I used to work at a company who used to do this, but I'd argue that it almost never does what you want re: concealing the input space. There are so many ways to recover the input space from the latents, and often they are constructed expressly to contain exactly enough information to do precisely that.

But how much effort would that require? ... is what I was going to say but in this age of "claude reverse engineer this doohickey" maybe the better question is how much compute and how many tokens it would require.

I'd be a bit nervous re: the legal implications of using "Pixel" in the name of the project (I thought this was an official Google project).

Companies regularly release their custom-made branding fonts under SIL / open licenses.

Out of curiosity, does anyone know of anything similar that will plug into an HDMI port directly à la Roku? Would be cool to stream my Jellyfin collection on something like this

You could look into Intel Compute Sticks[0]. They're small Bay Trail (quad core Atom) PCs with a male HDMI and a USB port. They will run Linux well enough but it's a bit of a pain to get it installed since they've only for a single USB port. They used to be really cheap so they may still be. IIRC they've got full VP9 and h.264 High Profile hardware acceleration.

[0] https://www.intel.com/content/www/us/en/ark/products/series/...

Why not run the Jellyfin client (or Kodi?) on a Firestick?

I would rather have control of the fullstack. I don't very much trust Amazon/Roku/etc... regarding my privacy.

I think your best option then is CoreELEC running Kodi. Look for a deal on Fire Cube 2nd gen on eBay, root it and install CE. Not as compact as a Firestick and a bit more expensive.

https://discourse.coreelec.org/t/best-device-for-coreelec-in...

My trouble with Pangram is that the calibration is totally off. It can tell you if there is a drop or an ocean of AI in the text you submit. The "X% of AI" message doesn't really mean much. You can test this yourself: Slowly add generated text to a human write paragraph and you'll see it hit "100% AI" quite quickly.

It analyzes in segments, and the percentage is of the segments, not of the words.

Yes but the segment sizes are huge and any amount of LLm generated sequences within that block cause all of it to be flagged as AI.

That's the state of the tech right now. It'll be nice if we can get it and keep it more granular than that, but even at the current state it's pretty impressive.

I think my issue is more about how the tool communicates it's result than with the tool itself. The people I speak with don't understand that this is a binary "YES/NO" for many of the passages they are testing.

Good?

I don't really know why you would expect a "X% chance of containing AI" number to rise slowly or smoothly alongside the percentage of text that has a 100% chance of containing AI.

Because it isn't an "X% chance". It's an "X% of words".

Surely not at the sub-paragragh level? OP said their test passage was a paragraph.

My understanding is that the main focus is on the engine and building a good embedding API so that folks can build browsers (and apps and web drivers etc...) with it.

That is my understanding too. In contrast Ladybird has really improved in the UI area and the recent addition of session restore, content blocking and performance makes it a very capable daily driver browser

Maybe The Gold-Bug by Edgar Allen Poe? Although the language might be a bit baroque.

I very much remember this, from when I was ten or so.

In the same vein, is the "The Adventure of the Dancing Men" [1]

[1] https://en.wikipedia.org/wiki/The_Adventure_of_the_Dancing_M...

this, but for a 10 year old consider getting the Great Illustrated Classics abridged version. i read it when i was 10, and it kinda blew my mind.

Thanks. We just finished Gulliver's Travels with him, so I bet he can handle it.

Would love a 27B model. Wonder if we'll get a distillation.

I'd like to see this benched against oh-my-pi because I've really been digging how it does compaction / snapshotting.

great idea. Currently working on updating benches on frontiers models, so far on glm-5.3 it's already showed great results. Can I have a link to oh-my-pi to make sure will sure use right one for it?

Also, quick question: Could I use octomind upstream of harnesses like omp and get the savings? Or do you primarily get the savings from using the octomind CLI?

Also regular pi, which is particularly minimal

How is oh-my-pi compaction different? Is there other harnesses that employ the same tricks?

did they stick with the image-based compaction (when vision is supported i guess)? It boggled my mind: "snapcompact, which archives older history into dense images for a vision-capable model"

For obvious reasons it underperforms on models that hide their reasoning chain. Other than that it works fine for a non-customizable method. But properly customized handoff is still better.

it's in the docs! never have used it though https://omp.sh/docs/compaction

I use it. It works flawlessly.

I imagine it is bean counters but like VC bean counters... You dump a butt load of cash into games hoping they'll be the next GTA V or Skyrim

I believe that with the budgets but it's crazy we don't see more risks from a creative perspective. In order for the VC model to work, you need to take risks for asymmetric payoffs. In the end, I guess management would rather follow the internal herd and play it safe.

Compare it more to movie production and book publishing. Where we also see a pattern of the big budget stuff being risk averse and the really creative stuff happening downmarket. Often so far downmarket that the creators themselves end up being the primary financiers.