Make the next few years good ones.

Post #442,866 by malraux 3/31/23 12:24:33 PM 3/31/23 12:24:33 PM Reply	Make the next few years good ones. I go back and forth. There are signs that AGI isn't "attainable". We're running out of training material, being able to produce language tokens doesn't remotely resemble thought or planning, and so on and so forth. However, any sort of AI alignment doesn't matter with the chuckleheads open sourcing everything so that bad state actors, millenialists, jihadists, and other extremist groups won't really need to work all that hard to misalign even a non-AGI AI. Could either be the Great Filter, or a Star Trek post-scarcity society (ha), or hopefully more likely than the worst case at least, the human species muddling through as always and making a hash of things but managing to pull its ass out of the fire after a few big screw-ups. Regards, -scott Welcome to Rivendell, Mr. Anderson.
Post #442,867 by Another Scott 3/31/23 2:46:35 PM 3/31/23 2:46:35 PM Reply	Yeah, it doesn't have to be excellent to scam people. I think DeLong made a good case that it's not actually good enough for expert work (or maybe even for good teaching) - yet. But for the monsters, it only has to be good enough to get people to give them money in excess of the costs. Presumably, most of the cost is in the training, and presumably MS and Google and the rest of the big players are going to keep their server farms of training data locked up. But anyone who wants to can scrape the web and steal everything that isn't locked down, so as storage continues to get cheaper, that will be less of a barrier. A lot of this stuff is still buzz-wordy and click-baity, but there is some useful substance out there. Phys.org: The cosmos would look a lot better if Earth's atmosphere wasn't photo bombing it all the time. Even images obtained by the world's best ground-based telescopes are blurry due to the atmosphere's shifting pockets of air. While seemingly harmless, this blur obscures the shapes of objects in astronomical images, sometimes leading to error-filled physical measurements that are essential for understanding the nature of our universe. Now researchers at Northwestern University and Tsinghua University in Beijing have unveiled a new strategy to fix this issue. The team adapted a well-known computer-vision algorithm used for sharpening photos and, for the first time, applied it to astronomical images from ground-based telescopes. The researchers also trained the artificial intelligence (AI) algorithm on data simulated to match the Vera C. Rubin Observatory's imaging parameters, so, when the observatory opens next year, the tool will be instantly compatible. While astrophysicists already use technologies to remove blur, the adapted AI-driven algorithm works faster and produces more realistic images than current technologies. The resulting images are blur-free and truer to life. They also are beautiful—although that's not the technology's purpose. [...] Cheers, Scott.
Post #442,868 by malraux 3/31/23 3:31:54 PM 3/31/23 3:31:54 PM Reply	Training is cheap now too. Stanford showed that you can cheaply train your own model given access to someone else's via an API. Regards, -scott Welcome to Rivendell, Mr. Anderson.
Post #442,869 by scoenye 3/31/23 7:06:16 PM 3/31/23 7:06:16 PM Reply	Samsung beat them to that punch... https://www.theregister.com/2023/03/13/samsung_fake_moon_pics/ They're trading a quantifiable method for a black box that does "something" to make it look better, but no one is able to explain how or why. I have a hard time calling that science.
Post #442,870 by drook 3/31/23 8:21:24 PM 3/31/23 8:21:24 PM Reply	It's already useful I just listened to Seth Godin's latest podcast. He trained GPT-4 on the entirety of his blog and created a personal chatbot. He asked it several questions - the answers to be in-the-style-of him - and the entire podcast was just him reading the answers. It clearly wasn't the most inspired writing, but little to give away that it wasn't human generated. Then at the end he revealed that he didn't read it, either. The entire episode was narrated by a language model of his voice. I didn't have a clue. It was entirely believable. tl;dr We can already generate reasonable answers based on specific people and have it narrated in their voice. -- Drew
Post #442,871 by malraux 3/31/23 10:36:16 PM 3/31/23 10:36:17 PM Reply	The copied voices are ok But if you know the person's voice well, or if you listen to them side by side, the difference is obvious. I'm not saying that won't fool a lot of people, just that it can be detected at this point. Regards, -scott Welcome to Rivendell, Mr. Anderson.
Post #442,872 by drook 4/1/23 5:29:35 AM 4/1/23 5:29:35 AM Reply	It's already effectively passing the Turing test Tell someone, "Listen closely to this and tell me if it's a person or synthesized," and you'll probably get a bunch who can tell ... as well as a bunch of false positives on the real people. But play it without prompting them to expect it and I think it's very unlikely anyone would notice. At least for non-dramatic reading of straight information. -- Drew
Post #442,873 by malraux 4/1/23 12:24:06 PM 4/1/23 12:24:06 PM Reply	As I said, it will fool a lot of people There are companies like play.ht that can do similar things, and in multiple languages. Regards, -scott Welcome to Rivendell, Mr. Anderson.
Post #442,874 by drook 4/1/23 3:10:16 PM 4/1/23 3:10:17 PM Reply	Not Welsh, I'll bet :-D -- Drew
Post #442,875 by malraux 4/1/23 3:39:24 PM 4/1/23 3:39:24 PM Reply	You'd lose that bet. https://play.ht/text-to-speech-voices/british-welsh/ Regards, -scott Welcome to Rivendell, Mr. Anderson.
Post #442,876 by drook 4/1/23 4:43:11 PM 4/1/23 4:43:11 PM Reply	Well that's incomprehensible But then, it's a different language. -- Drew
Post #442,879 by rcareaga 4/2/23 7:50:33 PM 4/2/23 7:50:33 PM Reply	language tokens &c being able to produce language tokens doesn't remotely resemble thought or planning True enough, but machine sentience, when and if it arrives, won’t necessarily resemble thought or planning as we understand these. I still think that the 2014 film Ex Machina made this point brilliantly: the audience is prepared throughout, along with the protagonist, to see the fembot “Ava” as a human female analogue, and at the end the film makes clear that it—not her—is neither female nor human. Even absent sentience and intentionality as we presently understand these, the (primitive, as measured by the standards we’ll apply before the end of next year, if not sooner) sundry iterations at large are beginning to simulate such elements, and the difference between the consequences of blind and of “directed” AI actions will, I predict, rapidly blur. cordially, and I-told-you-so,
Post #442,881 by malraux 4/2/23 8:32:49 PM 4/2/23 8:32:49 PM Reply	Oh, they don't need thought or planning to be dangerous. The difference between actual thought and planning, and a high fidelity impersonation of actual thought and planning, will be lost on those who are simulated out of existence. "Blindsight" by Peter Watts is a good exploration of such topics. Regards, -scott Welcome to Rivendell, Mr. Anderson.
Post #442,884 by boxley 4/2/23 9:32:42 PM 4/2/23 9:32:42 PM Reply	best description of ITIL I have seen yet The difference between actual thought and planning, and a high fidelity impersonation of actual thought and planning "Science is the belief in the ignorance of the experts" – Richard Feynman
Post #442,885 by malraux 4/3/23 9:26:39 AM 4/3/23 9:26:39 AM Reply	😂 Regards, -scott Welcome to Rivendell, Mr. Anderson.
Post #442,886 by drook 4/3/23 9:53:26 AM 4/3/23 9:53:26 AM Reply	Then ISO 9000 to thoroughly document the impersonation -- Drew

Welcome to IWETHEY!