Apropos of...
Now and then we all think of something relevant to something, but there is no channel for it. Apropos just means - "with reference to", including "apropos of nothing" when you just have a good thought you want to share. This is that place! Enjoy.
Pete Horne pfp

@horneps

My experience of A.I. amnesia. Around and round we went in circles, until the sky it fell. Crashing down like waves that pounded on a beach somewhere in hell. Jimmy Barnes.
0 reply
1 recast
6 reactions

Pete Horne pfp

@horneps

I have been doing the hard graft to get PQTR a real app and it's nearly there. Meanwhile I have been dreaming of my first Guggenheim installation.
0 reply
0 recast
9 reactions

Pete Horne pfp

@horneps

Awww. Thanks Claude!
1 reply
0 recast
19 reactions

Pete Horne pfp

@horneps

Consolidation of approach: 47D hypercube [0,1]^47 (dials) Holistic loss (spectral + per-hue) ACEO: eigenspace from prior Σ SPSA: gradient polish after ACEO Now into testing and refinement. Getting it done.
0 reply
1 recast
4 reactions

Pete Horne pfp

@horneps

Lol - a small squeak. I am sooooo bored about thinking about polynomials and linear trade offs. Let's get this thing done!
1 reply
3 recasts
14 reactions

Pete Horne pfp

@horneps

Well, what a journey. I had to go back to the human process of "vibing" a camera RAW to get the process right... deep into the refactor. Let's see!
2 replies
0 recast
47 reactions

Pete Horne pfp

@horneps

The most common question I get is "what are you doing?" Ok - it's really "simple". Pro DLSR Camera -> F1 pit lane shot -> Apply Art directors style instantly on phone -> send to social. Problem is - the Art Director wants studio quality shots instantly - their vibe. Their vibe means high fidelity raw + their lightroom style for that image type. So, I am working out how to optimally capture their vibe and transfer it to any RAW - instantly. The GeoS optimiser is the invention that knows how to capture the vibe and make it generalisable to any photo. And I am under strict instructions - no AI f around with my styles because AI gives you the average not my best style. They want (and deserve) the optimal. See - Easy!
0 reply
0 recast
12 reactions

Pete Horne pfp

@horneps

This is my favourite image class right now as I work on the GeoS optimiser. These are diff images of the optimal vs the reference. I am actually colour blind - so this helps me respond to differences mathematically (diff is really spectral diff numbers). The closer the images get to each other - it becomes obvious to me and the optimiser. So me and math are a better team than me and visual checks (BTW tone deaf sound engineers are recongnised as the best as they don't add their own preferred vibe - true!)
4 replies
4 recasts
43 reactions

Pete Horne pfp

@horneps

The tide rises. This one of my apps - it’s like Lightroom (actually more precise as it’s the dials for Hollywood standard vfx editing) and the image is produced by my own raw processing code. Finishing the optimiser now I invented called GeoS that matches vibes from any reference photo into the settings. More on that soon.
1 reply
0 recast
11 reactions

Pete Horne pfp

@horneps

I will be finally, after a year, surfacing my app soon. What I’ve done was ambitious - I wanted to make my own AI that uses other AIs strengths to do coding - that’s actually easy in the base weights model - but my goal is to handle the complexity of massive coding projects. My goal was on the f1 grid immediate image production - so my hard target was do Lightroom/darktable on the phone GPU. And I’m just about there on both.
2 replies
1 recast
7 reactions

Pete Horne pfp

@horneps

There won’t ever be one llm to rule them all. It’s a fools errand as there is more than one way to direct an llm to make the base model weights now - by llms. Also context is everything and so there is as many context goals as there are vertical ideas off weights models.
1 reply
0 recast
9 reactions

Pete Horne pfp

@horneps

Duelling AIs. Claude doing the work on the right, Gemini doing review and feedback on the left. 10x on 10x.
3 replies
0 recast
17 reactions

Pete Horne pfp

@horneps

I love how science creates its own art. This is my i9 + GPU having a good old think about the image it is processing.
1 reply
0 recast
13 reactions

Pete Horne pfp

@horneps

About... I don't know how many months of research completed and in the coding pipeline.
0 reply
3 recasts
20 reactions

makemake  pfp

@makemake

people turn off their brains when they scroll through the timeline so u can often abuse that to ur advantage because theyll take whatever you serve them at face value. you just have to package it nicely
3 replies
4 recasts
21 reactions