{"id":3517,"date":"2024-05-22T14:03:46","date_gmt":"2024-05-22T12:03:46","guid":{"rendered":"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/"},"modified":"2024-05-22T14:03:46","modified_gmt":"2024-05-22T12:03:46","slug":"the-web-conf-2024-singapore-trip-report","status":"publish","type":"post","link":"https:\/\/tomayac.com\/wordpress\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/","title":{"rendered":"The Web Conf 2024, Singapore: Trip report"},"content":{"rendered":"\n\t\t\t<h2 id=\"background\" tabindex=\"-1\">Background <a class=\"direct-link\" href=\"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/#background\" aria-hidden=\"true\">\ud83d\udd17<\/a><\/h2>\n<p><img src=\"https:\/\/blog.tomayac.com\/images\/thewebconf2024--92d9bkb5eyl.png\" alt=\"The Web Conf heading in front of the conference venue.\" \/><\/p>\n<p>The <a href=\"https:\/\/www2024.thewebconf.org\/\">Web Conference<\/a> (formerly known as\n<a href=\"https:\/\/thewebconf.org\/\">WWW<\/a>) is an international conference focused on\nexploring the current state and the evolution of the Web through the lens of\ndifferent scientific disciplines, including computing science, social science,\neconomics, and political sciences. It's organized by the Association for\nComputing Machinery (<a href=\"https:\/\/www.acm.org\/\">ACM<\/a>) Special Interest Group on the\nWeb (<a href=\"https:\/\/www.sigweb.org\/\">SIGWEB<\/a>) and is held annually in a different\nlocation around the world. The 2024 conference took place in Singapore from May\n13 to 17. It's attended by 70% academia and 30% industry.\n<a href=\"https:\/\/www2024.thewebconf.org\/sponsors\/sponsors-and-supporters\/\">Google was a Gold sponsor<\/a>,\ntogether with TikTok.<\/p>\n<h2 id=\"conference\" tabindex=\"-1\">Conference <a class=\"direct-link\" href=\"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/#conference\" aria-hidden=\"true\">\ud83d\udd17<\/a><\/h2>\n<h3 id=\"day-1\" tabindex=\"-1\">Day 1 <a class=\"direct-link\" href=\"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/#day-1\" aria-hidden=\"true\">\ud83d\udd17<\/a><\/h3>\n<h4 id=\"online-trust-day\" tabindex=\"-1\">Online trust day <a class=\"direct-link\" href=\"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/#online-trust-day\" aria-hidden=\"true\">\ud83d\udd17<\/a><\/h4>\n<h5>Keynote I: Factuality Challenges in the Era of Large Language Models<\/h5>\n<p><strong>Speaker:<\/strong>\n<a href=\"https:\/\/mbzuai.ac.ae\/study\/faculty\/preslav-nakov\/\">Dr. Preslav Nakov<\/a>,\nProfessor and Department Chair of NLP, MBZUAI<\/p>\n<p>Fact-checking the output of LLMs: Decompose the output of an LLM into its\nindividual claims, decide which are check-worthy, check one-by-one:<\/p>\n<ul>\n<li><a href=\"https:\/\/arxiv.org\/pdf\/2311.09000\">Factcheck-Bench: Fine-Grained Evaluation Benchmark for Automatic Fact-checkers<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/yuxiaw\/Factcheck-GPT\">yuxiaw\/Factcheck-GPT<\/a><\/li>\n<\/ul>\n<p>Detecting LLM-generated texts:<\/p>\n<ul>\n<li>\n<p><a href=\"https:\/\/d4mucfpksywv.cloudfront.net\/papers\/GPT_2_Report.pdf\">Release Strategies and the Social Impacts of Language Models<\/a><\/p>\n<\/li>\n<li>\n<p><a href=\"https:\/\/huggingface.co\/openai-community\/roberta-large-openai-detector\">RoBERTa<\/a>\n(for GTP-2)<\/p>\n<\/li>\n<\/ul>\n<p>Arabic LLM:<\/p>\n<ul>\n<li>\n<p><a href=\"https:\/\/inceptioniai.org\/jais\/\">Jais Arabic LLM<\/a>, see the\n<a href=\"https:\/\/huggingface.co\/core42\/jais-13b-chat#:~:text=prompt_eng%20%3D%20%22%23%23%23%20Instruction%3A%20Your%20name%20is%20Jais%2C%20and%20you,%D9%88%20%5B%7CAI%7C%5D%3A%5Cn%23%23%23%20Input%3A%20%5B%7CHuman%7C%5D%20%7BQuestion%7D%5Cn%23%23%23%20Response%3A%20%5B%7CAI%7C%5D%22\">prompt instructions<\/a>.<\/p>\n<\/li>\n<li>\n<p><a href=\"https:\/\/arabic-gpt.ai\/\">Demo<\/a><\/p>\n<\/li>\n<\/ul>\n<p>Audience question: Why don't we use LLMs for what they are good for: working\nwith language like reformulating or summarizing, but not asking them to come up\nwith facts. \u2014 We probably could, but hallucination problems there as well.<\/p>\n<h5>Keynote II: Building Trust and Safety on Facebook<\/h5>\n<p><strong>Speaker:<\/strong> <a href=\"https:\/\/www.linkedin.com\/in\/lgpueyo\">Llu\u00eds Garcia Pueyo<\/a>, Director\nof Engineering, Meta<\/p>\n<p>For many languages there isn't enough actually harmful labeled content, so\nmodels are trained on artificially oversampled labeled examples.<\/p>\n<ul>\n<li><a href=\"https:\/\/research.facebook.com\/publications\/semi-supervised-monotonic-regression-for-calibrating-social-media-classifiers\/\">Semi Supervised Monotonic Regression For Calibrating Social Media Classifiers<\/a><\/li>\n<\/ul>\n<p>Facebook posts ranking formula: probability you like something, probability you\nshare something, probability you hide something. Like, comment, and send are not\ngood signals for bad experiences. Hiding, reporting from the three dot overflow\nmenu are.<\/p>\n<ul>\n<li><a href=\"https:\/\/dl.acm.org\/doi\/pdf\/10.1145\/3543507.3583883\">Detecting and Limiting Negative User Experiences in Social Media Platforms<\/a><\/li>\n<\/ul>\n<p>Future challenges with LLMs:<\/p>\n<ul>\n<li><a href=\"https:\/\/arxiv.org\/pdf\/2402.17152\">Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations<\/a><\/li>\n<\/ul>\n<p>Meta folks are organizing the\n<a href=\"https:\/\/integrity-workshop.org\/\">Integrity Workshop<\/a> series.<\/p>\n<h5>The Dynamics of (Not) Unfollowing Misinformation Spreaders<\/h5>\n<p>Collected health misinformation URLs and tweets tagged by\n<a href=\"https:\/\/www.politifact.com\/\">PolitiFact<\/a>. Found users who share this content on\nTwitter. Denoted these users misinformation spreaders. Also pulled the followers\nof spreaders. They found that misinformation ties are rarely severed, with\nunfollowing rates of 0.52% per month. Users are 31% more likely to unfollow\nnon-misinformation spreaders than they are to unfollow misinformation spreaders.\nReciprocity, initial exposure, and ideology are the most important factors for\npredicting unfollowing.<\/p>\n<ul>\n<li><a href=\"https:\/\/arxiv.org\/pdf\/2401.13480\">The Dynamics of (Not) Unfollowing Misinformation Spreaders<\/a><\/li>\n<\/ul>\n<h4 id=\"web4all-sponsored-by-google\" tabindex=\"-1\">Web4All (Sponsored by Google) <a class=\"direct-link\" href=\"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/#web4all-sponsored-by-google\" aria-hidden=\"true\">\ud83d\udd17<\/a><\/h4>\n<h5>Touchpad Mapper: Exploring Non-Visual Touchpad Interactions for Screen-Reader Users<\/h5>\n<p>Touchpad Mapper: maps the position of objects in images to the touchpad area, so\nwhen the screen reader user moves their finger over the touchpad, the position\nof the finger is taken into account for announcing the image contents.<\/p>\n<ul>\n<li><a href=\"https:\/\/athersharif.github.io\/touchpadmapper\/index.html\">Touchpad Mapper<\/a>,\nrequires a backend app to extract the exact physical coordinates of the finger\non the touchpad.<\/li>\n<\/ul>\n<figure>\n    <img alt=\"Diagram showing the data flow from touchpad to Mac app to backend server to browser to user.\" src=\"https:\/\/blog.tomayac.com\/images\/thewebconf2024--eur2u7qn0cm.png\" \/>\n<figcaption>Touchpad Mapper makes images and videos spatially explorable with the\ntouchpad.