Posts by lynne@mk.pars.ee
(DIR) Post #B5ABBhxsvZsw1OpgXI by lynne@mk.pars.ee
0 likes, 0 repeats
Huge respect. Distilling so much metal, spending days in the lab.https://www.youtube.com/watch?v=qT5XLkATzSI
(DIR) Post #B5DYBFiepCuaq6LFEe by lynne@mk.pars.ee
0 likes, 0 repeats
@whitequark@social.treehouse.systems I looked into it when I got sick of SSHFS disconnecting, but went with NFS over WireGuard since it's faster. I even wanted to use NFSv3 over UDP to keep everything UDP but they've pretty much removed support for it, plus NFSv4 had more features.
(DIR) Post #B5FnGCJKoPHrLREVXc by lynne@mk.pars.ee
1 likes, 0 repeats
RCR finally reviewed a cybertruck!"Fisher Price brutalism". Beautiful.https://www.youtube.com/watch?v=L3oO510dyVI
(DIR) Post #B5OPjnwjJYAPzPsrqK by lynne@mk.pars.ee
0 likes, 0 repeats
An NVIDIA engineer shared he spends ~5000$ a month on Opus tokens (but really, they get free unlimited tokens). He mainly works on test programs, so it's non-critical, and if it's vibed properly, no one would complain.His division can't hire anyone, he couldn't before AI anyway. But they've pretty much frozen hiring new employees now. Same goes for AMD and Intel.So yeah, I'm pretty certain layoffs are coming in a few months, first the low-level folks. Those in better positions would just direct agents and goof around. And it's definitely not a bubble, they've grown so accustomed to it, even if the AI companies crash, they'll buy their models and self-host them for their own use. Though they all mainly use Anthropic, who are pretty well-off.
(DIR) Post #B5OPjoAYU9E4gILuOO by lynne@mk.pars.ee
1 likes, 0 repeats
Freelancers can't compete either. Not when companies have so much power to just do everything themselves. Doubt they'll upstream changes either, and I really doubt they'll help out maintainers.In short, we're done for.
(DIR) Post #B5lC7yBVyB4jyGVSNs by lynne@mk.pars.ee
0 likes, 0 repeats
Oh my, Claude can now interface with Blender directly.Wow.
(DIR) Post #B5lC7yNDGgQuYXyncO by lynne@mk.pars.ee
1 likes, 0 repeats
I've hardly used Claude and I can't really live without it. I have so many almost-complete abandoned projects since I get stuck on menial tasks needed to finish them.It was able to write very complex algorithms too if you give it enough contexts.It's nuts.
(DIR) Post #B5n9D5f5gcdB3v0hOK by lynne@mk.pars.ee
0 likes, 0 repeats
Honestly, I'd rather have well-integrated AI agents handling all customer queries than humans. Agents could tell you exactly what's happening instantly. Humans are idiots and would rather waste some other human's time.Also agents can be hacked and manipulated more reliably.
(DIR) Post #B5vJAEjK4Gpz7lwmTw by lynne@mk.pars.ee
0 likes, 1 repeats
What would be scary is if you ordered something from amazon in the night, and it arrives the same night after barely 2 hours.
(DIR) Post #B5w7SD77oJfn25qQIC by lynne@mk.pars.ee
0 likes, 0 repeats
It took them 3.5 hours to package and send it off to a delivery agency. At 23:00. At night. On Sunday. On a holiday.Yes. This is how things should run.
(DIR) Post #B6JhVDupLsozNCuH9k by lynne@mk.pars.ee
1 likes, 1 repeats
I finally reverse engineered the entire ProRes RAW decoder and display pipeline. You don't need to use Windows or OSX plus Resolve just to view, edit or process your camera raws!
(DIR) Post #B6SVISE7Nk9WxziK4e by lynne@mk.pars.ee
0 likes, 1 repeats
I wrote an APV decoder in Vulkan compute, and it got merged into FFmpeg!The VFX industry is moving towards APV, and several Android smartphones record in APV for very high quality recordings.
(DIR) Post #B6WqpuQAS6DpdTz1yi by lynne@mk.pars.ee
0 likes, 0 repeats
I never in my wildest dreams thought code would be the first thing to get automated. I thought copilot was a glorified autocomplete for years. I thought vibe coding last year was a more general but still a yassified autocomplete that required users to do a lot of manual work, doomed to build upon broken incorrect code.But, here we are. I guess the reason why none of us recognize this as cyberpunk or futuristic is that it's just not something most thought possible, until last year, really. And because the densest introspective fog isn't over the past, or the future, but the present.We'll likely end up with a superhuman coder by the end of this year, if it isn't already here, and I think it is, just lacking a harness. At that point, writing code manually will be harmful to a project. Human feedback and taste would stop helping if you can just burn more tokens and let the AI arbitrarily rewrite code if it gets stuck in a local minimum. At that point, maintainers will likely be more of a harm, and reviewing will become impossible.Sure, we can write code for fun. But any public code ends up being more training data. And agents ran by those with nothing better to do will find a way to contact you to nag you about your low-quality code.Human training data will become irrelevant in no more than a year, when synthetic data from majority AI-led projects becomes used in production, at which point anything proven to last a month or two with no bugs is good enough material to use in training.
