Show more

GitHub stars are to the upcoming dot ai bubble what clicks were to the dot com bubble

Hoo boy, another discussion about how mailing lists are terrible and therefore must be replaced by a single point of failure system with non-exportable archives.

Please don't clog up the Transgender Concerns Form with so many spurious reports that it becomes unusable, and the Missouri government gets Big Mad. Because that would be wrong. Here's a link to the form so you'll know exactly where not to do that: ago.mo.gov/file-a-complaint/tr

Show thread

Curious to know what are people intuition about it without getting into a computer science argument

Show thread

In your view, are a Twitter handle and their bird.makeup equivalent:

I would be nice if Mastodon admin could make an allow-list of accounts for other servers, this would allow them to only allow accounts of public interest for their community from bird.makeup and remove noise from the rest.

That feature might be too niche though.

yeah gonna need to suspend bird dot makeup, and go through and suspend all the twitter mirrors.

gross af.

1/ A rare disagree w/ the great Matt Levine. Crypto is not illegal like meth is. Crypto is under the same regime like TradFi, according to the SEC itself.

It would be deeply odd if the Chair of the SEC had never bought a single share of stock nor invested in a mutual fund.

@video_manager @dlakelan @marklemley

Stable Diffusion (as used by StabilityAI) has been trained on billions of images, each of which typically requires over a million bits ( 3×8×512×512 > 10⁶ ) to "store" in a way where you can recreate the image.

So if Stable Diffusion were copying the images, it would need at least a quadrillion bits for reasonable quality. (10⁹×10⁶=10¹⁵)

Instead, Stable Diffusion is conditioning less than one billion parameters, so the whole model needs less than 64 billion bits. (6.4×10¹⁰)

That's a ratio of less than 0.000064 model bits stored per input bit. Is it likely that it is storing 0.0064% of every image? No. It is conditioning the model to distinguish between "art" and "noise" and to be able to make input noise look more like the concepts of art that it has been trained on. One can imagine all possible 512×512 images lying in a 786432-dimensional space with the "art" images clustered together in a hard-to-imagine shape. The 890 million parameters of the model are about this shape and how parts of it are associated with certain keywords, not about recreating any particular image.

I'd really love to find a way to grow lemmy.ml/ using bird.makeup. I wonder if there is some types of data that I could forward that would help

Futurama's most lasting contribution to our culture? I think maybe "Technically correct — the best kind of correct!"

Important lesson for social movements:
You don't need to use the brick, but you must have one

youtube.com/watch?v=bmSSAu20f0

As a Canadian, this is half funny, and half worrying that if a fascist gets to power in the US, they’ll come to “liberate” us from our democratically elected representative

Seems like universeodon.com just banned bird.makeup...

Every once in a while a top-notch observation disguised as a meme comes around, and I think this is one.

I think we tend to overestimate the importance of software design as the main factor in explaining why social media feels “good” to us while underestimating the importance of good people posting good things. This comes to mind as folks wonder if Bluesky or Substack Notes or whatever new platform might be more “user friendly” than Mastodon and then therefore definitely “better.”

Show more
Librem Social

Librem Social is an opt-in public network. Messages are shared under Creative Commons BY-SA 4.0 license terms. Policy.

Stay safe. Please abide by our code of conduct.

(Source code)

image/svg+xml Librem Chat image/svg+xml