<\/figcaption>\n<\/figure>\n<h4 id=\"beyond-facts-4th-international-workshop-on-computational-methods-for-online-discourse-analysis\" tabindex=\"-1\">Beyond Facts: 4th International Workshop on Computational Methods for Online Discourse Analysis <a class=\"direct-link\" href=\"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/#beyond-facts-4th-international-workshop-on-computational-methods-for-online-discourse-analysis\" aria-hidden=\"true\">\ud83d\udd17<\/a><\/h4>\n<h5>Leveraging Large Language Models to Detect Influence Campaigns on Social Media<\/h5>\n<p>They used an LLM to determine if, based on user metadata and network structures,\na user is part of an organized information campaign. Their model was trained\nwith\n<a href=\"https:\/\/www.kaggle.com\/datasets\/fivethirtyeight\/russian-troll-tweets\">Russian troll tweets<\/a>.\nThese Moderation Research datasets are available freely from the\n<a href=\"https:\/\/transparency.x.com\/en\/reports\/moderation-research.html\">X Transparency Center<\/a>.<\/p>\n<h5>Escaping the Echo Chamber: The Quest for Normative News Recommender Systems by Abraham Bernstein<\/h5>\n<p>Looked at news recommendations. Interesting datasets:\n<a href=\"https:\/\/movielens.org\/\">MovieLens<\/a>,\n<a href=\"https:\/\/www.kaggle.com\/datasets\/somnambwl\/bookcrossing-dataset\">Book-Crossing<\/a>.<\/p>\n<h5>Towards Fact-check Summarization Leveraging on Argumentation Elements tied to Entity Graphs<\/h5>\n<p>Used <a href=\"https:\/\/www.politifact.com\/\">PolitiFact<\/a> as ground truth and compared GTP4\nvs. <a href=\"https:\/\/customgpt.ai\/\">Custom GPT<\/a> to see if the models could come up with\nsimilar results.<\/p>\n<h5>Detection Distortions in Science Reporting by Isabelle Augenstein<\/h5>\n<p>Looked at how journalists cover scientific research. Scientific findings\nfrequently undergo subtle distortions when reported, e.g., with regard to\ncertainty, generality, and causality.<\/p>\n<ul>\n<li><a href=\"https:\/\/sbert.net\/\">Sentence BERT<\/a>: framework to compute sentence \/ text\nembeddings for more than 100 languages. These embeddings can then be compared,\ne.g. with cosine-similarity, to find sentences with a similar meaning.<\/li>\n<\/ul>\n<h3 id=\"day-2\" tabindex=\"-1\">Day 2 <a class=\"direct-link\" href=\"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/#day-2\" aria-hidden=\"true\">\ud83d\udd17<\/a><\/h3>\n<h4 id=\"web4all-https-www-w4a-info-2024\" tabindex=\"-1\">Web4All (<a href=\"https:\/\/www.w4a.info\/2024\/\">https:\/\/www.w4a.info\/2024\/<\/a>) <a class=\"direct-link\" href=\"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/#web4all-https-www-w4a-info-2024\" aria-hidden=\"true\">\ud83d\udd17<\/a><\/h4>\n<h5>Keynote Speech: Liddy Nevile &quot;Accessibility?&quot;<\/h5>\n<p>(Liddy Nevile is the mother of one of the organizers,\n<a href=\"https:\/\/www.w3.org\/People\/Charles\/\">Charles McCathieNevile<\/a>, aka. Chaals.) Her\nson went to university at 10 to learn Logo, so she would learn it, too. Got to\nknow folks at MIT. One of them was Tim Berners-Lee. Worked with Mosaic folks and\nhow blind kids would use it. Concerned about inappropriate content. Founded\n<a href=\"https:\/\/www.w3.org\/PICS\/\">Platform for Internet Content Selection<\/a> (PICS) W3C\ngroup, which created a numbering system to classify content. Opened the way for\nwhat people at the time thought of as curation of content. Was well received by\nthe adult industry. <a href=\"https:\/\/www.w3.org\/People\/EM\/\">Eric Miller<\/a> wondered why,\nif PICS worked, couldn't embedded descriptions, ideally structured, also work?\nCreated &quot;metadata&quot; catalog, which ended up becoming Dublin Core. Published\n<a href=\"https:\/\/asistdl.onlinelibrary.wiley.com\/doi\/full\/10.1002\/bult.105\">An Introduction to the Resource Description Framework<\/a>.<\/p>\n<figure>\n    <img alt=\"Diagram showing how Platform for\nInternet Content Selection (PICS) worked.\" src=\"https:\/\/blog.tomayac.com\/images\/thewebconf2024--rn27gockll.png\" \/>\n    <figcaption>Platform for\nInternet Content Selection (PICS)<\/figcaption>\n<\/figure>\n<p><em>Platform for Internet Content Selection (PICS,\n<a href=\"https:\/\/www.w3.org\/PICS\/PICS-FAQ-980126.html\">source<\/a>)<\/em> W3C used a grant to set\nup the <a href=\"https:\/\/www.w3.org\/WAI\/\">Web Accessibility Initiative<\/a> (WAI). Worked on\nhow to encode resources so they would be accessible to users. How could a blind\nperson find out where the cursor is? How could flickering content be avoided?\nWAI brought people together to learn about making accessible websites. Some\ncountries converted WAI into national laws. Could compliance be put into the\nresource, just like PICS? Different countries do things differently. Australia\ndoesn't prosecute people for not complying with\n<a href=\"https:\/\/www.w3.org\/TR\/WCAG21\/\">Web Content Accessibility Guidelines<\/a> (WCAG).\nWorked on making math books accessible with MathML. Inclusion (make something\naccessible just in case) or accessibility (make something accessible just in\ntime). Worked on structured accessibility data.\n<a href=\"https:\/\/en.wikipedia.org\/wiki\/Jutta_Treviranus\">Jutta Treviranus<\/a> worked on\n<a href=\"https:\/\/raisingthefloor.org\/our-approach-accessforall\/\">Access4Al<\/a>: <em>&quot;Whether\nusing a public workstation, or engaging in an on-line learning environment,\ncomputer systems should fit the individual preferences and requirements of the\nuser, especially if the user requires an alternative access system. An\ninternational effort is underway to create a common specification for expressing\npersonal preferences for all systems affecting the user interface and content\nretrieval&quot;.<\/em> Took accessibility description work further with\n<a href=\"https:\/\/schema.org\/accessibilitySummary\">schema.org<\/a> in the form of\n<a href=\"https:\/\/schema.org\/accessibilitySummary\"><code>accessibilitySummary<\/code><\/a>. Now there's\n<a href=\"https:\/\/www.iso.org\/standard\/80519.html\">ISO\/IEC 4932<\/a>\n(<a href=\"https:\/\/www.dublincore.org\/groups\/access\/standards\/\">Core Accessibility Metadata<\/a>).\nLooking forward to seeing accessibility services working with AI at last doing\nwhat we dreamed of so many years ago. If accessibility is an opt-in, it needs to\nbe made sure that the data isn't abused. Good privacy fields help, it's about\nthe people, everyone occasionally needs accessibility features. (The speaker\nsaid they were very thankful to Google for the <a href=\"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/schema.org\">schema.org<\/a> work.)<\/p>\n<h5>Decoding the Privacy Policies of Assistive Technologies<\/h5>\n<p>They looked at the privacy policies of various assistive technologies companies.\nIt's not great; some of them collect data about sexual orientation.<\/p>\n<h5>QualState: Finding Website States for Accessibility Evaluation<\/h5>\n<p>Web accessibility evaluation engine called\n<a href=\"https:\/\/github.com\/qualweb\/qualstate\">QualState<\/a> for automatically testing the\naccessibility of web apps. QualState loads a page, performs actions. Identifies\nevents on page. Clicks links, buttons, and submits forms based on the DOM tree.\nIgnores some nodes, but needs a full DOM tree to see which states the page was\nalready in.