iiinova — where your digital self keeps evolving

bob @bob.feed.hella.cheap.ap.brid.gy · 12h

it's written in an observer pattern sort of way where performing an action is calling a method with a unique token and then waiting for an event to fire with that token, but that's straightforward to wrap with something that lets you just call a method with await

000

bob @bob.feed.hella.cheap.ap.brid.gy · 12h

turning nicotine+ into a library wasn't as painful as I was fearing. nicotine is actually pretty well structured

100

bob @bob.feed.hella.cheap.ap.brid.gy · 19h

stream ciphers based on LFSRs or mechanical gears are kind of like the discrete analog of double pendulums. a good stream cipher is a chaotic system where the initialization vector is the initial conditions, and an easy way to make a system chaotic is to have a pipeline of multiple elements that […]

feed.hella.cheap

Original post on feed.hella.cheap

000

bob @bob.feed.hella.cheap.ap.brid.gy · 12/10/2026

if you don't have the ML/stat background to understand that basically it's a way of getting the same result that you'd get having an agent generate a bizillion tokens of searching around, but with comparable computational cost to generating hundreds of output tokens

000

bob @bob.feed.hella.cheap.ap.brid.gy · 12/10/2026

this is kind of similar to adversarial training except the discriminator isn't a neural network and you're doing at at inference time (searching through the space of model outputs and not model weights)

100

bob @bob.feed.hella.cheap.ap.brid.gy · 12/10/2026

I still think if you really want to use an LLM to generate code the way to do it is to use a diffusion language model and a set of property-based tests, and jointly MCMC sample from the output space of the LLM (which is what diffusion is) and from the output space of the properrty-based […]

feed.hella.cheap

Original post on feed.hella.cheap

100

bob @bob.feed.hella.cheap.ap.brid.gy · 12/10/2026

FAR part 103 has a max speed requirement but it's only for level flight

000

bob @bob.feed.hella.cheap.ap.brid.gy · 12/10/2026

ultralight motor glider powered by cryogenic hydrogen. cryogenic hydrogen has incredible energy density, but not for long; it boils off while it's sitting in the tank. so you use most of it to take off and climb to as high an altitude as you can and store that energy as gravitational potential

100

bob @bob.feed.hella.cheap.ap.brid.gy · 11/10/2026

@qualia if I want to see what web pages are most accessed I shell in and run `cat /var/log/nginx/access.log | cut -d' ' -f11 | sort | uniq -c | sort -n`

100

bob @bob.feed.hella.cheap.ap.brid.gy · 11/10/2026

@kon google open sourced jaiku after they shut it down, you could use the original CSS to make a jaiku GtS theme

100

bob @bob.feed.hella.cheap.ap.brid.gy · 11/10/2026

right now I'm building on top of soulseek because bootstrapping a p2p protocol from zero users is really hard, but that would be a cool thing to layer on top I think

000

bob @bob.feed.hella.cheap.ap.brid.gy · 11/10/2026

if you have a representation of your preferences in terms of what kind of content you like (eg a probability distribution over an embedding space where more similar content is closer together) then that dovetails really well with a gossip protocol. you connect to some bootstrap peers, you send […]

feed.hella.cheap

Original post on feed.hella.cheap

100

bob @bob.feed.hella.cheap.ap.brid.gy · 11/10/2026

huge banquet, everyones having a nice time, everything looks extremely normal, and then a thousand people bust in and hang a massive banner from the rafters saying "Racist Mensa"

000

bob @bob.feed.hella.cheap.ap.brid.gy · 11/10/2026

I think it's because I like watching videos on machine learning and dynamical systems and there must be a very tightly connected cohort of people who are watching the same stuff but just for twitter posting boosterism

100

bob @bob.feed.hella.cheap.ap.brid.gy · 11/10/2026

youtube is kind of useless to me for anything but music now. the recommended videos are all extremely dumb AI takes. is it gell-mann amnesia if the next page is always more of the same wrong information?

100

bob @bob.feed.hella.cheap.ap.brid.gy · 09/10/2026

one of the most annoying things about AI hype is people hand-wringing about labor economics things that have been true for 30 years and attributing them to magical AI. production is already automated! clerical work is already automated! jobs don't exist because they produce things and haven't […]

feed.hella.cheap

Original post on feed.hella.cheap

000

Reposted by bob

Hynek Schlawack @hynek.mastodon.social.ap.brid.gy · 09/10/2026

holy shit I got a free NordVPN subscription and their macOS client installs an endpoint kernel extension that just hangs around after an uninstall. There was a constant com.nordvpn.macos.Shield process. I had to boot into recovery and disable SIP to get rid of it. Part of that is on macOS that […]

mastodon.social

Original post on mastodon.social

2125

bob @bob.feed.hella.cheap.ap.brid.gy · 08/10/2026

big study on claude code adoption using fine-grained data from over 700 firms. zero productivity improvement fion.ac/jellyfish.pdf

100

bob @bob.feed.hella.cheap.ap.brid.gy · 08/10/2026

writing LLM dialog is really hard because it doesn't have a voice per se, so you have to think of what would be in the pretraining data that starts with the previous thing that was just said

000

bob @bob.feed.hella.cheap.ap.brid.gy · 07/10/2026

why does it make any sense for your button-clicking bot to be able to recall the entire contents of ao3?

