Skip to main content

Wikipedia directory


Restricted access: How the internet can be used to promote reading and learning

2022-derksen.pdf: “Restricted access: How the internet can be used to promote reading and learning”⁠, Laura Derksen, Catherine Michaud-Leclerc, Pedro C. L. Souza (2022-03-01; ; similar):

  • We evaluate the impact of restricted internet access on education outcomes.
  • We randomize access to Wikipedia in Malawian secondary schools.
  • Online information appeals broadly to student interests.
  • Increased time spent reading leads to gains in English exam scores.
  • The internet can also improve Biology scores, especially for low achievers.

Can schools use the internet to promote reading and learning?

We provided Wikipedia access to randomly-selected students in Malawian boarding secondary schools [n = 301 vs n = 1,207 controls]. Students used the online resource broadly and intensively, and found it trustworthy, including for information about news and safe sex.

We find a 0.10σ impact on English exam scores, and a higher impact among low achievers (0.20σ). Students used Wikipedia to study Biology, and exam scores increased for low achievers (0.14σ).

Our results show that by restricting internet access to a source of engaging and accessible reading material, it is possible to encourage independent reading and affect educational outcomes.

[Keywords: Internet, information, education, development, reading, secondary school]

…Our experiment took place in 4 government boarding schools which serve students of mixed socioeconomic status⁠. Each school has ~500 students spread over 4 forms (grade levels). Government boarding schools are common in Malawi and across sub-Saharan Africa. They are more academically competitive than government day schools and most private schools (de Hoop 2010). However, even in these schools, many students do struggle academically. In particular, 1⁄4th of students had an English exam score below 50⁄100 in the year before the intervention. While government boarding schools attract good students, fees are not exorbitant.25 Indeed, according to our baseline survey, many students at our sample schools are of lower socioeconomic status: 42% do not have electricity at home, and 45% do not have running water. 1⁄3rd of students have at least one parent who did not complete primary school.

Boarding schools provide a controlled environment; students have no access to the internet outside of our intervention, allowing us to cleanly limit internet use to Wikipedia. At the time of the intervention, the school grounds had consistent 3G or 4G network coverage. However, students were not allowed to access the internet or use phones, even outside of class time, and being caught with a phone at school was grounds for suspension. Students sleep in dormitories, and are not permitted to leave the school grounds. In particular, they do not go home during the term, so those who do have home internet access cannot use it.26

We conducted a randomized experiment in government boarding schools in Malawi, a country with rapidly improving internet infrastructure, but where students have limited internet experience and no internet access at school. This setting allows us to isolate both treatment and control students from the broader internet. Students were allowed to use Wikipedia inside a classroom referred to as a digital library, using anonymous usernames. Students were aware that their browsing behavior was private, and that browsing histories could not be linked to individual students. The digital library was open evenings and weekends during one school year, and access was restricted to treated students. This design limits potential spillovers on English language skills and Biology exam scores. Students did not have any other internet access during term time.

…Students found the online material engaging, as evidenced by their frequent and broad use of Wikipedia. They spent, on average, 80 minutes per week online. Rather than relying on aggregate usage statistics, we observe individual browsing histories, which allows us to characterize demand for specific topics at the level of an individual. Each student browsed, on average, more than 800 different pages across a range of topics.

Students came to use and trust Wikipedia, particularly for topics which are important, prone to misinformation and often absent from school books, such as world news and safe sex. We find spikes in activity in the week surrounding world news events that occurred during the experiment. We also show that students with access to Wikipedia are able to find news information that control group students cannot. Young people are generally curious about sex, and we find that students spent 7% of their browsing time on topics related to sex and sexuality. While Wikipedia pages are informative, and access to accurate information about sex can be important (Dupas 2011; Kerwin, 2018; Derksen et al 2021), students may have browsed these pages not only for information but also as a form of entertainment. 1⁄3rd of the time spent browsing these topics overlapped with topics from the school syllabus, such as pregnancy and reproductive health. Students sought information on both news and sex and sexuality independently, without prompts or incentives.

