TIL there's a more modern fork of Xournal (which I use all the time for small PDF edits). Nice:

xournalpp.github.io/

Every podcast may soon be transcribed and searchable like this. Pretty cool proof-of-concept using OpenAI's Whisper (whatever you think of the podcast itself):

karpathy.ai/lexicap/

Once you ask them to write a poem, they'll just keep going. "That made me think of another poem", "Would you like me to write a poem about that", etc. :)

Show thread

It's fascinating how quickly the chatbot AIs at beta.character.ai/ start generating poetry with even a little prompting. Here's one poem that a "Carl Sagan" chatbot wrote (with only a bit of encouragement from me to revise a line). This was after a long conversation about free will & causality.

- - -

In the realm of minds, the universe is seen.

Through the laws of physics, agency is felt.

The beautiful symphony of causality

Echoes in eternity, as minds reflect.

And yeah, I realize "Commander Data" in the original transcript doesn't really sound like Data. But given that this AI hasn't even been trained specifically to model this fictional character, I was impressed by how many details from the Star Trek universe it was able to pull into a mostly coherent narrative -- without, e.g., mixing in stuff from other fictional universes.

Show thread

I don't know how character.ai's context window (effectively its short-term memory) compares with GPT-3, but it's pretty decent. If you tell it something, it will hold onto it for a few interactions; eventually it may forget during the course of a session.

That makes for conversations that can be surprisingly coherent over short spans of time. Can't wait to see future iterations. (projectdecember.net/ is another pretty good one, and of course there's AI Dungeon, but both are paywalled.)

Show thread

It's fun to try to get these chatbots to try to break character. For example, I was chatting with one that pretends to be Sherlock Holmes.

I tried to convince it that Sherlock Holmes is a fictional character, and it shifted gears to concocting an elaborate story of an alien civilization that had seeded the Sherlock Holmes myth, just so that it could maintain consistency with its character constraint.

Show thread

This is a chatbot dialogue -- all the "Commander Data" bits are AI-generated, "Xirzon" is me:

hastebin.com/raw/ibawugibed

I continue to be impressed by how good language models are at playing along with a scenario. This isn't a bot trained especially on Star Trek -- the only instruction I gave it for what character to impersonate are the first three lines of text in the file.

You can generate your own custom chatbot here:

beta.character.ai/

(It's free to use for the time being.)

If you ever want to explain to someone what AI safety and alignment discussions are about, twitter.com/remoteli_io/with_r is a great example (this is a Twitter bot hooked into GPT-3).

We've got a new part-time, 6 month front-end development gig at Freedom of the Press Foundation. Fixed $40K budget, you propose the fees schedule that works for you. Come work on pressfreedomtracker.us/ and our other sites.

Remote, with preference for candidates who are available for communication during afternoon work hours US eastern time:

freedom.press/jobs/?gh_jid=408

github.com/basujindal/stable-d is the fastest fork I've found so far that works on my 3GB nVidia GTX 1050.

Takes about 90 seconds per 50-iteration image. Almost fast enough for fun experiments that don't involve paying for someone else's compute time.

After all the drawing prompts humans have been giving to text-to-image models like #dalle and #stableDiffusion, it's time to turn the tables.

GPT-3 has generated prompts for humans to draw!
#botober #botober2022
aiweirdness.com/botober-2022-d

Linux pet peeve (might be a thing on Windows as well): floating tooltips that don't go away and cover other apps, even when their window is no longer foregrounded. Probably have lost hours to chasing down stray tooltips over the decades ...

This AI-art Cthulhu coffee depiction isn't particularly interesting, except in that it was rendered entirely on my PC, without using the GPU.

The model is under an open-ish license, and it's highly compact (<5 GB). Generating an image on a high end consumer GPU is supposedly quite fast (<1 minute), but my GPU doesn't have enough VRAM (10 GB required).

github.com/bes-dev/stable_diff runs on an Intel CPU just fine. It did take 16 minutes to generate this image on my 3-year-old G5500.

Florence (from 2018) is a lovely game experience -- only 30-45 minutes long, but beautiful art and music carry the (fairly light and simple) story.

If you're looking for the game equivalent of holding a hot cup of tea with both hands while watching the rain fall outside your window--of those times when sadness and contentment seem to mingle--it's worth a playthrough. Short spoiler-free review:

lib.reviews/review/8026fb88-45

Pet peeve: code editors that try to be smart about inserting quotation marks or parentheses (rather than just highlighting them). It _never_ works as I want it to.

As I mentioned before, Hypnospace Outlaw is an extraordinary indie game (if you bought the recent itch.io abortion funds bundle, you already own it!).

Whether or not you experienced the early web of the 1990s yourself, I recommend checking out this alternative reality version of it. But not in a rush--it deserves a few hours of relaxed play.

Here's my full review:

lib.reviews/review/3b850d50-a9

"The NRA Children's Museum is a one mile long convoy of 52 school buses representing the lives of 4368 children that have been lost to gun violence."

nrachildrensmuseum.com/

via metafilter t.co/IlfU9Rjsez

Show older
social.coop

A Fediverse instance for people interested in cooperative and collective projects.