riotnrrd

CVPR submissions were two weeks ago.

This is what we did at ILM, with our in-house software package Zeno. The lowest level of complete functionality was the "Sg" namespace (short for Scene Graph). Above that lived user-and-SDK-friendly commandline utilities layer (SgCmd) that was used for all scripting. Above that, finally, was the Oz layer -- the graphical UI. Oz actions called SgCmds, but never saw or touched the Sg layer.

In theory, Oz was UI toolkit agnostic but, as anyone who's ever done UI work will understand, it rapidly became wedded to the toolkit we used (a heavily customized version of Tk). This made the eventual switching to Qt very, very difficult. So difficult it nearly destroyed the artists' faith in the R&D group. (Extremely late delivery, buggy, and coming at the manpower cost of not doing any improvements on the tool itself for a year and a half.)

I love this game so much. First of all, it's kind of technologically astounding. You can get in your spaceship and take off from a planet and fly to another one and land, all without any loading screens. The planets themselves are sculptable isosurfaces (only to an extent, though; you can't drill a hole through to the other side) and littered with points of interest, unique animals and plants.

It's quasi-social -- you can play with friends or go on missions with randos if you want -- but fundamentally a mellow, solitary sandbox to play in however you'd like. I like to collect neat looking spaceships, some people build amazing bases, and others basically use it as an outer space dog fighting game.

It's absolutely worth picking up if you want to putter around for a hundred or so hours in outer space.

Not true. According to an article on Rock Paper Shotgun (https://www.rockpapershotgun.com/stop-blaming-teeth-for-citi...) the developers "have put out a statement denying this "bizarre story."

"Citizen lifepath feature does not tie to citizen geometry and does not affect the performance figures of the characters," the statement reads. "We know the characters require further work, as they are currently missing their LODs which affect some parts of performance. We are working on bringing these to the game along general LODs improvements across all game assets. Characters feature a lot of details that, while seemingly unnecessary now, will become relevant in the future of the project.""

Micro-Max (https://home.hccnet.nl/h.g.muller/max-src2.html) is a chess engine written in C (2000 characters!) and has an ELO of about 1900 and plays only legal moves.

If you're curious, you can read the relevant laws yourself. I suggest you start with the STOCK act of 2012 (S. 2038, 126 Stat. 291, enacted April 4, 2012). This law prohibits the use of non-public information for private profit, including insider trading by members of Congress and other government employees. You can also follow up with the relevant SEC definitions of "non public information."

I work as a researcher (ML for robotics) inside a product company so I can speak to this a little. The research I work on generally takes a year between conception and publication (three months of tinkering, six months of refinement and benchmarking, three months to write and submit a paper). Now, I work on more than one of these projects at a time but they all have a similar cadence (conference deadlines do that for you), and can be "measured" by management the same way a year-long engineering project is. The questions I'm asked (or ask myself) and the same as when I was a pure engineer. Are you making headway? Are you on schedule? Do you need help? Should this project be abandoned or rescoped if it's not succeeding? Etc.

> one thing about the generative models is that you could in theory get the model to recite copyrighted work, word by word.

Only if it's so badly trained as to be considered broken. It's not fair or smart to judge an entire method (LLMs) by some broken examples. I wouldn't claim "all cars leak oil" just because I saw a broke-down beater one day.

My main confusion was with the claim that one person did an entire "opening sequence" of a major film. That's simply not possible. It's like claiming a skyscraper was built by one guy.

> by someone that had done the opening sequence for a major film

I used to work in VFX (tools programming) and this makes no sense. Major films have VFX crews that number in the hundreds; nobody does anything on their own. Even if they were only doing a credit sequence, that's dozens of people's work over weeks or months.

And "never saw a cent"?? Sure, VFX studios don't get residuals or a percentage of gross, but they do get paid.

Although the simplicity is appealing, this theory isn't supported by the evidence. Read the 'Chemical Hunger' series at that site, starting here: https://slimemoldtimemold.com/2021/07/07/a-chemical-hunger-p...

I worked on tools for artists and you're totally right. The fast iteration was so great; it felt more like good ol' fashioned stay-up-all-night hacking for the joy of it than anything else I've ever done professionally.

I worked in VFX for almost ten years. It's incredibly fun! I absolutely loved making movies and I worked with some amazing artists. It pays terribly, but it was the most fun I've ever had in a job.