“A Large-Scale Characterization of How Readers Browse Wikipedia”, Piccardi et al 2021

“A Large-Scale Characterization of How Readers Browse Wikipedia”⁠, Tiziano Piccardi, Martin Gerlach, Akhil Arora, Robert West (2021-12-22; ; backlinks; similar):

Despite the importance and pervasiveness of Wikipedia as one of the largest platforms for open knowledge, surprisingly little is known about how people navigate its content when seeking information. To bridge this gap, we present the first systematic large-scale analysis of how readers browse Wikipedia. Using billions of page requests from Wikipedia’s server logs, we measure how readers reach articles, how they transition between articles, and how these patterns combine into more complex navigation paths. We find that navigation behavior is characterized by highly diverse structures. Although most navigation paths are shallow, comprising a single pageload, there is much variety, and the depth and shape of paths vary systematically with topic, device type, and time of day. We show that Wikipedia navigation paths commonly mesh with external pages as part of a larger online ecosystem, and we describe how naturally occurring navigation paths are distinct from targeted navigation in lab-based settings. Our results further suggest that navigation is abandoned when readers reach low-quality pages. These findings not only help in identifying potential improvements to reader experience on Wikipedia, but also in better understanding of how people seek knowledge on the Web.

[If users load an average of 1.5 pages per session, and almost all the subsequent 0.5 page loads are by following internal wiki links (and only 6% by alternative navigation methods like search), and sessions terminate at low-quality pages mostly, how much reading or lack of reading is due to the presence or absence of wiki links?

I notice that there are still a lot of missing wiki links on articles (even proper noun ones which are dead obvious: eg. John Eyre×Osman II). From a reader’s perspective, the absence of a link is evidence that they shouldn’t bother searching and they should halt there if that was what they wanted. Quality is in considerable part just accuracy and comprehensiveness of wikilinking. (See also impact of banner ads⁠/​latency; “Wikipedia Matters”⁠, Hinnosaar et al 2019; “Science Is Shaped by Wikipedia: Evidence from a Randomized Control Trial”⁠, Thompson et al 2017.)

If an average page has, say, 50 wikilinks and the expectation of another page is ~0.5 or 50% of a page, then each individual wikilink would on average be worth 1% of a pageview and one’d expect a marginal gain of <1% for each additional wikilink added to that page. That sounds ludicrously valuable if the real value is even a tenth of that, because adding wikilinks has traditionally not been a major focus of WP tooling or bot operator cause area (compared to disambiguation or vandal fighting). Can the user tracking estimate the value more directly? One could also look at analyzing the effects of the various semi-auto and auto-linking bots as natural experiments on the logged traffic.]

Externalities in knowledge production: evidence from a randomized field experiment

2022-hinnosaar.pdf: “Externalities in knowledge production: evidence from a randomized field experiment”⁠, Marit Hinnosaar, Toomas Hinnosaar, Michael E. Kummer, Olga Slivko (2021-09-01; ):

Are there positive or negative externalities in knowledge production? We analyze whether current contributions to knowledge production increase or decrease the future growth of knowledge.

To assess this, we use a randomized field experiment that added content to some pages in Wikipedia while leaving similar pages unchanged. We compare subsequent content growth over the next 4 years between the treatment and control groups.

Our estimates allow us to rule out effects on 4-year growth of content length larger than 12%. We can also rule out effects on 4-year growth of content quality larger than 4 points, which is less than one-fifth of the size of the treatment itself. The treatment increased editing activity in the first 2 years, but most of these edits only modified the text added by the treatment.

Our results have implications for information seeding and incentivizing contributions. They imply that additional content may inspire future contributions in the short-term and medium-term but do not generate large externalities in the long term.

[Keywords: knowledge accumulation, user-generated content, Wikipedia, public goods provision, field experiment]

“Scarecrow: A Framework for Scrutinizing Machine Text”, Dou et al 2021

“Scarecrow: A Framework for Scrutinizing Machine Text”⁠, Yao Dou, Maxwell Forbes, Rik Koncel-Kedziorski, Noah A. Smith, Yejin Choi (2021-07-02; ⁠, ; backlinks; similar):

Modern neural text generation systems can produce remarkably fluent and grammatical texts. While earlier language models suffered from repetition and syntactic errors, the errors made by contemporary models are often semantic, narrative, or discourse failures.

