Avalon local LLM

[Pocster]

Accusations from chat that I changed the code!

Accumulative wetness at very little gpu cost; I like my wetness .

This of course is leading to accumulated snow effect

[Pocster]

(expletive deleted)er just told me to stop arguing!!! WTF!

[Pocster]

accumulating snow actually build height, and I can make it piss down.

Best bit I can blend between weathers. So snow will melt as it gets wet as blended to rain. Rain will dry as its blended to a sunny day. TBH it does look AAA. Masses of optimisations to do as chat's been fibbing a bit.

[Pocster]

Then all of a sudden chat becomes shit. Quickly checking reddit seems a current issue. basically its lazy, misses obvious things, ignores evidence - so like a real person now ....

[Pocster]

Give me strength!. Discussions about an efficient looking tree.... and then it cant render a (expletive deleted)ing cube!

[Pocster]

and then I want more.

[Pocster]

And I still wanted more but performance number 1!. Really good lod on tree. Shadow types depending on distance from camera. Near trees leafs have shadows also. Tree sways in wind etc. I did find a glb for this and had to do some work to get it sane. Will probably compromise my snow accumulator now but for awesome looking tress I'll take the hit.

[Pocster]

Chat being a shit today. I told it multiple times to (expletive deleted) off. As it would "this isn't proceed able ", or "it's too many bodged patches". You can imagine my responses!

I told it to man the (expletive deleted) up, stop bitching and being shit because you've been down graded as 5.6 is coming out and to make a sentence from these letters 'FCUK OFF". 10 minutes later we progressed!

Doesn't look much I know. But its a procedural landscape chunk based i.e. it's deterministic and infinite. Colours etc wrong; terrain variation limited at the moment. But this will be beauty! then I can cover it in tress and (expletive deleted) the frame rate! (I wont of course (expletive deleted) the frame rate). 

I've noticed it's shit at efficient techniques. We were doing complex things and it worked; I asked for a full breakdown of the render pipeline . 90% was on the cpu!!!. I said "you (expletive deleted)ing prick!". few hours later 90% on the gpu. It wants to "keep" the cpu code in!. I guess it's designed not to remove legacy stuff. Took a few attempts to get it all out not just set a flag to FASLE to avoid it. I call it bodger most the time. Again it's very human like!

[SteamyTea]

Have I lost the plot a bit here.

I seem to remember that you wanted to automate the toilet light.

Now you are creating an alternative universe.

 

I have a tree in the garden that sways in the wind, and the shadows are in the right place (over my neighbours garden).

I left a light in and went to work, came back several hours later, light was still on. 3W x 13h = 39 Wh.

Edited Saturday at 09:18 by SteamyTea

[Pocster]

1 hour ago, SteamyTea said:

Have I lost the plot a bit here.

I seem to remember that you wanted to automate the toilet light.

Now you are creating an alternative universe.

 

I have a tree in the garden that sways in the wind, and the shadows are in the right place (over my neighbours garden).

I left a light in and went to work, came back several hours later, light was still on. 3W x 13h = 39 Wh.

The plot is flexible but the objective is the same. Super home assistant with voxel landscape backdrop!!!

Edited Saturday at 10:28 by Pocster

[-rick-]

2 hours ago, SteamyTea said:

Have I lost the plot a bit here.

I seem to remember that you wanted to automate the toilet light.

Now you are creating an alternative universe.

I think you’ve missed a key part of the specification which is to keep @Pocster occupied and provide an easy distraction from finishing his house. Everything following that is just part of the journey. 😜

[Pocster]

Just now, -rick- said:

I think you’ve missed a key part of the specification which is to keep @Pocster occupied and provide an easy distraction from finishing his house. Everything following that is just part of the journey. 😜

Exactly. Give yourself an unrestricted project and the house never gets done because you're too busy!

[SteamyTea]

18 minutes ago, Pocster said:

Give yourself an unrestricted erection

 

[Pocster]

look at chats (expletive deleted)ing code! 

 

   mediaArtworkReflectionClone = mediaArtworkGroup.clone(true);
1162-      mediaArtworkReflectionClone.userData.mediaArtworkReflection = true;
1163-      mediaArtworkReflectionClone.traverse((object) => {
1164-        object.userData.mediaArtworkReflection = true;

=== BODGE CHECK ===
bodge_absent=YES

=== NODE CHECK ===

=== COMPILEALL ===

 

BODGE CHECK absent = YES ffs!

[-rick-]

16 minutes ago, Pocster said:

look at chats (expletive deleted)ing code! 