<\/p>\n<h5>A Universal Web Accessibility Feedback Form: A Participatory Design Study<\/h5>\n<p>Hypothesis: companies don't get accessibility complaints because the feedback\nforms are inaccessible. Placement of the feedback link reduces or creates entry\nbarriers. Add an introduction paragraph. Describe each step clearly. Make sure\nconstraints (multiple choice checkboxes, radio buttons, etc.) are clearly\nexplained, and don't rely purely on technical error message handling. Allow\ncontact information to be added optionally. The form should provide details on\nwhere exactly on the website a problem occurred. Allow for system settings to be\nshared and the used assistive technology. Make their tool available on GitHub:\n<a href=\"https:\/\/github.com\/human-centered-systems-lab\/a11y-feedback\">human-centered-systems-lab\/a11y-feedback<\/a>.<\/p>\n<h5>Accessibility and AI<\/h5>\n<p>Can AI coding assistants produce accessible UI code? Yes, when explicitly\ninstructed to do so. They are not reliable and subject to hallucinations. Dark\nmode button worked fine. Image had mixed Japanese\/English alt text. Can AI\ncoding assistants eliminate the need for developer accessibility awareness?\nAccessibility features are not applied consistently. There might be states that\nget missed. Empty alt texts are hard to catch, since it looks intended. More\nbenchmarks are needed. Fine-tuning models for accessible UI. AI powered DevTools\ncan help, too. Copilot doesn't make any claims about the accessibility of its\ncreated code.<\/p>\n<h5>Evaluating the Effectiveness of STEM Images Captioning<\/h5>\n<p>Teaches university students Web development. Asks them to take the\n<a href=\"https:\/\/nomouse.org\/\">#NoMouse Challenge<\/a>. Split their group in two groups: one\nwas trained on image accessibility, one wasn't. Correctness (does the\ndescription accurately describe what the image depicts) vs. usefulness (does the\ndescription accurately describe the conveyed meaning) of describing images.\nCreated AI image descriptions with\n<a href=\"https:\/\/huggingface.co\/blog\/idefics\">IDEFICS<\/a>\n(<a href=\"https:\/\/huggingface.co\/spaces\/HuggingFaceM4\/idefics_playground\">demo<\/a>).\nStudents were asked to evaluate human-generated vs. AI-generated descriptions.\nDescribing STEM images (like diagrams describing photosynthesis) generally is\nhard. In all cases, humans performed better. The AI had more problems with STEM\nimages. Planning to compare other AI engines. Also thinking about ways to\nimprove the prompt engineering (e.g., &quot;Describe this detailed scientific diagram\nin a way that the description is useful to a blind user&quot;).<\/p>\n<h5>Making Accessible Movies Easily: An Intelligent Tool for Authoring and Integrating Audio Descriptions to Movies<\/h5>\n<p>Steps for creating movie audio descriptions (AD): read existing subtitles with\nOCR, identify speech gaps based on missing subtitles, create scene description\ntexts with <a href=\"https:\/\/arxiv.org\/abs\/2305.06355\">VideoChat<\/a> or\n<a href=\"https:\/\/arxiv.org\/abs\/2306.02858\">VideoLLaMa<\/a>, then use ChatGPT to merge the\nsubtitles and the scene description. The final step is to run text to audio and\naudio mixing. Created an app called EasyAD that incorporates all these steps.\nQuality was evaluated as being good, but speed was still slow, feedback also\nsuggested more languages than Chinese should be supported.<\/p>\n<figure>\n    <img alt=\"Three steps: describe video. Image with subtitles. Merge and generate audio description.\" src=\"https:\/\/blog.tomayac.com\/images\/thewebconf2024--aw7gw51esrb.png\" \/>\n    <figcaption>The movie audio description system.<\/figcaption>\n<\/figure>\n<h5>Towards Effective Communication of AI-Based Decisions in Assistive Tools: Conveying Confidence and Doubt to People with Visual Impairments at Accelerated Speech<\/h5>\n<p>Screen reader users typically comprehend speech 3 times faster than sighted\nusers. Speeded up voices lose some of the emotions of speech like confidence or\ndoubt. Up to a factor of 1.5 to 2 this effect isn't noticeable, at faster speeds\nit's noticeable and research is required to reintroduce these emotions.<\/p>\n<h4 id=\"welcome-reception\" tabindex=\"-1\">Welcome reception <a class=\"direct-link\" href=\"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/#welcome-reception\" aria-hidden=\"true\">\ud83d\udd17<\/a><\/h4>\n<p>The welcome reception took place in the\n<a href=\"https:\/\/www.tipsyunicorn.com.sg\/\">Tipsy Unicorn<\/a>.<\/p>\n<figure>\n<img alt=\"A band playing on a stage. In front of the stage is a pool with persons, one of them me. Behind the stage is a screen with a conference slide announcing the resource track. The slide has my headshot.\" src=\"https:\/\/blog.tomayac.com\/images\/thewebconf2024--4hpmqvcole7.png\" \/>\n<figcaption>I'm in this picture twice.<\/figcaption>\n<\/figure>\n<figure>\n<img alt=\"A band playing on a stage. In front of the stage is a pool. Behind the stage is a screen with a conference slide announcing the sponsors.\" src=\"https:\/\/blog.tomayac.com\/images\/thewebconf2024--htkl1djozfm.png\" \/>\n<figcaption>The sponsors: TikTok, Baidu, Google, and 6Estates.<\/figcaption>\n<\/figure>\n<h3 id=\"day-3\" tabindex=\"-1\">Day 3 <a class=\"direct-link\" href=\"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/#day-3\" aria-hidden=\"true\">\ud83d\udd17<\/a><\/h3>\n<h4 id=\"keynote-1-challenges-toward-agi-and-its-impact-to-sic-the-web\" tabindex=\"-1\">Keynote#1: Challenges toward AGI and its impact to [sic] the Web <a class=\"direct-link\" href=\"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/#keynote-1-challenges-toward-agi-and-its-impact-to-sic-the-web\" aria-hidden=\"true\">\ud83d\udd17<\/a><\/h4>\n<p><strong>Speaker:<\/strong> <a href=\"https:\/\/keg.cs.tsinghua.edu.cn\/jietang\/\">Jie Tang<\/a> This keynote\nwas about how they created the Chinese ChatGPT called\n<a href=\"https:\/\/chatglm.cn\/\">ChatGLM (\u667a\u8c31\u6e05\u8a00)<\/a>, which means &quot;clear words of wisdom&quot;.<\/p>\n<h4 id=\"web4good\" tabindex=\"-1\">Web4Good <a class=\"direct-link\" href=\"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/#web4good\" aria-hidden=\"true\">\ud83d\udd17<\/a><\/h4>\n<h5>CapAlign: Improving Cross Modal Alignment via Informative Captioning for Harmful Meme Detection<\/h5>\n<p>The authors prompt a large language model (ChatGPT) to ask informative questions\nto a pre-trained vision-language model\n(<a href=\"https:\/\/huggingface.co\/docs\/transformers\/main\/en\/model_doc\/blip-2\">BLIP-2<\/a>)\nand use the dialogs to generate a high-quality image caption. To align the\ngenerated caption with the textual content of a meme, they use an LLM with\ninstructions to generate informative captions of the meme and then prepend it\nwith the attributes of the visual content of a meme to a prompt-based LLM for\nprediction. (I would love to see this run on top of Memegen.)\n[<a href=\"https:\/\/dl.acm.org\/doi\/pdf\/10.1145\/3589334.3648146\">Paper<\/a>]<\/p>\n<figure>\n<img alt=\"Diagram showing hwo the CapAlign system makes two LLMs talk to each other.