That enables a lot of the above's computational properties. To reproduce that, a whole multi-layer ANN is needed to match the power of a single cortical pyramidal neuron:
"Single cortical neurons as deep artificial neural networks", Beniaguev, Segev, London 2021.
https://www.cell.com/neuron/fulltext/S0896-6273(21)00501-8
What's more, adding biological constraints to ANNs improve the latter:
"Do Biological Constraints Impair Dendritic Computation?", Jones and Kording 2021
https://www.sciencedirect.com/science/article/pii/S0306452221003900
Occult cyberpunk. Yap with me about decentralized systems, wearable computing, and biohacking.