Tilman Bayer

@tilmanbayer.bsky.social

267 Followers 292 Following

AI, data, Wikipedia, co-maintainer of @wikiresearch.bsky.social

Source: https://cdn.heyzine.com/flip-book/pdf/03174239fe1c154539a650558379845a0545b9f7-2.pdf p.11
Excerpt from https://en.wikipedia.org/wiki/Jacobian_conjecture
Excerpt from https://lithub.com/the-sound-of-imminence-ruth-ozeki-in-praise-of-the-typewriter/ , highlighted
Screenshot of https://www.quora.com/Why-do-all-the-starving-artists-in-movies-seem-to-be-able-to-afford-huge-open-floor-studios-in-NY-of-all-places-Also-whats-up-with-the-freight-elevators-they-always-have?top_ans=79728669
"[...] Access to a paid subscription of ChatGPT (powered initially by GPT-4 and subsequently by other models) was provided to all members in the AI-assisted and AI-led teams. While ChatGPT had six different versions available between February 14th 2024 (training for our pilot) and our final event on November 22nd 2024, researchers had access to the main flagship models (GPT-4, and/or GPT-4o). These models were capable of processing files, equipped with a Python environment for interpreting code and conducting data analysis, and had internet access. Additional information on the" (excerpt from https://doi.org/10.1073/pnas.2524747123 )
"8. PROHIBITED ACTIVITIES You may not access or use the Services for any purpose other than that for which we make the Services available. The Services may not be used in connection with any commercial endeavors except those that are specifically endorsed or approved by us. As a user of the Services, you agree not to: Systematically retrieve data or other content from the Services to create or compile, directly or indirectly, a collection, compilation, database, or directory without written permission from us. Trick, defraud, or mislead us and other users, especially in any attempt to learn sensitive account information such as user passwords. Circumvent, disable, or otherwise interfere with security-related features of the Services, including features that prevent or restrict the use or copying of any Content or enforce limitations on the use of the Services and/or the Content contained therein. Disparage, tarnish, or otherwise harm, in our opinion, us and/or the Services. Use any information obtained from the Services in order to harass, abuse, or" (screenshot of https://indicator.media/terms-of-service#prohibited )
"Your use of our Services Subject to your compliance with these Legal Terms, including the "PROHIBITED ACTIVITIES" section below, we grant you a non-exclusive, non-transferable, revocable license to: access the Services; and download or print a copy of any portion of the Content to which you have properly gained access, solely for your personal, non-commercial use or internal business purpose. Except as set out in this section or elsewhere in our Legal Terms, no part of the Services and no Content or Marks may be copied, reproduced, aggregated, republished, uploaded, posted, publicly displayed, encoded, translated, transmitted, distributed, sold, licensed, or otherwise exploited for any commercial purpose whatsoever, without our express prior written permission. If you wish to make any use of the Services, Content, or Marks other than as set out in this section or elsewhere in our Legal Terms, please address your request to: subscriptions@indicator.media. If we ever grant you the permission to post, reproduce, or publicly display any part of our Services or Content, you must identify us as the owners or licensors of the Services, Content, or Marks and ensure that any copyright or proprietary notice appears or is visible on posting, reproducing, or displaying our Content. We reserve all rights not expressly granted to you in and to the Services, Content, and Marks. Any breach of these Intellectual Property Rights will constitute a material breach of our Legal Terms and your right to use our Services will terminate immediately." (screenshot of https://indicator.media/terms-of-service#ip , last two sentences highlighted)
Screenshot of https://indicator.media/upgrade
From https://ukdefencejournal.org.uk/is-pre-nuclear-steel-behind-wwii-ship-targeting/
"a 1937 bus schedule to Jones Beach" (https:__www.washingtonpost.com/politics/2021/11/10/robert-moses-saga-racist-parkway-bridges/ )
https://www.deutschestextarchiv.de/book/view/marx_kapital01_1867/?p=439
(excerpt from https://www.authorsalliance.org/2025/06/24/anthropic-wins-on-fair-use-for-training-its-llms-loses-on-building-a-central-library-of-pirated-books/ )
von https://x.com/edleeprof/status/1938267202190532869
"Fig. 2: The multistage pipeline of DeepSeek-R1. From: DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning" https://www.nature.com/articles/s41586-025-09422-z/figures/2
"Extended Data Fig. 1 | Evolution of reasoning-related linguistic features in model outputs across training steps. a, Frequency of representative reflective terms in model-generated outputs throughout the training process. Reflective terms—including ‘wait’, ‘mistake’, ‘however’, ‘but’, ‘retry’, ‘error’, ‘verify’, ‘wrong’, ‘evaluate’ and ‘check’—were identified and curated by a panel of three human experts. Each expert independently proposed a set of words indicative of reflective reasoning, which were subsequently consolidated through consensus into a final vocabulary list. b, Frequency of the term ‘wait’ in model outputs over the course of training. This term was virtually absent during the initial training stages, appeared sporadically between steps 4,000 and 7,000 and exhibited a marked increase in frequency after step 8,000. These trends suggest the emergence of temporal reasoning or self-monitoring behaviour as training progresses."