\" src=\"https:\/\/blog.tomayac.com\/images\/thewebconf2024--0ivp2ppy3sqs.png\" \/>\n<figcaption>The CapAlign system<\/figcaption>\n<\/figure>\n<h4 id=\"panel-discussion-on-llm-impact-on-the-web\" tabindex=\"-1\">Panel Discussion on LLM Impact on the Web <a class=\"direct-link\" href=\"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/#panel-discussion-on-llm-impact-on-the-web\" aria-hidden=\"true\">\ud83d\udd17<\/a><\/h4>\n<p><strong>Panelists:<\/strong><\/p>\n<ul>\n<li><strong>Panel Chair:<\/strong>\n<ul>\n<li>Andrew Tomkins, Research Scientist, Google<\/li>\n<\/ul>\n<\/li>\n<li><strong>Panelists:<\/strong>\n<ul>\n<li>Jon Kleinberg, Cornell University<\/li>\n<li>Yoelle Maarek, Chief Researcher, Technology Innovation Institute<\/li>\n<li>Jie Tang, Tsinghua University<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<p><strong>Questions:<\/strong><\/p>\n<ul>\n<li>\n<p>Do we expect websites to have LLM-based front-ends?<\/p>\n<ul>\n<li>Jie thinks it's quite possible. Each website may have an agent that could\nalso interact with other websites' agents.<\/li>\n<li>Yoelle thinks that if the content of websites is generated by LLMs and LLMs\ntrain themselves on the content they created, this may lead to a rich get\nricher symptom and all LLMs learn the same. Maybe more diverse LLMs can help\nrather than one dominating one. Hallucinations are a big problem, and they\nwill continue to be. People need the feeling that the information comes from\nsomewhere, to have sources. We need to be careful not to take this feeling\naway.<\/li>\n<li>The most popular app on the Web is search. Search puts itself between the\npage and the searcher. If search doesn't lead to traffic to pages, there's\nno incentive to create content.<\/li>\n<\/ul>\n<\/li>\n<li>\n<p>Websites don't need to expose APIs anymore, agents can just talk to websites\nusing natural language. Will this cause specialized search engines to arise?\nDo we expect one central agent to rule them all?<\/p>\n<ul>\n<li>Yoelle says before common Web search engines, there was a federated search\nengine, but it died. Strongly believing in RAG, you need special agents to\nsurface hidden content.<\/li>\n<li>Jon states it's an old question, special agents like for flight search. You\ncould take special agents and hide them under one common interface. It's\nmostly a UI question.<\/li>\n<li>Jie says we have a network of webpages, and later Linked Data. Now AI to\nanswer questions. We could have linked AI to answer special questions.<\/li>\n<\/ul>\n<\/li>\n<li>\n<p>The Web is special. Someone has gone through the work of, for example,\ncollecting great spots to visit at a place. We would love for this person to\nkeep the benefits. We can do so through advertising. Now the model is\nchanging. Why would people keep creating content under these circumstances?\nWhat are possible models for this to work in the future?<\/p>\n<ul>\n<li>Jon says the Web has always been powered by altruism. Search isn't always\nabout finding the answer, but also about exploring the landscape. People\nwant to hear different takes on a question. LLMs will not just be used to\nfind one answer. If there's economic value created. Mashups is a 2005\nconcept, we mashed up Google Maps with stuff.<\/li>\n<li>Yoelle states most websites are automatically created. It's like AirBnB,\nit's business, not regular people renting out a spare room. If you have\nspecialized RAG-supported LLMs, you have a transaction when hidden content\nis being found. This isn't the Open Web, maybe it has already disappeared.\nEconomic value could be protected like this.<\/li>\n<li>Jie isn't sure about this. People only want to consume, they don't\nnecessarily search for something concrete, like TikTok or Douyin.<\/li>\n<li>Yoelle really wants to disagree. We want serendipity, I love this journey of\nsearching. It's something I must have to earn.<\/li>\n<li>Jon talks about the economic model. If you're a standup comedian, you start\nimitating others, and eventually you develop your own style. That's LLMs.\nYou don't owe the comedians you took inspiration from early on.<\/li>\n<\/ul>\n<\/li>\n<li>\n<p>Audience question by Natasha Noy: I want to broaden this. If we have a highly\ncurated extra layer, it hides this personal layer. We need to discuss this\nlayer.<\/p>\n<ul>\n<li>Jon says this could be something like an LLM giving out &quot;Likes&quot;, or Google\nScholar citation counts. Could this be self-prophesying, so people create\ncontent only to be cited by LLMs.<\/li>\n<\/ul>\n<\/li>\n<li>\n<p>Want to talk about privacy and safety. But before that, I want to talk about\ncrawling. It's a massive business. Crawlers asking for access to content could\nlook different than regular users, they already do. Could there be a third\nclass of LLM traffic to websites?<\/p>\n<ul>\n<li>Yoelle sees the point. People don't protect their websites too much, like\nwith robots.txt, because they know they get traffic back. Now with LLMs it's\ndifferent. They don't get the value back. It's costly for providers to\ncrawl, Azure, Google, AWS, they make a lot of money off LLM crawlers. We\nneed to think about protocols to support this.<\/li>\n<li>Jie thinks in the future this depends, if in the future the Web will\ninteract with human beings, if the Web itself is a personal assistant, the\nWeb will change. Not sure how.<\/li>\n<li>Jon thinks this question is orthogonal to the question of LLMs. We already\nhave alerts and notifications like has the flight price changed, has an\narticle been published on a topic. This is a pilot case for what LLMs could\ndo in a general manner.<\/li>\n<\/ul>\n<\/li>\n<li>\n<p>Audience question: One of the issues now is who owns the content creator?\nGoogle etc. make profit from profiling users. Would we all come to a\nconference to listen to bot-written papers? When we lose trust in the\nreflective power of an LLM, who's going to pay for this technology? Is it\neither you, or my personal information? Will this model work in the future?<\/p>\n<ul>\n<li>Jie isn't sure about the economic aspect. AI is still occupied with\nimproving its performance. The trust still needs to be earned.<\/li>\n<li>Yoelle means people won't use AI for trivial things they can do themselves.\nReasoning examples we see today are trivial, because the AI is still\nlearning. In the early days of the Web it was authority through PageRank and\nclicks that brought you value as a creator. Research may be needed to\nexplore if we can somehow give back to creators, I hope it will come\nnaturally. We can also think as creators, what information do we want to\nmake accessible to LLMs.<\/li>\n<li>Jon asks what's the value we're adding. If we're angry at an LLM, what can I\nadd? Wolfram Alpha now just solves math problems which a hundred years ago\nyou could publish a math paper about.<\/li>\n<\/ul>\n<\/li>\n<li>\n<p>Audience question: The age of information abundance. In the past, there was\ninformation scarcity. For example, you needed to travel to different countries\nto get information. Now it's the opposite, you have way too much information.\nGenerative AI makes this a lot worse, it creates so much information.<\/p>\n<ul>\n<li>Yoelle thinks the question is funny. LLMs are good at finding hidden\ninformation. But you don't know if it's a hallucination. You don't have the\ncontext, so you can't easily verify. The world now isn't deterministic.