2026-08-20T13:29:25Z https://nostr.ae Martin Uecker https://nostr.ae/npub1hyr65z796zgs0vc8p69njt3t6vgvc5uvxnpuhkr2wmastxev6pes8me57e https://files.mastodon.social/accounts/avatars/109/246/630/377/160/842/original/c1ed9de162ece70a.jpg https://files.mastodon.social/accounts/avatars/109/246/630/377/160/842/original/c1ed9de162ece70a.jpg https://nostr.ae/nevent1qqs89qd34fy0vjrmu46sdavduhs0ezkegl6gc99gakyxuqdevu00haczyzus02stchgfzpanqu8gkwfw90f3pnzn3s6v8j7cdfm0kpvm9ng8x00dnw3 In reply to <a href='/nevent1qqsxzq4g3hf9cxmepjaqu6rqyq5fp54hsznredq2vvh7u6c4aqv48csdpgam0'>nevent1q…gam0</a><br/>_________________________<br/><br/>Your link seems to say just the opposite. (BTW: I do research on diffusion models.) The idea that diffusion models reproduce training data is based on the (incorrect) idea that they would exactly learn the score, which is generally not true. Although by the same argument one could argue that an LLM trained to predict the next token also learns reproducing training data. 2026-08-15T16:49:39Z https://nostr.ae/nevent1qqswulfuq0wtjjcrq2ute6xhzu4ftdqvgt78laf6ju7x58yec9p4xxgzyzus02stchgfzpanqu8gkwfw90f3pnzn3s6v8j7cdfm0kpvm9ng8xkgk9z2 In reply to <a href='/nevent1qqswv0xy4jfrp2f3a88zssvtcgkm8ls0a26aqv33uwgnct3pscm2c8qw2jzsk'>nevent1q…jzsk</a><br/>_________________________<br/><br/>Unclear to me. But I am still curious why you think diffusion models are different? 2026-08-15T15:31:56Z https://nostr.ae/nevent1qqsxjjvwt4mdcn76xej3wpgyyqwahzs8du6p9pzudkdd9y3j5dvjmpszyzus02stchgfzpanqu8gkwfw90f3pnzn3s6v8j7cdfm0kpvm9ng8x6n630y In reply to <a href='/nevent1qqs8fwwaagp87njnpkq82jmrf7gqr9qpal7ms4zfgcfjzla8q5pfuhcd38ft3'>nevent1q…8ft3</a><br/>_________________________<br/><br/>I don&#39;t think that this &#34;neither intentional nor wanted&#34; by design is similar to a human being forbidden by some policy. I agree though that LLMs do not (usually) simply reproduce their training data and that they are trained on some data does not necessarily mean their output plagiarizes it. This does not imply though that training them on copyrighted works did not infringe on the rights of the original authors. This is far from clear at this point. 2026-08-15T15:15:01Z https://nostr.ae/nevent1qqsgu8wadzykhllqge78nydkduuzfqz725chhlvzkg52gew7w65rfcgzyzus02stchgfzpanqu8gkwfw90f3pnzn3s6v8j7cdfm0kpvm9ng8xm5ggxu In reply to <a href='/nevent1qqsysuqm4ea56237c7y45hdzyr5qk3j2jp9js5xtgg4z5stdfz9glcq8gs03c'>nevent1q…s03c</a><br/>_________________________<br/><br/>Nobody claims that they record all training data verbatim, just that this is possible, which is a known fact. (BTW: Shannon&#39;s theorem would apply to any kind of machine learning model.) 2026-08-15T13:08:36Z https://nostr.ae/nevent1qqsv2hdkm5rs5sq3urnz7cvrhk6ksgu3hmrkwuwkn0mn70kyyd9nddczyzus02stchgfzpanqu8gkwfw90f3pnzn3s6v8j7cdfm0kpvm9ng8x8flc7y In reply to <a href='/nevent1qqsyajqc5qhl3d7g3hxm2wwvww32x28hzt4hcu722m9h70kwhwfe69cuhujp5'>nevent1q…ujp5</a><br/>_________________________<br/><br/>Verbatim memorization is empirically known from LLMs. And why would transformers not be able to do this? 2026-08-15T13:01:28Z https://nostr.ae/nevent1qqsqserfwuwqvalfv7j8p76yd3s75w5dvnwserd96lj8j4nle0u63aszyzus02stchgfzpanqu8gkwfw90f3pnzn3s6v8j7cdfm0kpvm9ng8xfnzumc Debian will require reproducible builds!This is a huge milestone.<br/><a href="https://lists.debian.org/debian-devel-announce/2026/05/msg00001.html">https://lists.debian.org/debian-devel-announce/2026/05/msg00001.html</a><br/>#Debian #reproduciblebuilds 2026-05-10T09:53:16Z https://nostr.ae/nevent1qqs9k4kluycpj7h2tvl2c9durmnrntctg83x2987sn7yej2n00cjdtqzyzus02stchgfzpanqu8gkwfw90f3pnzn3s6v8j7cdfm0kpvm9ng8xnvpuvn In reply to <a href='/nevent1qqsq5cgmculylg25m2gkdt0x957ze5fr4ukswm6v0m5wxn35n8ev6ts28x9yv'>nevent1q…x9yv</a><br/>_________________________<br/><br/>The Chinese room argument is fairly unconvincing though. It would not be the person mindlessly following a rule book who can speak Chinese, but the system consisting of the rule book (software) and the processor (person in the room). Nobody would argue that a CPU / GPU is conscious either, but this does not logically preclude a computer with suitable software from being conscious. (no, I do not believe modern Chat bots are conscious, but they also do not pass a real Turing test) 2026-05-02T13:09:59Z https://nostr.ae/nevent1qqswmefmenyvfn84ceclqxzn4d8hww54qqmmmusf9ttv9l49x4r578szyzus02stchgfzpanqu8gkwfw90f3pnzn3s6v8j7cdfm0kpvm9ng8xupuwlq Sie hätten Hitler stoppen können<br/><a href="https://zeitung.faz.net/fas/politik/2025-03-23/d974bf1f6eef25ff6c6c0e6a23436a89/?GEPC=s10">https://zeitung.faz.net/fas/politik/2025-03-23/d974bf1f6eef25ff6c6c0e6a23436a89/?GEPC=s10</a> 2025-03-23T06:51:22Z https://nostr.ae/nevent1qqspf8hcx0mg8rxvg4fgzgcvxepc0yjqm0t0srqty448z7la8zvlehczyzus02stchgfzpanqu8gkwfw90f3pnzn3s6v8j7cdfm0kpvm9ng8xfh5t9m In reply to <a href='/nevent1qqs922c3xm79xkf0j5lcxa8yvtf8jfjkd2x7du0hdg2wr66ucrru9nshdth73'>nevent1q…th73</a><br/>_________________________<br/><br/>It is because the command line is stable and does not change all the time. GUIs also generally lack consistency and modularity. It could be different in theory. i don&#39;t know why GUIs arre so bad and need to be change all time. 2024-08-11T08:07:40Z