To facilitate research of these complex error types, we introduce a new structured, crowdsourced error annotation schema called Scarecrow. The error categories used in Scarecrow—such as redundancy, commonsense errors, and incoherence—were identified by combining expert analysis with several pilot rounds of ontology-free crowd annotation to arrive at a schema which covers the error phenomena found in real machine generated text.

We use Scarecrow to collect 13k annotations of 1.3k human and machine generate paragraphs of English language news text, amounting to over 41k spans each labeled with its error category, severity, a natural language explanation, and antecedent span (where relevant). We collect annotations for text generated by state-of-the-art systems with varying known performance levels, from GPT-2-small through the largest GPT-3-175b. We isolate several factors for detailed analysis, including parameter count, training data, and decoding technique.

Our results show both expected and surprising differences across these settings. These findings demonstrate the value of Scarecrow annotations in the assessment of current and future text generation systems. We release our complete annotation toolkit and dataset at Github⁠.

Figure 2: Average portion of tokens annotated with each span type (y-axis) across models (x-axis), with 95% confidence intervals.
Figure 3: Average portion of tokens covered by span annotations, broken down by span type. All models, including GPT-3, use the same apples-to-apples decoding hyperparameters: top-p = 0.96, temperature = 1, and no frequency penalty. We scale each span by its token length, normalize by generation token lengths, and remove severity-1 Grammar and Usage errors (see §C).
Figure 4: Taking the average span coverage (Figure 3) and removing reader issues (Technical Jargon and Needs Google), we plot values and 95% confidence intervals for all models, including all decoding hyperparameters we tested for GPT-3. We find a surprisingly large change in annotated errors depending on the decoding setting used.
  1. Scaling pays off to improve Encyclopedic, Commonsense, and Incoherent errors (Figure 2).

    These error categories decrease with in-domain training (GROVER) and larger model size (GPT-3). Human text still shows the fewest of these kinds of errors.

  2. Scaling benefits plateau for Off-Prompt, Bad Math, and Grammar & Usage errors (Figure 2).

    These 3 error categories see a model plateau in error reduction when scaling to GPT-3. Of these error types, humans still commit fewer Off-Prompt (more: §6.1) and Grammar & Usage errors, but Bad Math appears saturated for our domain.

  3. Self-Contradiction and Redundant errors exhibit more complex scaling behavior (Figure 2).

    We roughly categorize these trends as rising and falling: increasing for medium or large-scale models, but dropping for human-authored text. Further analysis (§6.2, §6.3) reveals these more complex patterns are affected both by interactions with other error types, as well how errors are counted.

  4. Human-authored text produces the most reader issues (Figure 2–3).

    The Needs Google and Technical Jargon span categories both have a humans highest trend, and both fall under reader issues: problems that are not necessarily errors, but that still prevent full comprehension or factual verification of the text (more: §6.4).

    Furthermore, human-authored text is not free from error annotations (Figure 3). This can serve either as a control for baseline error rates (more: §6.6), or as a mechanism for critiquing human writing.

  5. Decoding hyperparameters have a huge impact (Figure).

    For the previous findings, we fix the sampling configuration for all models to an apples-to-apples setup for fair comparison: top-p = 0.96, (softmax) temperature = 1, and no frequency penalty (ie. word repetition penalty; defined precisely in §5.2, Equation 1). To study the effects of these decoding settings, we annotate text generated by GPT-3 using a variety of values for top-p and temperature, both with and without a frequency penalty.

    To our surprise, the decoding hyperparameters considerably affected error rates (more: §6.5). As seen in Figure 4, the worst sampling procedure for GPT-3 (argmax sampling with no frequency penalty) performed even worse than GPT-2 XL. But the best sampling procedure (surprisingly, also argmax sampling, but with a frequency penalty) produced text with as few apparent SCARECROW error spans as those authored by humans (more: §6.6).