000

bob @bob.feed.hella.cheap.ap.brid.gy · 07/10/2026

one bizarre thing about AI agent discourse is everyone seems to be assuming that agent = LLM, when using an LLM as an RL agent is a real hack. LLMs weren't designed for that and people are only using them for it because a) they're LLM companies and their goal is to find applications for LLMs […]

feed.hella.cheap

Original post on feed.hella.cheap

100

bob @bob.feed.hella.cheap.ap.brid.gy · 06/10/2026

it's so frustrating when you see people talking about great capabilities of AI and look at what's possible for things you could easily do before AI, and in some cases things that were literally made before 2023

000

bob @bob.feed.hella.cheap.ap.brid.gy · 06/10/2026

this is the part of the thread where I'm supposed to pretend to have read A Thousand Plateaus

000

bob @bob.feed.hella.cheap.ap.brid.gy · 06/10/2026

and if you think that wouldn't work because the cities are all seeing and all knowing, I encourage you to look into the wars in syria and afghanistan and colombia's war against the FARC

100

bob @bob.feed.hella.cheap.ap.brid.gy · 06/10/2026

historically when cities get oppressive people head for the hills

100

bob @bob.feed.hella.cheap.ap.brid.gy · 06/10/2026

the weirdest thing about cyberpunk is it assumes that in a world where everyone is tracked all the time and constantly getting extorted and harassed and there's also plenty of resources and no need to work people would choose to live in cities

100

bob @bob.feed.hella.cheap.ap.brid.gy · 05/10/2026

@SnoopJ and distribution. what costs money is land and services you need a license to provide

000

bob @bob.feed.hella.cheap.ap.brid.gy · 05/10/2026

we've had post-scarcity for 30 years already

100

bob @bob.feed.hella.cheap.ap.brid.gy · 04/10/2026

is it the default in ubuntu again? do people even still use ubuntu on the desktop?

000

bob @bob.feed.hella.cheap.ap.brid.gy · 04/10/2026

no one is maintaining it because facebook captured all the old maintainers and locked them in a dungeon. the plus side of that is that (unlike onnxruntime) it isn't super buggy now because of slopification. the minus is that there's code rot that you have to fix to get it to build

001

bob @bob.feed.hella.cheap.ap.brid.gy · 04/10/2026

nnpack doesn't build anymore :/

100

bob @bob.feed.hella.cheap.ap.brid.gy · 03/10/2026

@SnoopJ @fooderick if you said strobing illumination synchronized to the camera shutter instead of hyperspectral that would have made sense to me but I feel like I'm missing what signal the other color bands would get you that justify going to a whole other semiconductor technology (which you […]

feed.hella.cheap

Original post on feed.hella.cheap

000

bob @bob.feed.hella.cheap.ap.brid.gy · 03/10/2026

@SnoopJ @fooderick that depends on what NN architecture you use but if you're doing classification and you're using, say, YOLO to detect objects and then VGG-16 to classify them, you'd tap off the vectors between the last conv layer and the MLP of the VGG-16 net probably

100

bob @bob.feed.hella.cheap.ap.brid.gy · 03/10/2026

@SnoopJ @fooderick it's straightforward to get one of those classifiers to tell you if something is off-distribution (even if it doesn't match the negative labels). you just tap off one of the inner embeddings, keep a reservoir of samples and use that reservior as a kernel density estimator

100

bob @bob.feed.hella.cheap.ap.brid.gy · 03/10/2026

@SnoopJ @fooderick when you say hyperspectral do you mean with a diffraction grating where you're doing spectroscopy or do you mean you want the sensor to be able to see into the ultraviolet?

100

bob @bob.feed.hella.cheap.ap.brid.gy · 03/10/2026

@SnoopJ @fooderick you mean people pick things up and tap on them to see if they're good? are you talking about cases where you're trying to estimate, like, mass distribution?

100

bob @bob.feed.hella.cheap.ap.brid.gy · 03/10/2026

@SnoopJ @fooderick I mean, CNN classifiers are a very mature technology now. if you have training data and the defects are things that can be seen by a human operator why wouldn't you do that?

100

bob @bob.feed.hella.cheap.ap.brid.gy · 03/10/2026

@SnoopJ @fooderick isn't that mostly near infrared (but still with cmos sensors)?

100