[links] Handcarved Ghibli figures / Sound of a pen plotter / ChatGPT is not a trickster
Also: AI knitting is still cursed, How to turn ChatGPT into an AI pirate, Moebius' La Parapsychologie et Vous, Realtime AI-postmodernism with Kendrick Lamars Voice, Organoid Intelligence and more.
“Studio Ghibli is bringing a new lineup of Totoro collectibles in collaboration with some of the best craftsmen from around Japan. The collection comprises Totoro figures, paper treasure boxes, flower vases, plates, handkerchiefs, and furoshikis (traditional wrapping cloth), teleporting children to the world of Ghibli through role playing goodies.” This stuff is gorgeous, get it here.
Parapsychology by Moebius: “La Parapsychologie et Vous, a book by Paule Salomon and Charlie Cooper, was published in 1980 and illustrated throughout by Moebius. According to a note by Jean-Marc Lofficier, Moebius had been introduced to Salomon by Jean-Paul Appel-Guéry, a French New-Age figure who I think may be the “guru” whose influence over Moebius in the 1980s is referred to rather scathingly by Jodorowsky in the Moebius Redux documentary.”
David Mrugala aka The Dot Is Black visualizes the sound of Axidraw-robots. No prints (yet), but i’d definitely hang one of those on my wall.
Your voice arguably has nothing to do with your face features, except maybe for the lips and overall volume of the jar. Yet, we get AI-models that learn that connection, to produce voices it correlated with face features. This is related to the pseudoscience of Physiognomy, where people believed you could determine personality traits from face features, a prerunner for eugenics, and we should be careful with AI-models aping such pseudoscience.
How to turn ChatGPT into an AI pirate via indirect prompt injection: "LLMs might process poisoned content retrieved from the Web that contains malicious prompts pre-injected and selected by adversaries."
Related paper: [2302.10149] Poisoning Web-Scale Training Datasets is Practical
Researchers are proposing Organoid Intelligence (OI) for “Wetware processors with organic neurons [which] will eventually replace many current AI training chips” (Daniel Jeffries). These OIs already learned to play the game Pong. I'm not really ready to use OI outside of a punk context but, well, here we are. Biopunk it is then.
In a thread a few weeks ago, I claimed that language paradoxa are a good benchmark for understanding in AI. I tried to get ChatGPT to produce a poem that is similar to the famous german paradoxical poem “Dunkel war’s, der Mond schien helle“ and I failed. I wrote that it “looks to me like a machine that is trained on internet-corpi and how words in those relate to each other is unable to really generate stuff the contradicts itself in a playful manner that still makes sense. You can get ChatGPT to be wrong, but you can't get ChatGPT to be wrong in a right way that makes poetic sense. ChatGPT is no trickster. In all seriousness, i think 'AI being a meaningful trickster' is a good benchmark, albeit maybe not one that we really want. Playful deception is a clear sign of intelligence. After all, dogs can absolutely do it.”
Now there’s a paper about pretty much exactly that: LLMs fail on the fringes of language: “We tested their ability to process grammatical illusions, semantic anomalies & complex nested hierarchies. Short answer: Even the much praised next-word prediction abilities of LLMs face issues of robustness, when pushed beyond training data.“ Here’s the Twitter-thread by Evelina Leivada.ChatGPT Is Nothing Like a Human, Says Linguist Emily Bender: ”We go around assuming ours is a world in which speakers — people, creators of products, the products themselves — mean to say what they say and expect to live with the implications of their words. This is what philosopher of mind Daniel Dennett calls ‘the intentional stance.’ But we’ve altered the world. We’ve learned to make ‘machines that can mindlessly generate text’, Bender told me when we met this winter. ‘But we haven’t learned how to stop imagining the mind behind it.’”
Youtube is building generative AI-tools into their video platform: YouTube’s new leader teases AI tools that can virtually swap outfits. I guess we’ll see tools for generative Vtubers soon on the platform too, which don’t just swap outfits, but whole virtual identities.
Hacker FM is “a fully AI generated podcast, released daily, featuring the latest hacker news”.
“Digital technology turns the world into editable grey goo ready to be remixed and tranformed into everything. This is realtime-postmodernism in Kendrick Lamars Voice.”
AI Tool Reveals How Celebrities' Faces Have Been Photoshopped
Artificial Intelligence generates new music to continue Nirvana's "On A Plain". Compare this to the quality from a few months ago. Deteriorates much less into noise now, still way to go for infinite AI-Nirvana.
In other news: AI knitting is still cursed: Crochet enthusiasts asked ChatGPT for patterns. The results are ‘cursed’. Former curses in AI knitting: SkyKnit: When knitters teamed up with a neural network and Yun Ying Hs AI Knitwear. Yes, AI-knitting.
New in mysterious sicknesses spreading via social contagion among diplomats: ‘Havana syndrome’ not caused by energy weapon or foreign adversary, intelligence review finds
Dance like nobody’s watching: Unique Identification of 50,000+ Virtual Reality Users from Head & Hand Motion Data
A new paper looks at Profiles of animal consciousness: A species-sensitive, two-tier account to quality and distribution, press release in german: Wie sich Bewusstsein bei Tieren erforschen lassen könnte. The researchers propose ten dimensions or aspects of consciousness, which maps pretty good with my toy-theory of a moiré-consciousness, which simply states that consciousness is the ability to compose complex interferrence-patterns from simpler patterns in brain activity with the goal to tell an attentionally guided story of ourselves, where this story is told in finegrained micro-behaviors like “looking at something“. Consciousness is a continuum, exactly like the complexity of Moiré interferrence patterns increases with the number of patterned layers. The more aspects of consciousness a living organism (and otherwise?) can develop, the more complex its conscious experience of the world.
Here’s the latest brain-2-image research: High-resolution image reconstruction with latent diffusion models from human brain activity. This technology might give us digital lucid dreaming at one point.
Do Trees Really Communicate Through Fungi? It May Be a Myth made up by the media for clicks: “I would say in very plain words: The [media] story is ahead of the science. That’s, I think, the basic message.“
After the blackest black and the pinkest pink, here’s the whitest white: Just How White Can White Paint Get: “A team from Purdue University is using nanotechnology to create paints that reflect 98% of sunlight.“
Make Joe Biden cool again: