Surface Cache GI running in a browser context. Much more detail than DDGI and generally a little more performant. Still gotta figure out what to do with residual disocclusion noise but it's nearly there. Runs pretty much at 60fps on all my test scenes on my laptop 3070.
Still too heavy for something low-end like mobile due to the software ray traversal strategy so working on a hybrid light mapper for low end platforms.
@Hartguy85@DJSnM@sjdhatters Never said "after one flight". Either way, even with just 2 starships (there will likely be more in simultaneous operation) the cost will still be dramatically less than any current alternatives
Wouldn't they still need to replace heat shield tiles, refurbish engines, constantly replace potentially unsafe structural damage after a few launches, maybe even faulty forward/aft flaps, etc? I don't imagine these things would require absolutely no maintenance, so they probably wouldn't be rapidly reusable on average. You would need thousands of starships first before you can pipeline rapid reusability.
In programming you typically specify a lot of code in terms of "operators" or "functions". Even under the hood an ISA is just a bunch of operators. The interesting thing about ML models is they learn new functions over a set of data. If you know exactly how you want to pipeline different disjoint sets of data and different data domains, you could program an entire piece of software by training tiny models continuously on the input / output that's expected, and then daisy chaining the models together.
In theory if you do this in a sort of online way you might be able to create an evolving, intelligent meta structure (i.e. a model) of the software in question.
I would imagine this could be like a middleware interpreter that continuously RLs away unwanted signals (no crashes, no slowdowns) and learns the correct representation of your software invariants in real-time and effectively becomes "self-healing" software when new data (from users) triggers certain signals, like internal asserts, error messages, etc.
The real far-fetched idea that would work best here would be something like molecular machines (i.e. deposit individual atoms, create tiny molecular factories, etc. instead of etch features by burning some substrate with light). Something like what biology already does but controlled by us, down to the atom, via some sort of programmatic framework. We are very far from this though.
@ChombaBupe This is what online learning is all about. Continually refit the distribution (as fast as possible) to consistently model the dynamic nature of reality. Not all AI is built the same.
@thefrontendcat Seems like you might have an issue with the virtual-to-physical mapping for lower clipmaps when writing shadow pages. Other than that it does run pretty snappy on my Pixel 10.
@mannyn83 Interesting, yeah I have something similar set up in my engine (software RT with per-instance TLAS/per-mesh BLAS) and instruction + data divergence is still a huge issue especially for moderate to large complexity scenes, where hit traversal alone can run upwards of 10ms.
As someone who's been working on an engine in JS + WebGPU for several years I agree on all accounts. You would need a custom V8 implementation to get close to C-like performance (and completely get rid of the GC), WebGPU still needs support for way more features, and browsers are too general purpose to be a good fit for games distributions (so you would need to make a custom, game-native browser in a sense).
Part 1: AI trends toward doing all economically viable work (including intellectual work), in which case you're free to pursue whatever creative interests you want with all your newfound free time! (remember that there are still people that love woodworking as a hobby!)
Part 2: AI uber-companies autonomously run the economy, so now we need to fix the government problem. Human government here would trend towards distributing the wealth in a disproportionate way (favoring the more influential/powerful) or totally fuck it up due to negligence, so how do we fix this? Autonomous government! One that can most effectively redistribute resources and enact legislation for the benefit of all, using a truly fault-tolerant democratic base.
Part 3: Abundance, prosperity and a truly well oiled socio-economic machine!
LOL yeah because we weren't already remaking the same handful of game ideas and game mechanics before AI. Revisiting the same "the world has fallen into darkness" hyper-serious fantasy tones again and again. The same shooter with little variation and a different skin. The exact same tropes of RTS, FPS, RPG, MMORPG, combat this, combat that, over and over again.
Yeah sure AI slop is a real thing, but human slop has existed for a much longer time. Taking into account the number of people on the planet, humans already have a hard time with "originality", with or without AI.
Super Smash Bros creator Masahiro Sakurai says AI-generated games just copy ideas and systems from existing games
He believes AI can't match the "originality" of human creators
@sampavlovic Single shot 🔥 so many questions: what does your (ir)radiance layout look like? Is it just world space probes or do you have screen space probes in the mix or are you just doing 1spp rays? Is it multibounce? Using HWRT? What is the performance like and on what machine?
My initial reaction to this was "very fucking cool but when am I gonna need to do short burst hops like this ever?"
And then it immediately hit me that we might be witnessing the beginning of not only earthbound, long range flying cars but the first step in the minituarization of personal space travel technology.