Yes. A side effect of the expansion of copyright enforcement pushed by larger corporations means that companies generally are walking on eggshells and have streamlined processes to remove content based on a standardized compliant process. Even more so in the last few years with the billion-dollar lawsuit against Cox working its way through the courts.

Solar is at 7%. It's very significant.

That’s my point. Solar is on track to eclipse that amount way before we would build all of this speculative geothermal.

Just make everything part of the mix, even dam to dam hydro on rural hillsides plays a part in stabilising grid edges.

The more interesting use of geothermal power, IMO, is pushing heat underground with excess solar while the sun is high and the days are long, and pulling out electricity from heat in the darker times .. done large enough that carries across summer to winter.

HN is owned by a startup accelerator and venture capital firm. They do growth hacking on the front page. And you probably know that since your throwaway account is several years old.

I found the KLD benchmark image at the bottom of https://unsloth.ai/docs/models/qwen3.6 to be very helpful when choosing a quant.

The previous version was named Amtrak-System-Map-020923.pdf and it's still available at https://www.amtrak.com/content/dam/projects/dotcom/english/p... and it really is about 1MB. Just zooming in, it seems to be a single lossy image.

The current one is called Amtrak-System-Map-081325.pdf and went live on the site sometime between August 19th and 20th (according to Wayback Machine snapshots). It looks like someone uploaded a working copy and forgot to export a nice small version.

It's somewhat akin to how documentation inside code is so often inaccurate; it may have been fine when it was written, but it doesn't get updated.

You didn't really explain what that does for you. Why do you paste it into an LLM?

I'm not sure if you're familiar with public domain texts from around the 19th or early 20th century, but they were not intended to be skimmed or speed-read the way we'd skim a modern text prior to getting into a more attentive close-reading. Even their short magazine articles were actually the near-equivalent to our scholarly papers, and were often read aloud at length in parlor gatherings. So having a LLM split the text into manageable sections for you and provide a hint of what each lengthy wall-of-text paragraph will be about is actually a huge gain in readability.

Oh well that was the whole point to me. If I wanted to read something that's not from 1911 I could just do that lol

The trick is to have a basic level of literacy and then you don't need the machine to chew it up for you like a mother bird.

Mostly from a bit further back but you might enjoy https://earlymoderntexts.com/texts

So before you were talking about summarizing whole articles and asking the LLM to find the things that would be "beyond the pale", but now you're just suggesting using it to insert paragraph breaks and section headings?

The LLM will easily do both for you. Particularly the thinking it does when constructing the summary generally involves a structured close reading of your text, and you can easily think of it as providing "paragraph breaks and section headings".

Sure it could do both, but the question is what are you suggesting?

If you're suggesting it alter the text beyond organizing it, people are going to be upset. And your first suggestion sounded like that.

I think the word "summarization" might be throwing people off. This is like an expansion.

> So having a LLM split the text into manageable sections for you and provide a hint of what each lengthy wall-of-text paragraph will be about is actually a huge gain in readability.

Perhaps your attention span needs improvement.

Would you mind not using the Lena test image? kthx

I bet it would still be interesting if it was the first car that had ever been made reliable enough to finish a half-marathon.

That sounds cool but how do you encode the image data?

The DCT of wavelengths that create the non-spectral colour.

No clue how you'd film in this format but the CGI and video games would be epic.

That's the easy part, just use a color space with imaginary primaries (see e.g. ProPhoto RGB), or use one with real primaries that allows for negative values – e.g. Windows uses floating point scRGB for HDR, which is just linear BT.709/sRGB, but negative RGB values can be used to cover the full range of real and imaginary colors.

But that's still mixing just a few primary wavelengths.

Just save pixel values as wavelength rather than RGB?

Maybe wavelets will make a comeback