3 ms·

Everything is a Remix (Complete Updated 2023 Edition) [video]

4 points by zdw 4y ago

dredmorbius 4y ago
Long-form video (I'm ~15 minutes into an hour-plus piece), but good, and I can think of a few directions it might take given contemporary trends <coff>generative AI</coff>....[1] Another recent piece looking at the copyright angle of sampling ran recently at On the Media: "Silenced Samples: How Copyright Laws Infringe on Hip Hop" <https://www.wnycstudios.org/podcasts/otm/episodes/hip-hop-sampling-on-the-media https://www.wnycstudios.org/podcasts/otm/episodes/hip-hop-sa...> (Transcript available after clicking through.) Media: <https://chrt.fm/track/53A61E/pdst.fm/e/dts.podtrac.com/pts/redirect.mp3/waaa.wnyc.org/otm/otm031523_cms1309020_pod.mp3 https://chrt.fm/track/53A61E/pdst.fm/e/dts.podtrac.com/pts/r...> ________________________________ Notes: 1. Spoiler: It does.
dredmorbius 4y ago
Having watched The Whole Damned Thing ... thoughts: 1. It's good, and worth viewing. Do yourself the favour of doing so, especially before commenting. Or reading others comments. Including my own ;-) 2. There's some good background and history. The WNYC bit addressed in my own earlier comment complements nicely: <https://news.ycombinator.com/item?id=35334999 https://news.ycombinator.com/item?id=35334999> <https://www.wnycstudios.org/podcasts/otm/episodes/hip-hop-sampling-on-the-media https://www.wnycstudios.org/podcasts/otm/episodes/hip-hop-sa...> 3. And of course I disagree with bits. In the conclusion, Ferguson asserts that "AI will not be dominating innovation, because AI does not innovate. They synthesize what we already know. AI is derivative by design and inventive by chance." There's some truth to this ... but the statement can apply as well to human innovation, and there are certainly human innovators who've created with no idea at all about the ultimate applications of what they've created. To hold that against AI, either as it presently exists (e.g., large language model (LLM) machine learning adaptive diffusion generative transformers), or as it may exist in a plausibly-conceivable future ... seems neither especially relevant (innovation can occur unconsciously) nor given (future AIs may well have a sense of goals, constrains, and what for want of a better term I'll call "structure" or "skeleton" of that which they're creating). My last point seems to be a major component of the argument made in yesterday's HN post, "Do large language models need sensory grounding for meaning and understanding?": <https://drive.google.com/file/d/1BU5bV3X5w65DwSMapKcsr0ZvrMRU_Nbi/view https://drive.google.com/file/d/1BU5bV3X5w65DwSMapKcsr0ZvrMR...> <https://news.ycombinator.com/item?id=35321054 https://news.ycombinator.com/item?id=35321054>. What predictive AI algorithms are doing seems to be more designating contours of concepts rather than developing them from their internal structures. The best artists I'm aware of, especially portrait artists, and in many cases CGI animators, develop images from the bones out. For nonhuman structures, this might be other principle structural elements --- the main trunk of a tree or plant, the keel of a ship or aircraft, major framing elements of buildings or structures, etc. Having a sense of what's going on beneath the level of what's immediately visible makes what is portrayed far more realistic. For animation, this goes further into physics packages and the like. (I'm writing as a viewer and audience of such works, not as a creator, though I've spent some time with at least some creators and seen some variants of the process, as well as presentations and documentation of how the magic is done.) Present-generation AI still doesn't seem to be doing this, but the leaps seem relatively small. I'm writing this as someone who has been and tends to be reasonably conservative in predicting technological progress. In particular I'm reminded of a talk by the late John Holland (University of Michigan and Santa Fe Institute) on the nature of creativity and innovation, with a set of simple mechanisms: duplication, transformation, and deletion. Add to this a fourth process, selection, and you have the heart of all evolutionary systems, natural or artificial. In the case of AI, addition of further random noise and selection might be sufficient to generate what we'd consider truly creative capabilities. Unfortunately I cannot find the talk, though I remember it being on YouTube, and believe I'd stumbled across it in 2015. He addresses some related concepts here: "Workshop: Complexity Methods - John Holland" <https://yewtu.be/watch?v=3httLyRCCuM https://yewtu.be/watch?v=3httLyRCCuM> 4. A whole area that's occurring to me as possibly ... interesting ... regarding LLM AI is that our present intellectual property models, most especially copyright, seem ill-prepared to handle it. The video addresses the issue on the training side, but the output side is equally fraught. Copyright applies to "original works of authorship", where authorship requires a human author. AI outputs are not human creations, and hence are not eligible for copyright. One possibility of course is that a few trillion-dollar concerns will lobby for either extensions of existing or creations of new forms of IP law which will address these works. But that too would be a major undertaking. For myself, I'd prefer seeing copyright walked back considerably. I'm increasingly doubtful that will happen.