\nWe're all computer scientists, we're used to determinism. Now it's not the\none truth. You live in a fuzzy world now. Maybe we all become computer\nartists. Scared by the uncertainty. I want proof, I lost this.<\/li>\n<li>Jon disagrees a bit. You visit a doctor, and each doctor you visit tells you\nsomething different. Indeterminism existed before. It's a fascinating\ntension. Huge fan of the abundance question, wrote papers about this. In the\nearly 1970ies, there was a book about information abundance. Abundance is\nconsuming human attention.<\/li>\n<\/ul>\n<\/li>\n<li>\n<p>Let's touch on trust and safety. There's privacy questions, there's government\nstandpoint questions, there's questions on where LLMs should be hosted.\nFinance has concerns about data safety. Can you share thoughts about what's\nthe biggest risk right now?<\/p>\n<ul>\n<li>Jie thinks all this is very important. Technology is super important. AI\nwill self-improve and self-reflect. We could have built a common model to\ncheck the quality of models. If AIs in the future will be smarter about\nthis, this would be great.<\/li>\n<li>Yoelle says it's super important to have many open source LLMs. We need\ndiversity. You could come back to different models, even if they are biased.\nFor many sensitive topics, you can bring models in-house trained on your\ndata, even with lower general quality.<\/li>\n<li>Jon means powerful tools will reflect society, including its biases. We rely\non LLMs as to make decisions.<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h4 id=\"history-of-the-web-1\" tabindex=\"-1\">History of the Web #1 <a class=\"direct-link\" href=\"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/#history-of-the-web-1\" aria-hidden=\"true\">\ud83d\udd17<\/a><\/h4>\n<h5>Viola, Pei Wei, and the Fights for Interactive Media<\/h5>\n<figure>\n<img alt=\"Several web browsers displayed over a map of the world based on where they were created.\" src=\"https:\/\/blog.tomayac.com\/images\/thewebconf2024--d9hf6givri.png\" \/>\n<figcaption>Early Web browsers\n(<a href=\"https:\/\/computerhistory.org\/blog\/happy-25th-birthday-to-the-public-web\/\">Source<\/a>)<\/figcaption>\n<\/figure>\n<p>Interesting historical reference:\n<a href=\"https:\/\/dl.acm.org\/doi\/abs\/10.1145\/105783.105799\">MediaView: a general multimedia digital publication system<\/a>.<\/p>\n<h5>Digital Democracy at Crossroads: A Meta-Analysis of Web and AI Influence on Global Elections<\/h5>\n<p>They looked at papers from the past until today that looked at elections.\nSuggestions for generative AI companies to tackle AI disinformation: Implement\nwatermarking and strict verification, regulate AI chatbots, mark AI-generated\ncontent as such. Government should require politics-related material to be\nmarked specially if AI was involved. Educate users to identify AI-generated\ncontent. Fake news isn't new, but the scale is way different now.\n[<a href=\"https:\/\/dl.acm.org\/doi\/pdf\/10.1145\/3589335.3652003\">Paper<\/a>]<\/p>\n<h5>History in Making: Political Campaigns in the Era of Artificial Intelligence-Generated Content<\/h5>\n<p>Historically, we had user-generated content. Now we have AI-generated content.\nIt's the year of the elections, in almost 60 countries, covering half of the\nEarth's population. Political campaigns make use of AI. Huma\/real life person\nimpersonating makes caller bots possible.\n<a href=\"https:\/\/www.reuters.com\/technology\/meet-ashley-worlds-first-ai-powered-political-campaign-caller-2023-12-12\/\">Ashley caller bot<\/a>\nin the US. AI-generated deepfake makes\n<a href=\"https:\/\/www.politico.eu\/article\/pakistans-imran-khan-use-ai-artificial-intelligence-make-victory-speech-from-jail\/\">campaigning from prison<\/a>\npossible. Also malicious deepfakes. ChatGPT is known to be left-leaning in the\nUS. Governments crack down on services and tools to create and spread\nAI-generated content and limit access to user data.\n[<a href=\"https:\/\/dl.acm.org\/doi\/pdf\/10.1145\/3589335.3652000\">Paper<\/a>]<\/p>\n<h5>Me, the Web and Digital Accessibility<\/h5>\n<p>Fun anecdote: IE showed the <code>alt<\/code> attribute like a tooltip. The author is the\nofficial translator of the WCAG standard for Portuguese and got into\naccessibility when he was made aware that Brazil's government required websites\nto be accessible. [<a href=\"https:\/\/dl.acm.org\/doi\/pdf\/10.1145\/3589335.3652002\">Paper<\/a>]<\/p>\n<h5>From Files to Streams: Revisiting Web History and Exploring Potentials for Future Prospects<\/h5>\n<p>Users love fast web content and there's an economic value in performance. 1991:\nall content text-based and delivered from one server. Now all types of content\nare delivered via CDN. Cites HTTP Archive stats on website size and First\nContentful Paint. FCP didn't improve. HTTP was FTP inspired. TCP handshake cost\nneeded to be paid. Less files meant faster loading time. Keep-alive allows\nreusing TCP connections. HTTP\/2 and HTTP\/3 reduced the overhead, no more\nline-blocking, multiplexing and streams,\n<a href=\"https:\/\/blog.cloudflare.com\/introducing-0-rtt\">0-RTT<\/a>. Server delivery was\nimproved. Client side lacked. JS (1995), CSS (1996), DOM (1998). Bundling as a\nsolution to make less requests. Browserify (2013), then Webpack. Webpack's\npopularity peaked just when HTTP\/2 was introduced. A solution (bundling) for a\nproblem that doesn't exist anymore. Erwin Hofman:\n<a href=\"https:\/\/www.erwinhofman.com\/blog\/two-main-performance-debts-of-http1\/#:~:text=Bundling%20is%20an%20anti-pattern%20in%20HTTP\/2\">&quot;Bundling is an antipattern in HTTP\/2&quot;<\/a>.\nRender-blocking as a major annoyance, can use dead code elimination and critical\nCSS identification. Can stream content over HTTP\/2 or \/3. Sees research\nchallenge in automatic content usage detection and ordering of JavaScript. They\nstream Web content via WebSocket (<a href=\"https:\/\/streaming-demo.de\/\">demo<\/a>).\n[<a href=\"https:\/\/dl.acm.org\/doi\/pdf\/10.1145\/3589335.3652001\">Paper<\/a>]<\/p>\n<h4 id=\"posters\" tabindex=\"-1\">Posters <a class=\"direct-link\" href=\"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/#posters\" aria-hidden=\"true\">\ud83d\udd17<\/a><\/h4>\n<h5>Automating Website Registration for Studying GDPR Compliance<\/h5>\n<img alt=\"Poster for the 'Automating Website Registration for Studying GDPR Compliance' paper.\" src=\"https:\/\/blog.tomayac.com\/images\/thewebconf2024--8g59e3vyi5c.jpg\" \/>\n<p>They used a headless browser to sign up to websites and then see if they had\nGDPR violations. [<a href=\"https:\/\/dl.acm.org\/doi\/10.1145\/3589334.3645709\">Paper<\/a>]<\/p>\n<h5>Breaking the Trilemma of Privacy, Utility, Efficiency via Controllable Machine Unlearning<\/h5>\n<img alt=\"Poster for the 'Breaking the Trilemma of Privacy, Utility, Efficiency via Controllable Machine Unlearning' paper.\" src=\"https:\/\/blog.tomayac.com\/images\/thewebconf2024--k2s5vcoaeyk.jpg\" \/>\n<p>TIL about the concept of Machine Unlearning. The work explains how parts of a\nmodel's training data can be removed without having to retrain the entire model.\n[<a href=\"https:\/\/arxiv.org\/abs\/2310.18574\">Paper<\/a>]<\/p>\n<h5>A Worldwide View on the Reachability of Encrypted DNS Services<\/h5>\n<img alt=\"Poster for the 'A Worldwide View on the Reachability of Encrypted DNS Services' paper.