herf

would love to try it again when you have updates

Unfortunately even with my 5070ti, llama.cpp seems to be about 20-30% faster at decode, running as:

set CUDA_VISIBLE_DEVICES=0 build\bin\Release\llama-server -hf google/gemma-4-12B-it-qat-q4_0-gguf -ngl 99 --no-mmproj-offload -mg 0 -c 262144 -fa on --host 0.0.0.0

If you can rent a dedicated 96-core Epyc for $1/hour (cheap dedicated host), the combination of IPC improvements (>2x) and core count make 7,010 "skylake core-years" cost only $175k, not $5M. On-demand cloud servers (which cost more than $1, maybe $5/hour) probably make the GPU cheaper, but it's closer than the author says.

We know the typical LMS responses for human cones (e.g., Stockman-Sharpe) that could easily be plotted instead of the bell-curve-like things they use here - why? Is it good enough for illustration? You could say, kind of but then all the colors are kind of wrong too. For example, D65 is very blue here.

The article takes a really long time to get to the thing everyone should know: which is that our cones responses overlap a lot, and we get color sensations by (kind of) differencing them. Doing a matrix transform after ten pages to get to chromaticity is technically correct, but it's not the most accessible/intuitive way to teach this.

Nice effort but it would be better if it were closer to "real" color math.

The JS code seems to output linear RGB directly to the webpage (missing gamma). I guess the difficulty with "adding gamma" to it is this: if the training data were also gamma encoded, then it might need to be refit with linear values.

It's hard to remember with all the ownership changes, but the Photobucket era was really a different time, of "it's your data, you're in charge, and we give you maximal control of it" - people would upload there to post elsewhere, and I recall they ran ads to monetize. But this era had the ethic that uploading was expensive, and you'd maybe want to do it once and have control of your stuff after that.

Now we have photo hosting services that barely work on the web (iCloud), or work only within a walled garden (Instagram), and I do miss the "it's your stuff, we're just a website" kind of attitude from the mid-2000s.

RGB values represent luminances against some adapted state, and a "zero" in a daylit scene is not "zero luminance" - it's just about 0.001x as bright as the brightest point - it's millions of photons, way more than zero. In a sense our eyes experience contrast on a sliding scale, and there is no absolute zero in the system. For example, broadcast systems historically used 16-235 as their luminance range for SDR. I think any argument that says "we must have zero" is going to have a bias, but I don't think zero is needed for most things.

https://www.thebignewsletter.com/p/did-a-private-equity-fire...

1. What is the search index?

2. The "description.md" example has things like "faces -> cluster_id". Is this from Davinci Resolve's face index? Things like faces+names and locations are really important with photo collections, but general LLMs don't handle them so well.