Because you communicate with it using natural language and real-world references and descriptions of what you want, you use emotion and emphasis (especially when re-prompting), you use examples and illustrative stories and common expressions. Understanding and interpreting all of that and replying in kind, to some degree, requires a large body of non-computation, cultural knowledge, or else the prompts are just meaningless words, and the replies will look like compiler output.
That sounds intuitively true, but I’m not convinced that it is actually the case. I don’t think we know enough about neural network training to say what training and how many parameters are necessary for what kind of performance on which tasks. To me it looks like we currently guess that more is better and try to throw as much compute and data at the problem as is economically feasible. There is little incentive for companies to invest into small model research since their moat is huge models that require special hardware to run.
For every ten kill X quests you get a Mankirk's Wife quest or a Hogger and it all works out great. 80% of almost everything is filler, like our bodies are 70% just water, but it's the taste that matters.
It remains to be seen how much of that social element is taken over by LLMs as well. There's already plenty of stories about people retreating into LLMs. Whatever shape and form this all ends up, we're just at the beginning, and the only limit will be how economically and socially sustainable it is (chances are: it's not).
For floors the perspective divide is once per row, just like for walls it's once per column.
The BSP may have led to some floor subdivisions, especially as it needs convex sectors. I don't remember if the engine would coalesce adjacent floor spans into a single one, but I hope it did.
That would only be true if a row (by which I mean a scanline) would be equidistant in view-space depth across its whole length, which is not quite true. While a column of pixels for a wall is (as long as you dont tilt the camera).
And it looks to me like we are mapping each row with a constant y, calculating the "distance" (thus scale factor) only once using just the vertical slope for the row.
For the record, there were constant-Z full 3D engines in the late 90s, which would find the correct axes on screen to render perspective-correct textured triangles using oblique spans of pixels. They were incredibly complicated and prone to holes at the slightest numerical accuracy, but they were a sight to behold. Constant-Z meant not just saving on perspective divides, but also easy Z-buffer and depth-based fog.
Author here. Yes, it is integral. I chose this approach to first show how to draw it from back to front, because the code is easier to understand this way.
I think it's worth noting that you can tilt with this method, but not roll.
Great for a helicopter game, Less so for general flight sim.
That was a large part of how games were designed back in the day. Start with what you have the ability to do, find the game that matches what you can do.
Notice the demo video from Comanche also shows roll.
Edit: To support roll, the renderer essentially rendered the voxel terrain into a frame buffer and then applied camera transformations that gave the appearance of a fully rotating viewpoint. The terrain itself was not being raycast through a true 3D voxel volume.
Reverse painters algorithm is still painters algorithm. You trade off the cost of a full screen clear before the frame, in return for eliminating overdraw
that's what the y-buffer is that the article mentions in the front-to-back rendering section.
it tracks how tall each columns write is so you can use it to only write the diff between it and the voxel behind it, skipping writing anything at all if the voxel behind is shorter than the current height.
So once you're done rendering front-to-back, you've got a y-buffer of highest-writes you can slap your blue sky across from highest-to-screentop on each line, avoiding the need to clear by write the sky to the full screen before starting the render.
The business need of the games industry is making money selling games and content and ads, just like the business need of Netflix or Spotify is making money via ads and subscriptions for music or movies or etc.
It's a consumer business much like any other. Just like most startups and major companies, they are not necessary for the world in the way utilities for example are.
The problem of videogames compared to startups and SV tech is that the long-term money potential is very limited at best, and rapidly becomes very brittle. Most startups pay bigger salaries for much easier work, because they burn the money investors are betting hoping the company will crack a new long-term market, not because they make money themselves. There's very little games market to crack, very little chance to turn your product into a long-lived platform to built on top of, meanwhile the upfront investment is huge.
Hundreds of millions of people from abroad shared that belief up until 2 decades ago or so. I don't think they believe it anymore. It's been like watching your awesome high school friend throw away their lives over time.
> Fewer people applying for patents, because the minute you apply for the patent, it's available to everybody, which means every model can train on it
We know LLM companies have, for lack of a better word, "sidestepped" the copyright on millions of works with their "transformative fair use" arguments. Are LLMs also a way to sidestep patents?
LLMs are accelerants. They enable people to do patent and copyright infringement at a much larger scale. As we know from previous examples, if you break the law enough as a company eventually they have to let you keep doing it.
I don’t see how? You can train on something pending patent, but what are the benefits? If it gets patented you open yourself to get sued, I don’t see how AI works around that, the idea itself is still patented? I think I’m missing something for the argument to make sense. Or is the idea that if too many people use your patented idea you won’t be able to enforce it? That sounds risky to me
Patents are public. Ingesting and innovating on them is the intended use. If you use an LLM to then make and market something that infringes on a patent, that isn’t the LLM doing any infringing, it’s you.
> are you even aware that you're infringing a patent?
Plenty of folks first learn they’re infringing when they get a demand letter. Unless you ask it, I’m not sure it’s on the LLM to search for prior art and patent conflicts.
What a funny perspective - they didn’t side-step copyright, they blatantly infringed without financial consequence. The interesting “upside” is none of the generated works are protected by copyright. So it’s a bizarre conundrum which goes to show the complete disconnection between the concepts of the original intent - to protect authors and creators - with the warped capitalist mechanics of “rights holders” like Disney buying political influence for regulatory market capture.
Sugar coating the discussion is for children and dishonest ethical rationalization, in my view.
reply