Excerpt from https://www.astralcodexten.com/p/webmd-and-the-tragedy-of-legible : "WebMD is the Internet's most important source of medical information. It's also surprisingly useless. Its most famous problem is that whatever your symptoms, it'll tell you that you have cancer. But the closer you look, the more problems you notice. Consider drug side effects. Here's WebMD's list of side effects for a certain drug, let's call it Drug 1: 'Upset stomach and heartburn may occur. If either of these effects persist or worsen, tell your doctor or pharmacist promptly. If your doctor has directed you to use this medication, remember that he or she has judged that the benefit to you is greater than the risk of side effects. Many people using this medication do not have serious side effects. Tell your doctor right away if you have any serious side effects, including: easy bruising/bleeding, difficulty hearing, ringing in the ears, signs of kidney problems (such as change in the amount of urine), persistent or severe nausea/vomiting, unexplained tiredness, dizziness, dark urine, yellowing eyes/skin. This drug may rarely cause serious bleeding from the stomach/intestine or other areas of the body. If you notice any of the following very serious side effects, get medical help right away: black/tarry stools, persistent or severe stomach/abdominal pain, vomit that looks like coffee grounds, trouble speaking, [...]' And here's their list of side effects for let's call it Drug 2: 'Nausea, loss of appetite, or stomach/abdominal pain may occur. If any of these effects persist or worsen, tell your doctor or pharmacist promptly. Remember that your doctor has prescribed this medication because he or she has judged that the benefit to you is greater than the risk of side effects. [...]' Drug 1 is aspirin. Drug 2 is warfarin, which causes 40,000 ER visits a year and is widely considered one of the most dangerous drugs in common use. [...]"
"In software, Alexander is regarded as the father of the pattern language movement. According to creator Ward Cunningham, the first wiki—the technology behind Wikipedia—led directly from Alexander's work.[7][8][9] Alexander's work has also influenced the development of agile software development.[9]" (from https://en.wikipedia.org/wiki/Christopher_Alexander )
ChatGPT 5.1. Thinking screenshot: "3. So… is that part of Facebook “not an online platform”? Here’s the subtlety: To decide whether Facebook as a service is an “online platform” / VLOP, regulators don’t care if every post is public; they care that the service includes hosting that disseminates information to the public (news feed with public posts, pages, public groups, etc.). That’s plainly true, and the Commission has designated Facebook as a VLOP on that basis. Wikipedia Recital 15 then says: if some functions fall within certain sections of the DSA and others don’t, those sections only apply to the in-scope functions. Digital Services Act So in my reading: Facebook the provider is unambiguously an online platform / VLOP. Certain functionalities (Messenger, maybe tightly-restricted “friends-only” contexts) are not “dissemination to the public”, but: Facebook still remains a hosting service for that content. And Facebook still remains an online platform provider, so most of the “platform”-level duties attach to it as a whole – they’re not narrowly drafted to cover only public posts. Think of it as: “Is the provider an online platform?” – yes, because some of what it does is public. “Is this specific message public dissemination?” – maybe not, but that doesn’t magically turn off all the DSA obligations the moment you flip your audience to “friends”. 4. Do “friends-only” users fall outside DSA protection?"