No. The VFX industry in the US is not unionized.

> To be fair, it's already been torpedoed due to GDPR

Not really. I'm a deep learning researcher and my work has to be palatable to the GDPR. You can remove personally identifying information from training datasets with some effort (replacing faces with GAN-generated fake ones, ditto with license plates and telephone numbers, etc.) It's just another aspect of data cleaning and not that big a deal.

I mostly have experience with Instant NGP, and it should work on older consumer NVIDIA cards. Their github page calls out Pascal cards as working, for example. 6 GB isn't much memory, though, so you may be limited in final resolution of the latent model and thus of the output.

Reinforcement learning with a dynamic world model. It's a pretty standard approach to training agents, although there are a lot of problems with this technique as well. People have used it to create agents for Minecraft, racing games, Nethack, lots of old arcade games and so forth.

It's not an issue of "alignment." The reward function was poorly written. That's it. It's like claiming the Ariane 5 explosion shows that FORTRAN needs "alignment."

Yeah, the Frugal Man's Bullet-time is a DSLR, Instant NGP from NVIDIA, and COLMAP.

Shutterstock and Getty are both well positioned to sell pre-trained (on their own data, natch) generative models to customers, who can then fine-tune with their own datasets. I think the market for "pretrained models with guaranteed no IP issues" is going to be lucrative. Maybe not enough digits to excite VCs, but more than enough to make up for a decrease in demand for stock images.

It's not related. The Florida move was to be Imagineering; these layoffs do not touch that group.

This paper is very good and the results are quite impressive. It's especially good news for smaller labs and hobbyists, as these small language models take significantly fewer resources to train. To quote the paper: "the training of SLMs on TinyStories can typically be done in less than a day on a single V100 GPU." This is opposed to LLMs such as ChatGPT which can take weeks to train on hundreds or thousands of GPUs.

The neural network inference (and training) is all done in compiled C libraries called by Python. The inefficiencies you're seeing are because you're doing it on a CPU, probably in one thread, rather than on a GPU with tens of thousands of "threads."

Not the OP, but the podcast opens with a pull-quote of sorts (spoken by the host) about how LLMs, for the first time, allow people to program via natural language. This is an exaggeration, but not much of one. I've generated usable Python code with ChatGPT, for example.

I think it's important to note that the whole point of having an opening like this is to intrigue you into listening to the podcast. So of course it's going to be a slightly shocking quote.

They are licensed: https://nvidianews.nvidia.com/news/shutterstock-teams-with-n...

I've been trying to figure out what system this is or, if it's a homebrew hack, what it's based off. Not Call of Cthulu, as they note in the article. Not GURPS (wrong type of stats and 'Faith' isn't a GURPS stat), or World of Darkness either. Obviously not D&D of any edition, nor its variants (DCC, AS&SH, LotFP). Not Troika, Vaesen, or Mork Borg or any indie system I'm familiar with.

It could be an indie system that doesn't have much Enlgish-language presence, but the player was using English names for abilities.

I'm a deep learning researcher so I can speak to this a little. Until roughly 2010 neural networks were barely used and broadly considered to be less efficient and less accurate than rival techniques (decision trees, support vector machines, etc.). For decades, far more work was put into these other approaches but the amazing thing about neural networks is that, even in a relatively primitive and neglected state, they outperformed traditional techniques. AlexNet is the poster-child of this phenomenon, but look at facial recognition, speech recognition, etc. These are all fields where older approaches are just dead. Nobody is working on SVMs for facial identification or object pose estimation anymore.

It's more accurate to say that neural networks succeeded against the research darlings of the time despite being relatively ignored for decades.

Neural radiance fields (NeRFs) is a technique to create models of complex 3d geometry from a set of photo/camera pose pairs. The way it usually works is that the 3d geometry you're interested in (a car, a tree, etc.) exists within a static, bounded space (e.g. a unit cube). This research changes that requirement, allowing you to create larger spaces without the memory and computational drawbacks of "just making the bounded space bigger."

Not to get into the unrelated weeds here, I think it's the specificity. We say "nuked" but not "Hiroshima'd."

Oh, please. This is such a dismissive, bad faith statement it should be deleted. Nobody at the NYT "loathes" the "average" American. They do not get sexually aroused when reporting tragedy. Be a better person.