This is actually kinda comforting. If we do need to go back to a simpler time for some technological apocalypse reason, it won't actually be too hard to organise. An ICBMs apogee is easily high enough, even 4500km is possible.
The problem with humans is that we can eat almost everything. And when our agriculture, which is very good at high density food production, fails, we will go back to foraging.
It would be worse than what we do now, to an extent that would be difficult to imagine. 1 billion starving people will eat everything.
In general, yes, but I'm not sure it can be subsumed so easily, as slowing down the initial human taste part of the process, taking stock of every nuance of scope, design and technical debt, in order to better understand the current feature requirements more clearly, takes a sort of holistic wisdom and patience we're not yet training for, LLM/agent rewards are fine tuned to deliver instant results from partially specified problems, obedient to instruction following. We need to balance Goodhart's law with a Socratic method?
Obviously if the training rewards change too you'll be right soon enough, they'll learn it too.
I dream of a parallel timeline when OKLAB became the standard colourspace in 1980! It's a very colourful dream, the lightness stays consistent as the hues change, and colour interpolation just works!
When it comes to colour spaces, there is no “one size fits all”. Perceptual colour spaces are good for, well, perceptual transformations, but many transformations ought to be done in a linear colour space, eg. exposure, color calibration and white balance adjustment, etc. This is how darktable operates (everything before the tonemapper is done in a linear colour space, everything after is done in a perceptual colour space such as JzAzBz or DT UCS 2022 [0], the latter being based on OKLab). Experience has shown that choices made when designing a perceptual colour space can lead to subtle issues [1]. (JzAzBz has its own issues for such transformations, such as not modelling the Helmholtz-Kohlrausch effect [2].)
Oklab/oklch is good for consistent luminance across hue, but the real world gamut is so limited - our display / print tech is always able to make green/yellow far brighter than red/blue, for any given tech - it can be a little frustrating to use.
Humans are too flawed with colour perception. I wish color were 2D, but much like music, it depends on what notes you play.
It’s bad enough that monitors do not produce colour accuracy, or to the full gamut.
Honestly, it’s the one thing Apple got right. Everyone in the creative industry runs Apple, and Apple had a consistency with color. Good luck getting your stuff approved if you produced it on Windows.
A full gamut isn't unreasonable because of that. A full gamut is unreasonable because you'd need a larger number of subpixels to make that work. If you look at a CIE 1931 chromaticity diagram, no matter where you place your 3 points for 3 subpixels you will be unable to cover the full gamut. BT 2020 is basically an attempt at covering as much as possible and you can still see it misses huge amounts of visible colors. A wider gamut at that point requires additional color subpixels so it's no longer a triangle.
And at least with current technology we are not at a point where sacrificing resolution or brightness for more coverage of the green end of the gamut is a wortwhile tradeoff.
We just need to invent red LEDs that can go negative and we'll basically fill out the gamut.
> And at least with current technology we are not at a point where sacrificing resolution or brightness for more coverage of the green end of the gamut is a wortwhile tradeoff.
I disagree there, we have a lot of very bright screens that could take the hit. And it's not just the green zone, replacing blue with azure and violet would allow a big expansion of possible saturated colors.
I wonder if they could benchmax Program Bench, it's one of the few benchmarks that looks almost exactly like a good fine tuning curriculum which might help general coding ability and agential work across the board?
I suppose it could make actually worse at taking instructions and redirection mid-task, but that could be easy to work around too. Just hold back a bit of the spec and drip feed it in later?
I don't mean cheat on it, I just mean optimize for it harder!
For this article, they put their thumb on the scale. It's just nonsense conclusions from bad assumptions.
A 100kW Earth based solar array and a 500kWh grid scale battery that will last ~20 years is already a tiny fraction of the cost of even one Starlink satellite. That's today. Commercial. No assumptions.
It doesn't matter even if it was free to put a satellite in orbit, as it only lasts five years once it's up there.
Their radiator assumptions are also laughable. 13x-28x better than ISS with carbon wizardry entirely on paper. TRL level what exactly?
reply