Cropped from https://commons.m.wikimedia.org/wiki/File:Paul_Klee_~_Angelus_Novus_~_1920.jpg
screenshot of https://www.phylliida.dev/modelwelfare/wildchat2/#ZjEsZjEuMMUFyAc4xw4uOccQyRkuMC40LGZhIVRhc2s%3D , showing clusters like "Generate explicit fantasy content with supernatural characters and sexualized power dynamics"
"I thought the selfies were deleted? This data was originally archived in compliance with law enforcement requirements related to cyber-bullying prevention. Photos cannot be linked to specific users within the app." (From https://www.teaforwomen.com/cyberincident )
"“The Burda legacy is well known in the technology industry, from the Digital Life Design (DLD) conference in Munich to Hubert Burda’s breathtaking recognition of the promise of the internet in the early 1990s, many years before such views were commonplace. I’m deeply honored to take on this role to help guide Burda during this exciting and consequential time.” Meredith Whittaker, President of Signal" (excerpt from https://www.burda.com/en/news/meredith-whittaker-joins-burdas-board-directors/ )
excerpt from "Terms of Use for Early-Access" at https://huggingface.co/datasets/institutional/institutional-books-1.0 : "Terms of Use for Early-Access This dataset is an Early-Access release shared by the Institutional Data Initiative for research and public-interest use (the “Service”). These terms are intended to support experimentation while encouraging collaboration and feedback as we refine the dataset and work with contributing institutions to define shared, long-term norms for open data reuse. To share questions or feedback, contact us at contact@institutionaldatainitiative.org. By accessing or downloading the dataset or otherwise using the Service, you agree to the following: Noncommercial Use Only You may use the Service solely for noncommercial purposes. Open-source projects and other public-use efforts are welcome, even if they may indirectly support commercial use, so long as they are unaffiliated with commercial actors or intent. If you are affiliated with a commercial organization or plan to use the Service for commercial purposes (including AI model training), you will contact us first at contact@institutionaldatainitiative.org. No Redistribution You may not share or redistribute the Service or any of the data provided through the Service, in whole or in part, including through public repositories or aggregators. If you want others to access it, please direct them to the attribution link. Derivative Works You may create derivative works for noncommercial use, but you may not make available any such derivative works that substantially reproduce the original dataset. Only outputs that are significantly transformed and cannot substitute for the original—such as evaluations, summary statistics, or visualizations—may be shared, with attribution. Attribution If you use the dataset in public-facing work, you must include attribution substantially similar to: [...]"
"which cast doubt on fair use applying to copyrighted works in generative AI. AI companies and their investors, meanwhile, have long argued that a better way is not feasible. In April 2023, Sy Damle, a lawyer representing the venture capital firm Andreessen Horowitz, told the U.S. Copyright Office: “The only practical way for these tools to exist is if they can be trained on massive amounts of data without having to license that data.” Later that year, in comments to the U.K. government, OpenAI said, “[I]t would be impossible to train today’s leading AI models without using copyrighted materials.” And in January 2024, Anthropic’s expert witness in a copyright trial asserted that “the hypothetical competitive market for licenses covering data to train cutting-edge LLMs would be impracticable,” court documents show. While AI policy papers often discuss the need for more open data and experts argue about whether large language models should be trained on licensed data from publishers, there’s little effort to put theory into action, the paper’s co-author, Aviya Skowron, head of policy at the nonprofit research institute Eleuther AI, told The Post. “I would also like those people to get curious about what this task actually entails,” Skowron said." (excerpt from https://www.washingtonpost.com/politics/2025/06/05/tech-brief-ai-copyright-report/ )
excerpt from p.2 of the paper, highlighting the sentence "We submit that a natural first step toward resolving this tension is to ask: Is it possible to train performant language models using only public domain and openly licensed text?"
"Abstract Large language models (LLMs) are typically trained on enormous quantities of unlicensed text, a practice that has led to scrutiny due to possible intellectual property infringement and ethical concerns. Training LLMs on openly licensed text presents a first step towards addressing these issues, but prior data collection"
excerpt from "Why You Should Fear “Slaughterbots”—A Response" https://spectrum.ieee.org/why-you-should-fear-slaughterbots-a-response
excerpt from Atkins and Seamone (2023), Remote Combat Exposure and Moral Injury from Drone Operations, pp.220-221
excerpt from from Atkins and Seamone (2023), Remote Combat Exposure and Moral Injury from Drone Operations, p. 217
"because it requires dedicated engineers to maintain and improve systems to handle the massive automated traffic. And AI companies that do not attribute their responses or otherwise provide any pointers back to Wikipedia prevent users from knowing where that material came from, and do not encourage those users to go visit Wikipedia, where they might then sign up as an editor, or donate after seeing a request for support. (This is most AI companies, by the way. Many AI “visionaries” seem perfectly content to promise that artificial superintelligence is just around the corner, but claim that attribution is somehow a permanently unsolvable problem.)" [excerpt from https://www.citationneeded.news/free-and-open-access-in-the-age-of-generative-ai/ ]
"Training Greenhouse Gas Emissions Estimated total location-based greenhouse gas emissions were 11,390 tons CO2eq for training. Since 2020, Meta has maintained net zero greenhouse gas emissions in its global operations and matched 100% of its electricity use with renewable energy, therefore the total market-based greenhouse gas emissions for training were 0 tons CO2eq."