…We notice that a greater portion of errors in human-authored text were due to artifacts present in the text-only format of the Common Crawl. For example, links to other articles or advertisements sometimes appear in the middle of an article’s text. While annotators were quick to mark these spans, they reflect errors in formatting, not in writing. We partition these errors separately and exclude them from the subsequent calculations. GPT-3’s generations also sometimes exhibited what appeared to be formatting errors due to training on web-scraped text, though more rarely. For example, some generations contained Which? after vague noun phrases, which appear to be learned from Wikipedia, where under-specified information is tagged by an editor with this word. For fairness, we removed these errors from GPT-3’s tally as well, though they were few enough we do not plot them separately.

“WIT: Wikipedia-based Image Text Dataset for Multimodal Multilingual Machine Learning”, Srinivasan et al 2021

“WIT: Wikipedia-based Image Text Dataset for Multimodal Multilingual Machine Learning”⁠, Krishna Srinivasan, Karthik Raman, Jiecao Chen, Michael Bendersky, Marc Najork (2021-03-02; ; similar):

The milestone improvements brought about by deep representation learning and pre-training techniques have led to large performance gains across downstream NLP, IR and Vision tasks. Multimodal modeling techniques aim to leverage large high-quality vision-linguistic datasets for learning complementary information (across image and text modalities). In this paper, we introduce the Wikipedia-based Image Text (WIT) Dataset (https:/​/​​google-research-datasets/​wit) to better facilitate multimodal, multilingual learning. WIT is composed of a curated set of 37.6 million entity rich image-text examples with 11.5 million unique images across 108 Wikipedia languages. Its size enables WIT to be used as a pretraining dataset for multimodal models, as we show when applied to downstream tasks such as image-text retrieval. WIT has four main and unique advantages. First, WIT is the largest multimodal dataset by the number of image-text examples by 3× (at the time of writing). Second, WIT is massively multilingual (first of its kind) with coverage over 100+ languages (each of which has at least 12K examples) and provides cross-lingual texts for many images. Third, WIT represents a more diverse set of concepts and real world entities relative to what previous datasets cover. Lastly, WIT provides a very challenging real-world test set, as we empirically illustrate using an image-text retrieval task as an example.

“The Pile: An 800GB Dataset of Diverse Text for Language Modeling”, Gao et al 2021

“The Pile: An 800GB Dataset of Diverse Text for Language Modeling”⁠, Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He et al (2021; ; backlinks; similar):

[torrent download] Recent work has demonstrated that increased training dataset diversity improves general cross-domain knowledge and downstream generalization capability for large-scale language models. With this in mind, we present the Pile: an 825 GiB English text corpus targeted at training large-scale language models.

The Pile is constructed from 22 diverse high-quality subsets—many of which derive from academic or professional sources. [Common Crawl⁠, PubMed Central, Bibliotik (Books3), OpenWebText2, arXiv, Github⁠, FreeLaw, Stack Exchange, USPTO Backgrounds, PubMed Abstracts, Gutenberg (PG-19), OpenSubtitles, English Wikipedia, DeepMind Mathematics, Ubuntu IRC, BookCorpus2, EuroParl, Hacker News⁠, YouTubeSubtitles, PhilPapers, NIH ExPorter, Enron Emails]

Our evaluation of the untuned performance of GPT-2 and GPT-3 on the Pile shows that these models struggle on many of its components, such as academic writing. Conversely, models trained on the Pile improve substantially over both Raw CC and CC-100 on all components of the Pile, while improving performance on downstream evaluations.

Through an in-depth exploratory analysis, we document potentially concerning aspects of the data for prospective users. We make publicly available the code used in its construction.

“The Intrepid Mother and Son Who Unraveled a Geographic Hoax: Atlas Obscura Had a Page for Something Called Moose Boulder, Until Fan Roger Dickey Called Us on It.”, Taub 2020

“The Intrepid Mother and Son Who Unraveled a Geographic Hoax: Atlas Obscura had a page for something called Moose Boulder, until fan Roger Dickey called us on it.”⁠, Matthew Taub (2020-03-10; backlinks; similar):

…What had brought them there, and into this rather dicey situation, was something called Moose Boulder, a kind of geological Matryoshka doll. Here’s what makes Moose Boulder special, from the outside in: Lake Superior is the world’s largest freshwater lake, and its largest island is Isle Royale, whose largest lake is called Siskiwit, whose largest island is called Ryan. According to Wikipedia, at least, Ryan Island is home to a seasonal pond called Moose Flats that, when flooded, contains its own island—Moose Boulder. This makes it “the largest island in the largest lake on the largest island in the largest lake on the largest island in the largest lake in the world.” Spoiler: Mother and son made it out alive, but it wasn’t because they stumbled on a geological/​hydrological anomaly that they could use to get their bearings. They couldn’t have, because, despite what the internet has to say, Moose Boulder almost surely doesn’t exist.