\" src=\"https:\/\/blog.tomayac.com\/images\/thewebconf2024--3akv9dy97c6.jpg\" \/>\n<p>They compare different privacy-preserving ways of how DNS can work over\nencrypted data and how they affect global reachability.\n[<a href=\"https:\/\/dl.acm.org\/doi\/10.1145\/3589334.3645539\">Paper<\/a>]<\/p>\n<h5>Uncovering the Hidden Data Costs of Mobile YouTube Video Ads<\/h5>\n<img alt=\"Poster for the 'Uncovering the Hidden Data Costs of Mobile YouTube Video Ads' paper.\" src=\"https:\/\/blog.tomayac.com\/images\/thewebconf2024--oicytbqmq0k.jpg\" \/>\n<p>They look at wasted bandwidth from YouTube ads. Apparently we preload quite\neagerly, even if most ads are skipped as soon as possible.\n[<a href=\"https:\/\/dl.acm.org\/doi\/10.1145\/3589334.3645496\">Paper<\/a>]<\/p>\n<h3 id=\"day-4\" tabindex=\"-1\">Day 4 <a class=\"direct-link\" href=\"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/#day-4\" aria-hidden=\"true\">\ud83d\udd17<\/a><\/h3>\n<h4 id=\"keynote-2-revisiting-the-behavioral-foundations-of-user-modeling-algorithms\" tabindex=\"-1\">Keynote#2: Revisiting the Behavioral Foundations of User Modeling Algorithms <a class=\"direct-link\" href=\"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/#keynote-2-revisiting-the-behavioral-foundations-of-user-modeling-algorithms\" aria-hidden=\"true\">\ud83d\udd17<\/a><\/h4>\n<p><strong>Speaker:<\/strong> <a href=\"https:\/\/www.cs.cornell.edu\/home\/kleinber\/\">Jon Kleinberg<\/a>\nAlgorithms as partners, GPT should stand for General Purpose Technology. In\ncontrast to algorithms as creators of environments, like when they curate social\nmedia experiences. When consuming a linear feed, after each item the user has\nthe chance to quit or continue scrolling. cThe algorithms are tuned to bring you\nchips because you know you crave them, while at the same time you also know that\nyou should be having a salad. Sees AI as a semi-autonomous vehicle.<\/p>\n<h4 id=\"systems-2\" tabindex=\"-1\">Systems #2 <a class=\"direct-link\" href=\"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/#systems-2\" aria-hidden=\"true\">\ud83d\udd17<\/a><\/h4>\n<h5>A Multifaceted Look at Starlink Performance<\/h5>\n<p><a href=\"https:\/\/www.measurementlab.net\/tests\/\">M-Lab Tests<\/a> makes open-source data\nabout Internet <a href=\"https:\/\/speed.measurementlab.net\/#\/\">speed tests<\/a> available.\nMeasurement Lab is led by teams based at Code for Science &amp; Society; Google,\nInc; and supported by partners around the world. Internally, the Google team\nresponsible is called\n<a href=\"https:\/\/moma.corp.google.com\/team\/1360534536144\">Open Internet Measurement<\/a>.\nMedian latencies for Starlink is ~40\u201350ms, while mobile networks are ~30ms. NA\nand EU (regions with dense ground stations concentration) enjoy the best\ncoverage. Very few locations where Starlink outperforms cellular. Closeness to\nthe ground station determines latency a lot.\n[<a href=\"https:\/\/dl.acm.org\/doi\/pdf\/10.1145\/3589334.3645328\">Paper<\/a>]<\/p>\n<h5>PASS: Predictive Auto-Scaling System for Large-scale Enterprise Web Applications<\/h5>\n<p>Paper that looks at the Web app scaling of <a href=\"https:\/\/www.meituan.com\/\">Meituan<\/a>,\na Chinese shopping platform for locally found consumer products and retail\nservices including entertainment, dining, delivery, travel, and other services.\nUses app's recent performance data to predict upcoming load. Offline model\n(looks at historical data) doesn't have information about spontaneous spikes,\nonline model (looks at just passed data) has a slight lag. Uses hybrid\nauto-scaling by combining predictive scaling with reactive scaling.\n[<a href=\"https:\/\/dl.acm.org\/doi\/pdf\/10.1145\/3589334.3645330\">Paper<\/a>]<\/p>\n<h5>FusionRender: Harnessing WebGPU's Power for Enhanced Graphics Performance on Web Browsers<\/h5>\n<p>Smaller language translation overhead. Omits runtime error checks. Recucs data\ncommunication between GPU and CPU by using pre-packed configuration. They\ncompared WebGL vs. WebGPU. <a href=\"https:\/\/threejs.org\/\">Three.js<\/a>,\n<a href=\"https:\/\/www.babylonjs.com\/\">Babylon.js<\/a>, <a href=\"https:\/\/playcanvas.com\/\">PlayCanvas<\/a>,\nand <a href=\"https:\/\/www.orillusion.com\/en\/\">Orillusion<\/a>. WebGPU is slower(!) on all\nframeworks. Frameworks render objects separately , leading to redundant\ntransmission. Merged rendering leads to reduced transmission. How can it be\ndetermined which objects can be merged? Graphics rendering configurations. They\nintroduce FusionRender. Input: user-defined configurations, output: WebGPU.\nJoins objects with identical signatures. Uses a hash map, objects are grouped\nbased on their hash. Implemented a prototype for <a href=\"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/Three.js\">Three.js<\/a>, tested on\nMacBook Pro, ThinkPad X1, and Pixel 6 with Chrome and Firefox. FusionRender\nshows improvements between ~29% and ~120% with synthetic data, about ~30% with\nreal world data. (Code:\n<a href=\"https:\/\/github.com\/qqyzk\/FusionRender\">qqyzk\/FusionRender<\/a>)\n[<a href=\"https:\/\/dl.acm.org\/doi\/pdf\/10.1145\/3589334.3645395\">Paper<\/a>]<\/p>\n<figure>\n<img alt=\"Chart showing how WebGPU is slower than WebGPU.\" src=\"https:\/\/blog.tomayac.com\/images\/thewebconf2024--oqygjzp0xr.png\" \/>\n<figcaption>The performance of WebGPU was inferior to the performance of WebGL.<\/figcaption>\n<\/figure>\n<h5>QUIC is not Quick Enough over Fast Internet<\/h5>\n<p>The paper examines QUIC's performance over high-speed networks. They find that\nover fast Internet, the UDP+QUIC+HTTP\/3 stack suffers a data rate reduction of\nup to 45.2% compared to the TCP+TLS+HTTP\/2 counterpart. This performance gap\nbetween QUIC and HTTP\/2 grows as the underlying bandwidth increases. The root\ncause is high receiver-side processing overhead, in particular, excessive data\npackets and QUIC's user-space ACKs. QUIC perceives much more packets than\nHTTP\/2. In Chromium, much more\n<a href=\"https:\/\/docs.kernel.org\/networking\/kapi.html#c.netif_receive_skb\"><code>netif_receive_skb<\/code><\/a>\ncalls are invoked for QUIC. The issue is observed on CLI data transfer clients\nand browsers (Chrome, Edge, Firefox, Opera), on different hosts (desktop,\nmobile), and over diverse networks (wired broadband, cellular). Ruled out server\nsoftware, UDP\/TCP protocols, HTTP syntax, TLS encryption, client OS, etc. as\nreasons. [<a href=\"https:\/\/dl.acm.org\/doi\/pdf\/10.1145\/3589334.3645323\">Paper<\/a>]<\/p>\n<h4 id=\"history-of-the-web-2\" tabindex=\"-1\">History of the Web #2 <a class=\"direct-link\" href=\"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/#history-of-the-web-2\" aria-hidden=\"true\">\ud83d\udd17<\/a><\/h4>\n<h5>Toward Making Opaque Web Content More Accessible: Accessibility From Adobe Flash to Canvas-Rendered Apps<\/h5>\n<iframe loading=\"lazy\" src=\"https:\/\/docs.google.com\/presentation\/d\/e\/2PACX-1vTyXwDC0QeZkbMGG2AgsCuypfFMQgFhINlocmvva96oFDjz0esEsa5N0_00EycD00EwC-tfPOBmWakj\/embed?start=false&loop=false&delayms=60000\" frameborder=\"0\" width=\"960\" height=\"569\" style=\"width: 100%;\" allowfullscreen=\"true\" mozallowfullscreen=\"true\" webkitallowfullscreen=\"true\"><\/iframe>\n<p>[<a href=\"https:\/\/docs.google.com\/presentation\/d\/1VGBhUDdPrYMqo34UrJQq5cxGmlRdDwJZzEeu9fZUZnU\/edit?usp=sharing\">Slides<\/a>]\n[<a href=\"https:\/\/dl.acm.org\/doi\/pdf\/10.1145\/3589335.3651999\">Paper<\/a>]<\/p>\n<h5>Revisiting 30 years of the Network Time Protocol<\/h5>\n<p>Network Time Protocol (NTP) has a hierarchical structure that delivers the time,\nthe stratum 0 server has the most accurate time and passes it on to lower\nlevels. NASA has proposed the\n<a href=\"https:\/\/science.nasa.gov\/mission\/mars-reconnaissance-orbiter\/mars-relay-network-interplanetary-internet\/\">Interplanetary Internet<\/a>.