Avalon local LLM

[Pocster]

Oh yes and the amazing cube marched clouds. VERY expensive! need to optimise again soon !. Lit!, can have a nice storm with them! Which means a detour (as usual) to .... lightening!!!

[Pocster]

Pocster's

Unreal

Kafkaesque

Engine

 

P.U.K.E (tm)

[Pocster]

2 hours ago, Pocster said:

Pocster's

Unreal

Kafkaesque

Engine

 

P.U.K.E (tm)

PUKE now does God rays, lightening storms, ambient occlusion , terrain can influence cloud colours and likewise; same for water etc. So all in the same render pipeline. Limits in webgpu are no more than 16 sampler passes for the march. This limit has been a constant problem. Also back at a crippling frame rate if you have max clouds/max rain and a lightening storm and decide to look at a grazing angle over water!.

[Nickfromwales]

9 hours ago, Pocster said:

Also back at a crippling frame rate

Time to upgrade the machine then........

[Pocster]

6 hours ago, Nickfromwales said:

Time to upgrade the machine then........

Amateur ! Time to analyse shaders ! Optimise 💪

[Nickfromwales]

3 hours ago, Pocster said:

Amateur ! Time to analyse shaders ! Optimise 💪

And then realise your FPS is still crap, and then it’s…….

 

10 hours ago, Nickfromwales said:

Time to upgrade the machine

 

[-rick-]

Just now, Nickfromwales said:

And then realise your FPS is still crap, and then it’s…….

 

Most code these days is horrifically optimised. Someone who knows what they are doing can often find orders of magnitude speed ups if they look hard enough. IIRC @Pocster used to work in game development and to get this far even with AI is, I think, showing that experience. There are likely a lot of speedups still to be found even in webgl which is hardly the fastest way to do 3d graphics.

 

What I want to know though is what happened to the original plan? This is all very pretty but I don't think the mountains are telling @Pocster what's for dinner or where he left his socks. Which I thought was the starting point of all this

[Nickfromwales]

I thought it was just to water his plants.

[Pocster]

35 minutes ago, -rick- said:

 

Most code these days is horrifically optimised. Someone who knows what they are doing can often find orders of magnitude speed ups if they look hard enough. IIRC @Pocster used to work in game development and to get this far even with AI is, I think, showing that experience. There are likely a lot of speedups still to be found even in webgl which is hardly the fastest way to do 3d graphics.

 

What I want to know though is what happened to the original plan? This is all very pretty but I don't think the mountains are telling @Pocster what's for dinner or where he left his socks. Which I thought was the starting point of all this

Theres always ways to optimise. My main lever is of course the shaders (even though Claude wrote them) and pre-baking - 256Gb to burn so we create lots once at runtime to use later and save on some computations. The original premise is still there - just got a bit distracted. But yes the Alex/Chatgpt from Avalon 1 will be imported into this. The robot running over these luscious landscapes will tell you when your chicken is ready and then you're playing Ed Sheeran AGAIN! 

Edited 8 hours ago by Pocster

[Pocster]

34 minutes ago, Nickfromwales said:

I thought it was just to water his plants.

That was an esp32 project and works!. I'll integrate that into Avalon later. One thing at a time! ( well lots of things at lots of times really - so never get bored!)

[Pocster]

38 minutes ago, -rick- said:

get this far even with AI is, I think, showing that experience

Tell you what - that's a massive understatement. What I mean is I've created a custom 3d engine that would be commercial grade - yet not 1 line of code. That is the real power of AI. I see plenty of youtube/reddit "wrote this in 5 hours with Fable from 1 prompt". Some are believable some are slop. But most are relatively simple (still impressive that Ai can take "write me a really good minecraft game with dinosaurs in it" and still produce something). But use it properly from experience and build it up layer by layer rather than the final product from 1 prompt and its a phenomenal tool. The accomplishment here is WELL beyond my technical ability. But the "difficult" parts the code and horrific maths aren't my burden 99% of the time. I'm still in awe of what's possible. Must also admit it can take some effort - occasionally days on 1 feature to be correct. Or bastard bugs... we had a few mountain tops going black as you approach. No obvious reason some would and some wouldn't. Check all the render stuff/textures/lighting everything me and Claude could think of. Nothing obvious. Found it out of frustration!. The ray march on water can shoot a ray so far that it can hit a mountain top. Essentially at that point its a NaN so plots black! - what a bast!. Water affecting a mountain maybe 4km away in real space. That kind of bug finding without Claude (as I've done many a time when a software engineer) its pure torture and can literally take weeks to find...