…It’s doubtful that any of these other hikers, however, had consulted Atlas Obscura. Had they done so, Dickey soon realized, they would have found the precise coordinates: 48.0088°, −88.7720°. They would have seen that some people had marked visiting it on their Atlas Obscura profiles. Dickey had to get creative to actually contact these people. “I did Google reverse image search for their profile photos”, he says, which led him to two people with social media presences, neither of whom responded to his messages…Naturally, as they all do, the Atlas Obscura entry for the site had an image—albeit a grainy one—of a lonely little rock, cautiously jutting out of the water, feebly sprouting some weeds…Many photos that users upload to Atlas Obscura link to their original sources, but this one was a dead end. Using the Wayback Machine, Dickey found that it had come from a defunct website that appeared to document a geological research expedition to Ryan Island…The supposed photographic evidence had indeed come from that expedition, but it was merely a photo of an ordinary rock, off the coast of Isle Royale itself and not of the Inception of islands deep inside it.

By now, the odds seemed overwhelming to Dickey that Moose Boulder was a myth, a spasm of the Internet’s imagination that had managed to proliferate and live on. But still, something didn’t quite add up. There was a missing piece to the puzzle that stopped Dickey short of declaring it all a hoax. He had found another article about Moose Boulder, published in 2009, that cited Wikipedia as its source of information. But the information about Moose Boulder had been added to Siskiwit Lake’s Wikipedia page in 2012. It was like a scene in a bad horror movie in which someone gets a phone call from a dead person. Dickey joked with his girlfriend that perhaps Moose Boulder does exist, but only in some kind of “temporal anomaly.”…Here’s the rub: Wikipedia is a nesting doll, too. Before a page for Siskiwit Lake had been added to the site, the page for Isle Royale had pointed readers to Moose Boulder, and had been doing so since 2009. It was put there by a different user than the one who added it to the Siskiwit page in 2012. Either way, that’s where the trail goes cold, and there’s no other evidence that the place exists. The identity of that first Wikipedia user to write about it—with those completely unrelated sources—remains a mystery, but all available evidence suggests that it was a person having a laugh, nothing more.

“REALM: Retrieval-Augmented Language Model Pre-Training”, Guu et al 2020

“REALM: Retrieval-Augmented Language Model Pre-Training”⁠, Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, Ming-Wei Chang (2020-02-10; ; backlinks; similar):

Language model pre-training has been shown to capture a surprising amount of world knowledge, crucial for NLP tasks such as question answering. However, this knowledge is stored implicitly in the parameters of a neural network, requiring ever-larger networks to cover more facts.

To capture knowledge in a more modular and interpretable way, we augment language model pre-training with a latent knowledge retriever, which allows the model to retrieve and attend over documents from a large corpus such as Wikipedia, used during pre-training, fine-tuning and inference. For the first time, we show how to pre-train such a knowledge retriever in an unsupervised manner, using masked language modeling as the learning signal and backpropagating through a retrieval step that considers millions of documents.

We demonstrate the effectiveness of Retrieval Augmented Language Model pre-training (REALM) by fine-tuning on the challenging task of Open-domain Question Answering (Open-QA).We compare against state-of-the-art models for both explicit and implicit knowledge storage on three popular Open-QA benchmarks, and find that we outperform all previous methods by a substantial margin (4–16% absolute accuracy), while also providing qualitative benefits such as interpretability and modularity.


