I've read quite a lot about salsa, but missed that post! Thanks for sharing. I've definitely thought about adding some adaptive walk-minimizing stuff to hornpipe. It would help a lot since walking around the graph increases thread contention a ton. But I've higher priority stuff for MVP lol
Wow, I love this! I implement the push-pull algorithm pretty much exactly as described in gitlab.com/samsartor/ho.... Although yesterday I experimented with changing the dirtied flags to logical timestamps, to try and fix some concurrency-related glitches.
Just published docs.rs/weakref/, a Copy+static reference type which checks liveness at runtime!
I saw verdagon.dev/blog/surpris... a few years back and started using a similar weakref implementation in my own weird projects. But now it has its own crate 📦
The point is, hallucinations aren't some willful conspiracy by ML people to ignore the philosophers. We just don't have all the tools yet. It is a very young field.
There are a bunch of subtle reasons why no one has figured it out. RL algorithms suck. LLMs are fundamentally incapable of learning from experience. Optimization on symbolic logic sucks. Data of (world model state, language/image/video) pairs is practically non-existent. And so on...
Combining symbolic and learned models is like ML's version of combining general relativity and quantum mechanics. Everyone wants to do it. Every first year PhD thinks they'll be the one (I sure did). And it turns out to be really hard.
People I know would be thrilled to have LLMs on top of some sort of explicit world model and so correct all hallucinations. You'd win best paper at NeurIPS easily, and make every existing LLM obsolete overnight. Loads of explicit world models have turned up in brainstorming meeting I've been to.
Just read open.substack.com/pub/garymarc... and there is a take really I want to unpack.
> For what I think are mostly sociological reasons, people who have built neural networks such as LLMs have mostly tried to do without explicit [world] models
We Don't Want to Talk by Friday Pilots Club has such a baseline I'm obsessed. They are on tour rn!
youtu.be/k0hcWl9pNm0
the real paperclip maximizer was the MBAs all along
I really don't like Three Body Problem. I was promised metaphysical dread in the face of a fundamentally impossible unpredictable world, and got <spoiler>it was aliens doing a prank lol</spoiler>. Does anyone know a book that executes on the "science suddenly breaks" premise for real?
I'm also curious about the coordinate systems that particle sims are using. Most rooms seem to be very long rectangles projected onto curvy walls. But a couple obviously wrap around. And while some rooms had wall/floor seams, they were very well hidden. Also, one room is volumetric! Cool!
Question for any CG folks who've seen it: how are the large scale projections handled? Is one machine doing all the rendering with like 50 HDMI cables out to the projectors? Or are multiple synced up somehow?
Room full of orbs
Starting at SIGGRAPH and ending at LeamLab Borderless has been really nice. Computer graphics book-ends my Tokyo trip!
Honestly, this was the most disappointing outcome of the election for me. I was sorta numbed to a Trump win, but I've been preaching RCV my whole adult life. And after the asskicking it got this year, I doubt we will ever see it on the ballot again.
I setup btrfs on my Linux laptop to get all the cool timeshift stuff. I didn't expect it would also be 100× easier to mount when dual-booting windows!
github.com/maharmstone/... is badass
Ok hear me out, C is typescript for assembly
I looked it up and that lawsuit is still actively ongoing www.courtlistener.com/docket/68117...
Regardless, I think it is forcing AI companies to be a lot more careful about memorization
It also seems that transformer-based models only copy training data when the same token sequence appears repeatedly in the corpus. SD1-era models were notorious for memorizing popular artists, but I haven't seen anything similar out of post-NYT-lawsuit models like Flux.
The "ai-generated art is stolen" argument came up in a conversation today. I think it is bad because regardless of training data I'd still still object to:
- People passing off AI art as their own
- Bots flooding the world with "content"
- Automating away fun parts of life
I'm glad Pieter isn't on Bluesky, or he'd be giving me grief for not exporting a video hours ago lol
"Don't worry, people can open the slide deck fine! Just make sure they launch this VS project first"
I went mad today trying to embed a WebGL canvas into PowerPoint! Surprisingly learn.microsoft.com/en-us/office... supports WebGL just fine. But the logistics of actually inserting into a slide and saving plugin state are insane.
Don't follow me into this nightmare
Yah, to me "synthetic data" is still "this big pile of blender renders". But I think we lost the terminology battle on this one
I honestly gave up on PIL/skimage/iio/torchvision/cv2 and started doing all my image processing in Rust with the amazing image crate, then yeeting a numpy array over the wall with pyo3. I had to write my own transforms, but that wound up being really useful anyway.
My first grad school paper (samsartor.com/matfusion) was so much more work, but way less fun to look at. Funny how that goes!
The actual project (github.com/samsartor/content_aware_tiles) is all pytorch shenanigans. But I made the video in TS with motioncanvas.io. The tiles are rendered and animated with fragment shader crimes!
Since I'm bootstrapping a profile here, I figure I should share the coolest thing I made recently
Sam Sartor, Pieter Peers
Content-aware Tile Generation using Exterior Boundary Inpainting
https://arxiv.org/abs/2409.14184