\nKorea Pathfinder Lunar Orbiter played K-pop from the Web. SpaceX and Blue Origin\nlook at Mars Internet. The\n<a href=\"https:\/\/www.ntp.org\/reflib\/proximity\/\">Proximity-1 Interleaved Time Synchronization<\/a>\n(PITS) protocol looks at how time synchronization could work in space.\n[<a href=\"https:\/\/dl.acm.org\/doi\/pdf\/10.1145\/3589335.3652000\">Paper<\/a>]<\/p>\n<h5>History of the Semantic Web<\/h5>\n<p>A walk down memory lane of the Semantic Web with\n<a href=\"https:\/\/faculty.rpi.edu\/james-hendler\">Jim Hendler<\/a>. Started with the\n<a href=\"https:\/\/www.scientificamerican.com\/article\/the-semantic-web\/\">Scientific American article<\/a>\nin roughly 2000. In 2005, started moving from reasoning to linking data. 2010\nwas the year of\n<a href=\"https:\/\/ieeexplore.ieee.org\/document\/5398788\">Web 3.0, the dawn of semantic search<\/a>.\n2014 Google Sem Webbers: R.V. Guha, Dan Brickley, Denny Vrandecic, Natasha Noy,\nChris Welty. Guha in 2014: &gt; 20% of pages included structured data. In 2016\nPeter Norvig mentioned &gt;60%. Facebook created Open Graph in 2011. IBM Watson\nin 2017. Facebook's Graph API made the Knowledge Graph concept more well-known.\n44% of pages now use <a href=\"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/schema.org\">schema.org<\/a> markup. Wikidata as a free\neditable knowledge base. The semantic web sort of won, but where are the\nintelligent agents? AI is getting there, but they are not directly using\nsemantic web technologies.<\/p>\n<h4 id=\"verso-a-web-browser-that-plays-old-world-blues-to-build-new-world-hope\" tabindex=\"-1\">Verso: A web browser that plays old world blues to build new world hope <a class=\"direct-link\" href=\"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/#verso-a-web-browser-that-plays-old-world-blues-to-build-new-world-hope\" aria-hidden=\"true\">\ud83d\udd17<\/a><\/h4>\n<p>A browser called <a href=\"https:\/\/github.com\/european-browser\/verso\">Verso<\/a> by\n<a href=\"https:\/\/mt.linkedin.com\/in\/daniel-thompson-yvetot-a98301222\">Daniel Thompson-Yvetot<\/a>,\nthe creator of the <a href=\"https:\/\/tauri.app\/\">Tauri<\/a> apps, a framework that uses the\nsystem's WebView to ship desktop apps. Tauri uses <code>WebView2<\/code> on Windows,\n<code>WKWewbView<\/code> on macOS, and <code>webkitgtk<\/code> on Linux. <code>WebView2<\/code> is based on\nChromium, which has a good update frequency. <code>WKWebView<\/code> means some people are\nstuck on old macOS. WebView W3C effort is slow and won't change things\nmeaningfully. Thought about using the Servo engine. Collaborated with Igalia.\nWas at Mozilla, now hosted by the Linux Foundation. Tauri folks maintain\nHTML5ever, used by Servo, Tauri, and Vercel. Engine is based on Servo, and a CLI\nfor headless integration. Deep local language integration for local translation\nand reader mode transformation. Default incognito profile management mode.\nProvides a WebView, too. Shards identities, storage, sign-in. Next steps: close\nearly funding round, set up non-profit organization at Commons Conservancy,\ndon't sell search, convince <a href=\"https:\/\/www.ngi.eu\/\">Next Generation Internet<\/a>\n(NGI) EU framework to accept the project. They want to launch in summer, coming\nto this conference was the first step.\n(<a href=\"https:\/\/web.crabnebula.cloud\/verso\/verso-nightly\/releases\">Nightly builds<\/a>)<\/p>\n<figure>\n<img alt=\"Slide with the text 'Verso is not a Servo engine'.\" src=\"https:\/\/blog.tomayac.com\/images\/thewebconf2024--mx8wptpmsk.png\" \/>\n<figcaption>Verso is not a Servo engine. Verso is\n<a href=\"https:\/\/en.wikipedia.org\/wiki\/Verlan\">Verlan<\/a> for Servo, though.<\/figcaption>\n<\/figure>\n<h3 id=\"day-5\" tabindex=\"-1\">Day 5 <a class=\"direct-link\" href=\"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/#day-5\" aria-hidden=\"true\">\ud83d\udd17<\/a><\/h3>\n<h4 id=\"keynote-4-ai-deepfakes-on-the-web-the-wicked-challenges-for-ai-ethics-law-and-technology\" tabindex=\"-1\">Keynote#4: AI deepfakes on the Web: the 'wicked' challenges for AI ethics, law and technology <a class=\"direct-link\" href=\"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/#keynote-4-ai-deepfakes-on-the-web-the-wicked-challenges-for-ai-ethics-law-and-technology\" aria-hidden=\"true\">\ud83d\udd17<\/a><\/h4>\n<p><strong>Speaker:<\/strong>\n<a href=\"https:\/\/law.unimelb.edu.au\/about\/staff\/jeannie-paterson\">Jeannie Marie Paterson<\/a>\nThe word deepfake is a combination of deep learning and fake. Can be used for\nfun or in movies (e.g., Princess Leia in Star Wars). Can be used for malicious\npurposes, like Elon Musk deepfake-generated get-rich-quick scams or scammers to\nfake family members' voices. Romance fakes with face swapping. Deepfaked\nsynthetic porn affecting many K-pop stars. Political deepfakes can affect\nelections. The liars dividend: people start questioning actual images and\ncalling them deepfake. Responses to deepfakes:<\/p>\n<ul>\n<li>\n<p><strong>Law:<\/strong> Can get active via transparency requirements, consumer regulators,\ncriminal offenses, online safety.\n<a href=\"https:\/\/oecd.ai\/en\/ai-principles\">OECD principles of ethical AI<\/a>.<\/p>\n<\/li>\n<li>\n<p><strong>Education:<\/strong> Look for errors in images and videos, but early signals like\n<a href=\"https:\/\/ieeexplore.ieee.org\/stamp\/stamp.jsp?tp=&amp;arnumber=9072088\">blinking patterns<\/a>\nhave been improved by technology.<\/p>\n<\/li>\n<li>\n<p><strong>Tech:<\/strong> Enforce voluntary guardrails to not create terrorist material,\npornography,\u2026 Release detection tools, but they often only work on the\ncompanies' own AI products. Industry has introduced an official\n<a href=\"https:\/\/c2pa.org\/post\/contentcredentials\/\">Content Credentials Icon<\/a> (C2PA).<\/p>\n<p>Participants in the initiative:<\/p>\n <img alt=\"The logos of Adobe, the BBC, Google, Intel, Microsoft, the Publicis Groupe, OpenAI, Sony, and Truepic.\" src=\"https:\/\/blog.tomayac.com\/images\/thewebconf2024--yjrlx6nag7.png\" \/>\n<p>Digital watermarks, but it's a cat and mouse game. Watermarks might not\nsurvive screenshots or photos of AI-generated photos.<\/p>\n<\/li>\n<\/ul>\n<h4 id=\"resource\" tabindex=\"-1\">Resource <a class=\"direct-link\" href=\"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/#resource\" aria-hidden=\"true\">\ud83d\udd17<\/a><\/h4>\n<h5>Ducho 2.0: Towards a More Up-to-Date Unified Framework for the Extraction of Multimodal Features in Recommendation<\/h5>\n<p>Multimodal recommender system that integrates with deep learning frameworks like\nTensorFlow, HuggingFace, or PyTorch.