(DIR) Post #B6amqSFZ3R6JN0rtKK by lynne@mk.pars.ee
1 likes, 0 repeats
I'm pretty sure Chapter 1 Sentence 1 of all HDR specs say "Note: This spec cannot be implemented, if you believe you have implemented it, you're wrong".Are you sure it's actually working? If so that's clearly a spec issue that must be fixed.
(DIR) Post #B6b5mpIFKvGgoB6dhw by lynne@mk.pars.ee
0 likes, 1 repeats
Honestly its nuts that I can ask an agent to build a mesh-shader based, raytraced minecraft clone (avoiding vertex/fragment shaders altogether), and it runs.Code is free, experimentation is free, I wonder what could be achieved like this?
(DIR) Post #B6b5mpSAk1CxIxkZBA by lynne@mk.pars.ee
0 likes, 0 repeats
Honestly, the OpenWorm codebase is so bad that vibecoding a new codebase from it, using modern SNN emulation tools like EDEN and a modern physics simulation code with vibe coding would be a considerable improvement.@niconiconi@mk.absturztau.be
(DIR) Post #B6kTAo9bkVcaxLrMMC by lynne@mk.pars.ee
0 likes, 0 repeats
@niconiconi@mk.absturztau.be Now I know why OpenWorm got stuck - they never implemented proprioception, the feedback loop between muscles bending and neurons firing. All their demos just use forced sinusoidal motion.In fact, there never was any feedback given to the nervous system. Even worse, it was just calling python on every single step!The vibe coding experiment is going pretty great. I rely entirely on EDEN (state of the art spiking neural network simulator) to generate the SNN network that the c302 model outputs. The EDEN output is an XML of nerves, receptors, types, raw equations and algebra (with ), which get converted into Vulkan compute shaders with multiple passes (integration, spiking, gating, etc.).Unit tests (hundreds of very small model networks from the NeuroML database) get compared to experimental results to make sure we remain grounded.The idea is to scale this to the full Drosophila model later on, so rather than doing expensive fluid physics sims, simple friction is just used to approximate a fluid/sand-like environment. So rather than waiting 10 seconds per step, we do 12k steps per second, since everything is kept on the GPU.Opus 4.7 knows a lot about biology so its very neat seeing how it guides itself to a solution while debugging by just looking at how the membrane voltage evolves directly.Right now I'm implementing proprioception, hopefully I'll be the first ever to see naturally generated locomotion.
(DIR) Post #B6kTAoT6a1DXvoyvkO by lynne@mk.pars.ee
0 likes, 0 repeats
@niconiconi@mk.absturztau.be I'm really pleasantly surprised that c302 is still maintained, and that the XML it outputs is able to be processed by EDEN.I mean they both expect and operate on NeuroML 2, but I expected the classic "we don't talk to each other at all ever" problem that science has when implementing a spec.
(DIR) Post #B6kiYpVNsLLDPRxRjs by lynne@mk.pars.ee
0 likes, 1 repeats
@niconiconi@mk.absturztau.be Full c302 D-model (Multi-compartmental cells) simulation. Had to patch c302 to support B-class motor-neurons from Boyle's 2012 paper. After that, I had to add support for modelling all cell types and cell groups, as well as handle post-spike Ca-channel depletion. And fine-tuning to make sure all the membrane voltages stay and spike within realistic ranges.This is the actual full C. Elegans nervous system driving the locomotion. Sibernetic/OpenWorm just generated sine waves to drive the muscles and pretended it was a simulation. I can't believe no one else has done this before.
(DIR) Post #B6yUrdYHn5ZcI6QScC by lynne@mk.pars.ee
0 likes, 1 repeats
@niconiconi@mk.absturztau.be don't blame OpenWorm developers for lack of progress, C. Elegans is a very shitty organism to simulate. Evolution hired the DOGE team feat. Elon Musk to cut everything out that didn't need to be there, including a CPG, which means the nervous system isn't "alive", ever, really, since there's no "clock" signal to drive anything.All it does is spike on inputs, do some processing, and output. It doesn't/can't do anything on its own. In my earlier post, I had tonic stimulation leak into the nervous system via the noise injection stage of the SNN sim, which got filtered through the SNN to become a sine wave locomotive drive. In a way, that was the real SNN doing the driving, but it was stimulated, rather than emergent.The reason the whole field is so dead is because the only connectome we've had has been of the world's most useless/frustrating creature to simulate. We got the fruit fly connectome at least, which has a CPG and can do something useful on its own, but it took decades. If only there had been a better creature with an actual consolidated nervous system.