popups.js”, Achmiz 2019

popups.js: popups.js⁠, Said Achmiz (2019-08-21; ⁠, ; backlinks; similar):

popups.js: standalone Javascript library for creating ‘popups’ which display link metadata (typically, title/​author/​date/​summary), for extremely convenient reference/​abstract reading, with mobile and YouTube support. Whenever any such link is mouse-overed by the user, popups.js will pop up a large tooltip-like square with the contents of the attributes. This is particularly intended for references, where it is extremely convenient to autopopulate links such as to​​Pubmed/​PLOS/​​Wikipedia with the link’s title/​author/​date/​abstract, so the reader can see it instantly.

popups.js parses a HTML document and looks for <a> links which have the link-annotated attribute class, and the attributes data-popup-title, data-popup-author, data-popup-date, data-popup-doi, data-popup-abstract. (These attributes are expected to be populated already by the HTML document’s compiler, however, they can also be done dynamically. See wikipedia-popups.js for an example of a library which does Wikipedia-only dynamically on page loads.)

For an example of a Hakyll library which generates annotations for Wikipedia/​Biorxiv/​Arxiv⁠/​PDFs/​arbitrarily-defined links, see LinkMetadata.hs⁠.


wikipedia-popups.js: wikipedia-popups.js⁠, Said Achmiz (2019-07-29; backlinks; similar):

wikipedia-popups.js: standalone Javascript library for creating ‘popups’ for links to English Wikipedia articles when the user mouse-overs the link. The tooltip-style popup displays the summaries/​introductions/​ledes to Wikipedia articles as returned by the Wikipedia API (see and All summaries are loaded on page load so as to have minimal latency (on-mouseover summary loading is noticeably slow). If a page has many Wikipedia links on it, this can result in quite a few requests; the summaries can instead be provided statically, encoded into data attributes. (This also allows encoding summaries/​previews of arbitrary websites by whatever is compiling the HTML.) See /static/js/popups.js for a JS library which takes that approach instead.

Wikipedia Matters

“Wikipedia Matters”⁠, Marit Hinnosaar, Toomas Hinnosaar, Michael Kummer, Olga Slivko (2019-07-14; ; backlinks; similar):

We document a causal impact of online user-generated information on real-world economic outcomes. In particular, we conduct a randomized field experiment to test whether additional content on Wikipedia pages about cities affects tourists’ choices of overnight visits. Our treatment of adding information to Wikipedia increases overnight stays in treated cities compared to non-treated cities. The impact is largely driven by improvements to shorter and relatively incomplete pages on Wikipedia. Our findings highlight the value of content in digital public goods for informing individual choices.

[Keywords: field experiment, user-generated content, Wikipedia, tourism industry]

The wisdom of polarized crowds

2019-shi.pdf: “The wisdom of polarized crowds”⁠, Feng Shi (2019-03-04; similar):

As political polarization in the United States continues to rise, the question of whether polarized individuals can fruitfully cooperate becomes pressing. Although diverse perspectives typically lead to superior team performance on complex tasks, strong political perspectives have been associated with conflict, misinformation and a reluctance to engage with people and ideas beyond one’s echo chamber.

Here, we explore the effect of ideological composition on team performance by analysing millions of edits to Wikipedia’s political, social issues and science articles. We measure editors’ online ideological preferences by how much they contribute to conservative versus liberal articles. Editor surveys suggest that online contributions associate with offline political party affiliation and ideological self-identity.

Our analysis reveals that polarized teams consisting of a balanced set of ideologically diverse editors produce articles of a higher quality than homogeneous teams. The effect is most clearly seen in Wikipedia’s political articles, but also in social issues and even science articles. Analysis of article ‘talk pages’ reveals that ideologically polarized teams engage in longer, more constructive, competitive and substantively focused but linguistically diverse debates than teams of ideological moderates.

More intense use of Wikipedia policies by ideologically diverse teams suggests institutional design principles to help unleash the power of polarization.