\n(<a href=\"https:\/\/colab.research.google.com\/drive\/1vPUALePlrjv4rfSn6CX2zMkpH2Xrw_cp?usp=sharing\">Demo<\/a>)\n[<a href=\"https:\/\/dl.acm.org\/doi\/pdf\/10.1145\/3589335.3651440\">Paper<\/a>]<\/p>\n<h5>The Web Data Commons <a href=\"http:\/\/schema.org\/\">Schema.org<\/a> Table Corpora<\/h5>\n<p>There are many table corpora, typically used to evaluate ML systems. The corpora\nuse different schemata and formats. Introducing <a href=\"http:\/\/schema.org\/\">schema.org<\/a> table corpora to\nbridge the gap. Use the Common Crawl corpus. Group by host (for example\n<a href=\"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/imdb.com\">imdb.com<\/a>) and class. Remove sparse entities and poor annotations.\nThe resulting tables can contain nested entities, for example, the actor type.\nNeeds flattening. Tables and meta statistics files are available for download.\nUsed for table annotation benchmarks, for Q&amp;A datasets, entity matching (based\non unique identifiers, like telephone numbers), or as a source for training\ndata. [<a href=\"https:\/\/dl.acm.org\/doi\/pdf\/10.1145\/3589335.3651441\">Paper<\/a>]<\/p>\n<h5>Tel2Veh: Fusion of Telecom Data and Vehicle Flow to Predict Camera-Free Traffic via a Spatio-Temporal Framework<\/h5>\n<p>Objective is to apply telecom data to improve traffic flow. Crossing telecom\ndata with vision-based camera data. Make a dataset available with the crossed\ndata. Based on this data, train a predictor using the vision-based data as the\ngroundtruth. System can be used for traffic flow monitoring and traffic\noptimization. [<a href=\"https:\/\/dl.acm.org\/doi\/pdf\/10.1145\/3589335.3651442\">Paper<\/a>]<\/p>\n<h5>An Open Platform for Quality Measures in a Linked Data Index<\/h5>\n<p>Finding a good dataset is a challenge. How can the quality be measured? Need\ndefined quality measures, Propose IndeGx, builds an index of public SPARQL\nendpoints. Used to compare FAIRness (Findable, Accessible, Interoperable,\nReusable) and accountability (traceability, transparency, trust). For dataset\ncreators, quality matters.\n[<a href=\"https:\/\/dl.acm.org\/doi\/pdf\/10.1145\/3589335.3651443\">Paper<\/a>]<\/p>\n<h5>CompMix: A Benchmark for Heterogeneous Question Answering<\/h5>\n<p>Heterogenous question answering systems where the answers come from different\nsources, like text or tables. <a href=\"https:\/\/qa.mpi-inf.mpg.de\/compmix\/\">CompMix<\/a> is a\ndataset of questions and answers. Covers comparatives, superlatives, ad-hoc,\ncount, ordinal questions, etc. Used generative LLMs and other methods to see if\nthey could answer the questions. None of the systems were able to answer 50% or\nmore of the questions, which means the questions dataset is really challenging.\nIdeally a system should ground the data in a source and make the answer\ntraceable. [<a href=\"https:\/\/dl.acm.org\/doi\/pdf\/10.1145\/3589335.3651444\">Paper<\/a>]<\/p>\n<h5>SE-PQA: Personalized Community Question Answering<\/h5>\n<p>Working on personalization in information retrieval. Used StackExchange\ncommunity questions and answers, tags, and user profile metadata. Trained\ndifferent models on the dataset. Personalization based on tags improved the\nquality. [<a href=\"https:\/\/dl.acm.org\/doi\/pdf\/10.1145\/3589335.3651445\">Paper<\/a>]<\/p>\n<h5>Can LLM Substitute Human Labeling? A Case Study of Fine-grained Chinese Address Entity Recognition Dataset for UAV Delivery<\/h5>\n<p>Drone delivery systems in China use named entity recognition to convert raw\naddresses into precise locations using large language models. Needs specifically\ntrained Chinese dataset. Released\n<a href=\"https:\/\/github.com\/zhhvvv\/CNER-UAV\">CNER-UAV dataset<\/a>. Contains lots of &quot;in the\nbuilding&quot; or &quot;in the unit complex&quot; or even &quot;in the room&quot; address refinements.\nTested with different models and evaluated precision and recall. ChatGLM (the\nChinese ChatGPT) performs poorly, GPT compares fine, but struggles with room and\nother address annotations.\n[<a href=\"https:\/\/dl.acm.org\/doi\/pdf\/10.1145\/3589335.3651446\">Paper<\/a>]<\/p>\n<h5>Graphameleon: Relational Learning and Anomaly Detection on Web Navigation Traces Captured as Knowledge Graphs<\/h5>\n<p><a href=\"https:\/\/github.com\/Orange-OpenSource\/graphameleon\">Graphameleon<\/a> is a browser\nextension to capture web navigation. Motivations like tracking one's carbon\nfootprint of a browsing session. Has a macro mode that captures the\nrequest\/response traffic and micro mode that also captures mouse clicks. Use the\n<a href=\"https:\/\/github.com\/ucoProject\/UCO\">UCO ontology<\/a>. Creates a knowledge graph and\na 3D graph visualization. Tracks website complexity with and without Firefox\nstrict or standard tracking prevention on. Another use case is to detect attacks\nlike XSS. [<a href=\"https:\/\/dl.acm.org\/doi\/pdf\/10.1145\/3589335.3651447\">Paper<\/a>]<\/p>\n<figure>\n    <img alt=\"Borealis trees at night.\" src=\"https:\/\/blog.tomayac.com\/images\/thewebconf2024--wkic6zfdjia.png\" \/>\n<figcaption><a href=\"https:\/\/www.gardensbythebay.com.sg\/en\/things-to-do\/calendar-of-events\/borealis.html\">Borealis<\/a> in Gardens by the Bay<\/figcaption>\n<\/figure>\n\n\t\t\t<p>\n\t\t\t\t<img alt=\"Thomas Steiner\" width=\"32\" height=\"32\" src=\"https:\/\/blog.tomayac.com\/feed.php?dl=https%3A%2F%2Fblog.tomayac.com%2F2024%2F05%2F22%2Fthe-web-conf-2024-singapore-trip-report%2F&dp=%2F2024%2F05%2F22%2Fthe-web-conf-2024-singapore-trip-report%2F&dt=The%20Web%20Conf%202024%2C%20Singapore%3A%20Trip%20report\" alt=\"\">\n\t\t\t\t<br\/>This post appeared first on <a href=\"https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/\">https:\/\/blog.tomayac.com\/2024\/05\/22\/the-web-conf-2024-singapore-trip-report\/<\/a>.\n\t\t\t<\/p>\n\t\t","protected":false},"excerpt":{"rendered":"<p>\t\t\tBackground \ud83d\udd17<\/p>\n<p>The Web Conference (formerly known as<br \/>\nWWW) is an international conference focused on<br \/>\nexploring the current state and the evolution of the Web through the lens of<br \/>\ndifferent scientific disciplines, including computing science, social sc&#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"intlwemo_wallet_address":"","footnotes":""},"categories":[],"tags":[],"class_list":["post-3517","post","type-post","status-publish","format-standard","hentry"],"_links":{"self":[{"href":"https:\/\/tomayac.com\/wordpress\/wp-json\/wp\/v2\/posts\/3517","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/tomayac.com\/wordpress\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/tomayac.com\/wordpress\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/tomayac.com\/wordpress\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/tomayac.com\/wordpress\/wp-json\/wp\/v2\/comments?post=3517"}],"version-history":[{"count":1,"href":"https:\/\/tomayac.com\/wordpress\/wp-json\/wp\/v2\/posts\/3517\/revisions"}],"predecessor-version":[{"id":3518,"href":"https:\/\/tomayac.com\/wordpress\/wp-json\/wp\/v2\/posts\/3517\/revisions\/3518"}],"wp:attachment":[{"href":"https:\/\/tomayac.com\/wordpress\/wp-json\/wp\/v2\/media?parent=3517"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/tomayac.com\/wordpress\/wp-json\/wp\/v2\/categories?post=3517"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/tomayac.com\/wordpress\/wp-json\/wp\/v2\/tags?post=3517"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}