INSANE Voxel Renderer meets Fable 5.2
I had a feeling it maybe could do it. So Fable 5.2 ported my Cuda voxel renderer to Vulkan today. It made mistakes at first, as I made it do things certainly not in its training data. Though I have been sat on the techniques for many years - and it still did some things wrong that I have to fix😁
Running at 4k on a RTX5090.
Realtime path tracing of massive armies.
Opus 5.5 is great for quick experiments!
- Custom Vulkan rendering engine
- 4k 60+fps on a RTX5090
- 64 000 characters
- Nvidia DLSS-RR
- Clustered LOD
- Not fully optimised yet
Still figuring the right balance of hacks for path tracing quality vs performance. Using raytracing limits the size of armies possible.
@BeyondTheBoxCN The closer you are to them on the ground level means they occlude ones behind more, and the less of them fill the frame up close. So its actually faster in that case.
Quite likely something like that. But I would frame it the other way around.
As an evolution of conceptual spaces for problem solving, whose equivalent was previously confined to our thoughts in conjunction with tool use.
Note that an AI's latent spaces aren't necessarily (or usually) equivalent to our own.
"It's not committed to the depot" - not yet..
Respectfully you are too focused with the source code being the 'thing'. It's one possible manifestation of a much richer conceptual representation in the grand scheme of things. It's an artefact and a minor detail for future models.
You think it is only reasoning about the code - which is not what I am describing. In the same way the code you write is an expression of your understanding of a problem, which is a much richer representation.
"The higher level language is the context the agent uses to reason against." - No because it does the real reasoning in whatever latent space it uses.
They are going to use and cache the 'real context' more efficiently in a latent space of its choosing. Then manifest to in machine code or your high level language, to communicate with the execution hardware and yourself.
Get ready for the future of programming languages being mostly 'read only'.
Where agents directly write/edit machine code without the need of a compiler/decompiler. On demand converting it to your language of choice for you to read and discuss with them.
Earlier this year I already had experiments with pre-neutered Fable directly editing executables this way. Todays models probably already have the ability to do this if they were specifically post trained for it.
INSANE Voxel Renderer meets Fable 5.2
I had a feeling it maybe could do it. So Fable 5.2 ported my Cuda voxel renderer to Vulkan today. It made mistakes at first, as I made it do things certainly not in its training data. Though I have been sat on the techniques for many years - and it still did some things wrong that I have to fix😁
Running at 4k on a RTX5090.
Perhaps this sentence does not translate well for you: "Fable 5.2 edges closer to AAA". The emphasis on the 'edges'.
AAA games are just a description of budget at the end of the day. Not a guarantee that the game has any 'soul', in fact quite often the opposite.
There were plenty of games since the 1980's that have been visually polished but utterly devoid of soul - ironically a lot that were advertisements for fast food!
I think you should consider that AI can liberate a group of people to focus on the substance and soul even more than before. And it allows people to explore creative avenues that they did not have the resources for previously. Sure there will be lots of junk too, there always is.
Fable 5.2 edges closer to AAA
The game:
- Tried to sortof match an example doing the rounds that seems sus..
- Custom experimental renderer (4k path traced DLSRR)
- C/C++, Vulkan, Box3d, Miniaudio, SDL3, base framework
- Not using my Vibemaker engine
- Downloaded SFX every other asset is Fable
- About 14 hours with multiple prompts
- Aiming/movement/feel need tuning more
On Fable 5.2, the good:
- Coding noticeably improved
- Texturing quality massively better
- 3d understanding for modelling improved
- 3d sculpting has slightly advanced
- Seems more token efficient now
The not so good: It seems to have lost some aesthetic quality in 3d?
It feels like whatever data the GPT models are trained on that makes their results subpar has been given to it - proportions/ratios/volumes. GPT models naturally tend to uniform boxy proportions.
In conclusion good they called it 5.2 not 5.5!
@richgel999@toncijukic Sorry, I don't follow. I was actually saying that my current understanding aligns with Hinton's.
My question was whether you are aware of research providing a more concrete scientific basis for what Hinton said - as I simply haven't come across that so far.
@richgel999@toncijukic Yes, so far I see it as pretraining creating a nutrient rich primordial soup. And the RL post training etc creates a useful focused identity/construct from that.
I am not aware of anyone knowing for certain anything past that at this stage?