“Inside the Secret Sting Operations to Expose Celebrity Psychics: Are Some Celebrity Mediums Fooling Their Audience Members by Reading Social Media Pages in Advance? A Group of Online Vigilantes Is out to Prove It”, Hitt 2019

“Inside the Secret Sting Operations to Expose Celebrity Psychics: Are some celebrity mediums fooling their audience members by reading social media pages in advance? A group of online vigilantes is out to prove it”⁠, Jack Hitt (2019-02-26; similar):

…Collectively, the group, which has swelled to 144 members, has researched, written or revised almost 900 Wikipedia pages. Sure, they take on the classics, like debunking “spontaneous human combustion”, but many of their other pages have real-world impact. For instance, they straightened out a lot of grim hooey about the teen-suicide myth “blue whale game”⁠, and they have provided facts about the Burzynski Clinic⁠, a theoretical treatment for cancer operating out of Houston.

Most recently, Gerbic’s members have focused on what they call “grief vampires”, that is, the kind of middlebrow psychics who profit by claiming to summon the dead in shows in venues ranging from casinos or any old Motel 6 conference suite to wine vineyards or the Queen Mary permanently anchored in Long Beach. Some regional favorites may sound familiar—Theresa Caputo, the Long Island medium; or Chip Coffey, the “clairvoyant, clairaudient and clairsentient” psychic.

These are good, extremely profitable days for the ectoplasm-related industry. According to one market analysis, there are nearly 95,000 psychic “businesses” in America, generating some $2 billion in revenue in 2018. Lately, technology has changed the business of talking to the dead and created new kinds of openings for psychics to lure customers but also new ways for skeptics to flip that technology right back at them.

For instance, many psychics still rely on “cold readings”⁠, in which the psychic uses clues, like your clothes or subtle body signals, to make educated, but generally vague, guesses about your life and family. But the internet has popularized a new kind of “hot reading”, in which the psychics come to their shows prepped with specific details about various members of the audience. One new source of psychic intel is Facebook, which has become a clearinghouse for the kind of insider, personal detail that psychics used to have to really sweat for. If anything, “the psychics have just gotten lazier”, a team member told me.

…These Guerrilla Skeptics are hoping to catch Fraser on tape spewing intimate Facebook details that are totally false about the person the psychic is addressing. In fact, the details aren’t true about anyone, because they will be entirely fabricated by people like Michelle of Humpty Doo. At this stage, on this Skype call, the group’s only task is to create and maintain these fake Facebook profiles. These need to look normal—with regular updates of, say, a good New Yorker cartoon or a gif of Will Ferrell dancing, along with vintage Polaroid pics or posts expressing sly life sentiments. (“Still a little pissed I can’t fly or set things on fire with my mind!”)…Once the psychic has been stung, the team will write up an account and then post the evidence—video or sound—onto a website dedicated to a particular debunking mission, which Gerbic gives a memorable name. In future events, other skeptics can simply slip into performances and just leave cards with these odd operation names printed on them.

…Gerbic told me that the group’s previous hot-read sting—Operation Pizza Roll—worked perfectly back in 2017. She and the other skeptics spent 10 days creating Facebook profiles in advance of Thomas John’s visit to southern Los Angeles. Gerbic used her Facebook sock puppets, “Susanna and Mark Wilson”, to register herself and her pal Edward.

John is a well-known figure on the psychic circuit. He names people’s pets and dead relatives with breathtaking first-attempt accuracy. He has a thriving practice on Madison Avenue, and on the West Coast, his press materials tout a host of Hollywood clients, including Sam Smith, Courteney Cox and Julianne Moore. His audiences admire him, but then they probably haven’t Googled past the first page of results to learn that before he popularized his gift for talking to the dead, he was Lady Vera Parker, a drag queen in Chicago who later got into some trouble when Thomas John Flanagan (his legal name) was charged with theft, fined and sentenced to probation—precisely what the specific charge was for, his lawyer explained in a statement, the psychic can no longer remember.

On the appointed night of the show, in came Susanna and Mark Wilson, dressed in fancy clothes and toting third-row VIP tickets and unobtrusive recording equipment. Because Susanna’s Facebook page mentioned her losing her twin brother, Andrew, to pancreatic cancer, Gerbic arrived clutching a handful of tissues, a tactic she encourages because it sends the psychic the message that you will be an emotional and entertaining reading. Right away, Thomas John said he was tuning in to a twin brother who wanted to speak to his sister. Gerbic raised her hand.