 

   mediaArtworkReflectionClone = mediaArtworkGroup.clone(true);
1162-      mediaArtworkReflectionClone.userData.mediaArtworkReflection = true;
1163-      mediaArtworkReflectionClone.traverse((object) => {
1164-        object.userData.mediaArtworkReflection = true;

=== BODGE CHECK ===
bodge_absent=YES

=== NODE CHECK ===

=== COMPILEALL ===

 

BODGE CHECK absent = YES ffs!

maybe swearing at a computer program isn’t the most effective tactic

[-rick-]

6 minutes ago, -rick- said:

maybe swearing at a computer program isn’t the most effective tactic

and this is not just making a joke. It’s likely that the training corpus for the llms has the best responses in relation to professionals who don’t tend to use that language. By swearing at it you likely trigger different statistical paths within the training that have less optimal outcomes.

[Pocster]

Just now, -rick- said:

and this is not just making a joke. It’s likely that the training corpus for the llms has the best responses in relation to professionals who don’t tend to use that language. By swearing at it you likely trigger different statistical paths within the training that have less optimal outcomes.

I think I get better paths! 

[Pocster]

1 hour ago, -rick- said:

and this is not just making a joke. It’s likely that the training corpus for the llms has the best responses in relation to professionals who don’t tend to use that language. By swearing at it you likely trigger different statistical paths within the training that have less optimal outcomes.

But it would seem it genuinely does give you cheapest compute answer. Equally it has knowledge of technioques etc but it never makes the leap to them as a suggestion. But as soon as you say "why dont we use....." suddenly it kicks a whole new path in to explore. Problem is it thinks all my suggestions are good - as it's designed to pamper human ego. What you really want is to say "no, this is the wrong approach because...". I tell it off it should tell me off.

Edited Saturday at 15:04 by Pocster

[Pocster]

my world! I mean home custom Alexa intelligent system.... 

[SteamyTea]

 

 

 

[Pocster]

An interesting day

Took 2 hours to "undo" some of its shitty code.

I didn't look at the code but it's so hard coded and engraved that removing it broke everything. But doing it in baby steps worked.

Then of course it blew my mind when I defined the structure for our procedural world. Implemented the Skeleton no problem.

All good but 6fps!!!. I couldn't work it out. Told it that we are lead SE for Nintendo on a Switch 2 Super Zelda game that destroys Breath of the Wild!!!

It now says things like "yes, this would be correct for a AAA title"

(expletive deleted)ing idiot!

After much work found (christ knows why) for a certain shadow it was sampling the terrain mesh 7 times PER SHADOW!!!). That is a frame rate killer.

Now, we are in a new league of brilliance!. But I have to be on the ball big time for optimisations; which is good because thats my domain.

Edited 18 hours ago by Pocster

[SimonD]

13 hours ago, Pocster said:

But I have to be on the ball big time for optimisations; which is good because thats my domain.

Keeps on coming back to this doesn't it? Domain knowledge is first key, second is then how to translate that into useful prompts. I was playing with DeepSeek this morning and pretty blown away by how far it's come in a few months. It was fascinating that after inputing some questions, I explained I had domain knowledge and it they said that totally changed the situation and provided a much more complex and deep answer that someone without the knowledge or experience wouldn't understand. It's crazy this. I now want to go explore the smaller, more efficient models that are designed for particular tasks as that really seems like the way to go.

Edited 5 hours ago by SimonD

[Pocster]

13 minutes ago, SimonD said:

Keeps on coming back to this doesn't it? Domain knowledge is first key, second is then how to translate that into useful prompts. I was playing with DeepSeek this morning and pretty blown away by how far it's come in a few months. It was fascinating that after inputing some questions, I explained I had domain knowledge and it they said that totally changed the situation and provided a much more complex and deep answer that someone without the knowledge or experience wouldn't understand. It's crazy this. I now want to go explore the smaller, more efficient models that are designed for particular tasks as that really seems like the way to go.

100% . It does the task in a crap inefficient way but it works . You then say some technical term and instantly it jumps on it and you’re off in a much more technical direction . I designed my project such that 90% just editing 1 file i.e deliberately modular .

If you don’t ( like my terrain , weather and other experiments ) it lashes it up like shite .

I asked it how Minecraft generates its procedural landscapes . It gave a basic overview . I tried to use its understanding of it . Poor results - it then complains it can’t do the tuning and perhaps we leave it for now !!! . I read up for 10 minutes . Then give it a cohesive structured prompt on exactly what we’re doing and the techniques to use and it then stuns me ! .