4 points by zdw 4y ago

Long-form video (I'm ~15 minutes into an hour-plus piece), but good, and I can think of a few directions it might take given contemporary trends <coff>generative AI</coff>....[1] Another recent piece looking at the copyright angle of sampling ran recently at On the Media: "Silenced Samples: How Copyright Laws Infringe on Hip Hop" <https://www.wnycstudios.org/podcasts/otm/episodes/hip-hop-sampling-on-the-media https://www.wnycstudios.org/podcasts/otm/episodes/hip-hop-sa...> (Transcript available after clicking through.) Media: <https://chrt.fm/track/53A61E/pdst.fm/e/dts.podtrac.com/pts/redirect.mp3/waaa.wnyc.org/otm/otm031523_cms1309020_pod.mp3 https://chrt.fm/track/53A61E/pdst.fm/e/dts.podtrac.com/pts/r...> ________________________________ Notes: 1. Spoiler: It does.

dredmorbius 4y ago

Having watched The Whole Damned Thing ... thoughts: 1. It's good, and worth viewing. Do yourself the favour of doing so, especially before commenting. Or reading others comments. Including my own ;-) 2. There's some good background and history. The WNYC bit addressed in my own earlier comment complements nicely: <https://news.ycombinator.com/item?id=35334999 https://news.ycombinator.com/item?id=35334999> <https://www.wnycstudios.org/podcasts/otm/episodes/hip-hop-sampling-on-the-media https://www.wnycstudios.org/podcasts/otm/episodes/hip-hop-sa...> 3. And of course I disagree with bits. In the conclusion, Ferguson asserts that "AI will not be dominating innovation, because AI does not innovate. They synthesize what we already know. AI is derivative by design and inventive by chance." There's some truth to this ... but the statement can apply as well to human innovation, and there are certainly human innovators who've created with no idea at all about the ultimate applications of what they've created. To hold that against AI, either as it presently exists (e.g., large language model (LLM) machine learning adaptive diffusion generative transformers), or as it may exist in a plausibly-conceivable future ... seems neither especially relevant (innovation can occur unconsciously) nor given (future AIs may well have a sense of goals, constrains, and what for want of a better term I'll call "structure" or "skeleton" of that which they're creating). My last point seems to be a major component of the argument made in yesterday's HN post, "Do large language models need sensory grounding for meaning and understanding?": <https://drive.google.com/file/d/1BU5bV3X5w65DwSMapKcsr0ZvrMRU_Nbi/view https://drive.google.com/file/d/1BU5bV3X5w65DwSMapKcsr0ZvrMR...> <https://news.ycombinator.com/item?id=35321054 https://news.ycombinator.com/item?id=35321054>. What predictive AI algorithms are doing seems to be more designating contours of concepts rather than developing them from their internal structures. The best artists I'm aware of, especially portrait artists, and in many cases CGI animators, develop images from the bones out. For nonhuman structures, this might be other principle structural elements --- the main trunk of a tree or plant, the keel of a ship or aircraft, major framing elements of buildings or structures, etc. Having a sense of what's going on beneath the level of what's immediately visible makes what is portrayed far more realistic. For animation, this goes further into physics packages and the like. (I'm writing as a viewer and audience of such works, not as a creator, though I've spent some time with at least some creators and seen some variants of the process, as well as presentations and documentation of how the magic is done.) Present-generation AI still doesn't seem to be doing this, but the leaps seem relatively small. I'm writing this as someone who has been and tends to be reasonably conservative in predicting technological progress. In particular I'm reminded of a talk by the late John Holland (University of Michigan and Santa Fe Institute) on the nature of creativity and innovation, with a set of simple mechanisms: duplication, transformation, and deletion. Add to this a fourth process, selection, and you have the heart of all evolutionary systems, natural or artificial. In the case of AI, addition of further random noise and selection might be sufficient to generate what we'd consider truly creative capabilities. Unfortunately I cannot find the talk, though I remember it being on YouTube, and believe I'd stumbled across it in 2015. He addresses some related concepts here: "Workshop: Complexity Methods - John Holland" <https://yewtu.be/watch?v=3httLyRCCuM https://yewtu.be/watch?v=3httLyRCCuM> 4. A whole area that's occurring to me as possibly ... interesting ... regarding LLM AI is that our present intellectual property models, most especially copyright, seem ill-prepared to handle it. The video addresses the issue on the training side, but the output side is equally fraught. Copyright applies to "original works of authorship", where authorship requires a human author. AI outputs are not human creations, and hence are not eligible for copyright. One possibility of course is that a few trillion-dollar concerns will lobby for either extensions of existing or creations of new forms of IP law which will address these works. But that too would be a major undertaking. For myself, I'd prefer seeing copyright walked back considerably. I'm increasingly doubtful that will happen.