the way I'm doing optical flow works well, is super fast, and I think might actually be novel. it runs in real time on full res video in the browser on my ten year old laptop

it's been really interesting finding out who tries to figure out what's true and who tries to triangulate between the beliefs of the people around them

knitting 3D models https://www.youtube.com/watch?v=4rld9zOaUzU

why isn't running a classical computer vision heuristic, treating the output of that as a noisy result and running it through a denoising U-net done more often? a little bit of old-fashioned CV code can replace a ton of neural net compute

median software is worthless in the same way that median novels are worthless. there's no reason to do things that have been done when the marginal cost of replicating existing stuff has always been zero https://mastodon.social/@glyph/116422588020459518

the argument against investing in fusion power isn't that the technology can't work, it's that the advantage fusion power has over fission power is that fusion can extract more energy per kg from its fuel, and the fuel being too heavy or too expensive isn't the thing preventing more fission […]

whenever I ask someone who says claude code made them way more productive what they're producing the answer is either "refactoring" or the agent equivalent of making a really elaborate emacs config

the most disappointing thing about silicon valley right now isn't the widespread mania or the energy use or the rampant tone policing and groupthink, it's that no one cares about making things for people anymore. what made things work in the late 2000s was that when you're making consumer […]

the underlying LLMs haven't improved in like two years, systems like claude code perform better than claude does by using the LLM to do guided search, but there's a widespread belief that the LLMs have gotten better for some reason. when you point this out people tend to react by huffing and […]

new av

third party firmware that turns e-ink digital price tags into e-readers

I'm reading "There Is No Antimimentics Division" and I have to say I don't like it. tempted to write an essay titled "'There Is No Antimimetics Division' is Cthulu for liberals"

why does immersion usually get framed as subjectively being in a different environment but with the normal type of sensing and embodiment, when the point of art is usually to lose yourself

> Exclusive: OpenAI ‘Stargate’ Exec Peter Hoeschele Leaves Company > RIP

academics studying platform media really need to understand growth metrics. social media companies literally have a mathematical model that they use to make decisions about how their platforms are built and it doesn't seem to be widely known about outside silicon valley. to the extent that it's […]

it's funny that now that it's consensus that the datacenter build-out isn't going to continue as planned people are rushing to frame it as happening for reasons that aren't insufficient demand for AI (supply chain constraints, war in iran, etc)

"Try the New Experience" - a short story by github

AI is 1% algorithms, 8% data and 90% priming

> (33.41 kWh/gal) / (50 miles/gal) * 60 mph / 500W > 80.183999999999999995 m^4 s^-6 kg^2 (solar panels) this is why electric cars need batteries

how hard would it be to convert a geo metro to EV by putting four of those 12kw e-scooter motors on it

I love how everyone just forgot that anthropic got their seed funding from alameda research

"AI" is mostly fraud but there is actually a next thing and it's LFP batteries

while I was digging holes in the ground to find water I had a neat idea for a thing to do on my property there's a little gap between two hills (that was a river at some point but not in living memory) and I could put ground anchors at the top of the two hills, string steel cable between them […]

programmers are the people most susceptible to propaganda in the world

I'm now confident that at least the wasm backend for onnxruntime-web is just broken (at least for some combinations of ops). like, memory corruption broken. outputting uninitialized memory followed by a huge block of NaN regardless of what the input is

should I start making video essays about the same sort of stuff I make blog essays about? I think I figured out a visual style for them that I'd like (kinetic text, jump cuts, creative use of winamp visualizers)

two funhouse mirrors pointed to face each other is a generative AI system

why are raised beds so expensive? spending $100 on a piece of bent steel seems a little silly

old and busted: appollonian vs dionysian new hotness: affine vs hyperbolic

the five periods of american history (unironically)

shadertoy raymarching shader for minecraft thing where each voxel is a gaussian kernel

I have a little U-net type net that I'm trying to run in onnxruntime-web and it keep outputting NaNs which I never get from the pytorch version of the model. it's slowly dawning on me that this might be an onnxruntime bug

staring at a garbled image trying to divine how I screwed up the memory layout is my least favorite part of doing image processing

are there any horror games that use eye tracking? why isn't that more of a thing?

[brain rot] https://lore.kernel.org/git/CA+KyZp7KK74em6ad70kEd1xv9tLcK3vkC6momv4+2H_gyNrW1g@mail.gmail.com/T/#u

> About a half-dozen institutional investors — including hedge funds and venture capital firms that hold large stakes — approached his company in recent weeks looking to sell about $600 million of OpenAI shares. Last year, they would have been snatched up within days. But now, no one’s biting. […]

I don't think people appreciate how much of the AI bubble was funded by the gulf states

watching a video about medieval english forest law and realizing "forest" ala robin hood is just medieval english BLM land

hiking at the bottom of lake powell https://www.youtube.com/watch?v=s9q-LJ0x8p0

the little 130kb CNN is converging, you love to see it

people that are actually building talk about what they're building not about how productive they are