“Somebody is making me aware of cancer?” John asked, and Gerbic choked up, yes, yes. John reeled her in: “I’m getting something right in here”, and pointed to his abdomen, “stomach or pancreas?” Gerbic acted emotional. And John went straight down the rabbit hole, all the while being careful not to bring the crowd down. He said of Gerbic’s fictional dead brother: “First off, he is making fun of you, teasing you for being here with me! He’s laughing about it!” And the audience laughed, too.

Over the course of the reading, John comfortably laid down the specifics of Susanna Wilson’s life—he named “Andy” and amazingly knew him to be her twin. He knew that she and her brother grew up in Michigan and that his girlfriend was Maria. He knew about Susanna’s father-in-law and how he died.

But about 2⁄3rds of the way through John’s riffing, he seemed to sense something was fishy. All of which is, in fact, part of the experiment. Gerbic knows only some of the facts of her character’s life. Her thinking is that if John knows even more details than she does, then it’s absolute proof that he’s looked through the Facebook posts. Gerbic’s sting is placebo-controlled, double-blind. On the tape, it’s easy to catch the precise moment when John sensed that something was wrong. John was talking about the dead brother when he suddenly asked, “And ‘Buddy,’ who is that?”

Gerbic had no idea and improvised, “my father”, when in fact, Buddy was her fictional dead brother’s fictional dog. John kept up the reading and then interrupted himself: “Oh, I understand—OK, so I am being drawn over here”, and with that, he walked away.

“What Is the Commons Worth? Estimating the Value of Wikimedia Imagery by Observing Downstream Use”, Erickson et al 2018

2018-erickson.pdf: “What is the Commons Worth? Estimating the Value of Wikimedia Imagery by Observing Downstream Use”⁠, Kristofer Erickson, Felix Rodriguez Perez, Jesus Rodriguez Perez (2018-08-22; ; similar):

The Wikimedia Commons (WC) is a peer-produced repository of freely licensed images, videos, sounds and interactive media, containing more than 45 million files. This paper attempts to quantify the societal value of the WC by tracking the downstream use of images found on the platform.

We take a random sample of 10,000 images from WC and apply an automated reverse-image search to each, recording when and where they are used ‘in the wild’. We detect 54,758 downstream uses of the initial sample, and we characterise these at the level of generic and country-code top-level domains (TLDs). We analyse the impact of specific variables on the odds that an image is used. The random sampling technique enables us to estimate overall value of all images contained on the platform.

Drawing on the method employed by Heald et al 2015⁠, we find a potential contribution of $28.9 billion from downstream use of Wikimedia Commons images over the lifetime of the project.

…We find an overall quantity of 54,758 downstream uses of images from our sample. We estimate a series of logistic regressions to study variables that are statistically-significant in the odds of uptake of WC images. Overall, we find that license type is a statistically-significant factor in whether or not an image is used outside of the WC. Public domain files and licenses (those without attribution or share-alike clauses) are associated with increased odds of downstream use. This is consistent with other economic studies of the public domain ([2] [6]). We also find that for commercial use, prior appearance of the file elsewhere on Wikipedia has a statistically-significant positive effect, suggesting that human curation and selection are important in promoting key images to widespread use. We suggest further experimentation using a purposive sample of ‘quality’ and ‘valued’ images to test for the impact of human curation on the WC.

…This paper has tracked downstream digital use of images hosted on the WC. We find a mean rate of online use of 5.48 uses per image. Using commercial TLDs as a proxy for commercial use, we estimate a mean commercial usage of 2.99 per image. The odds that a given image from the WC will be used is statistically-significantly influenced by the license type issued by its uploader. Images with attribution and share-alike licenses have statistically-significantly reduced odds of being used externally compared to images fully in the public domain.

The actual societal value of the WC is likely considerably greater, and would include direct personal uses as well as print, educational and embedded software applications not detectable by our reverse image search technique. Getty routinely charges license fees of $650 or more for creative use (such as magazine covers), considerably higher than the rate for editorial use. Our valuation method could be improved with more information about usage rates of commercial stock photography as well as potential qualitative differences between stock and Commons-produced imagery.

“Generating Wikipedia by Summarizing Long Sequences”, Liu et al 2018

“Generating Wikipedia by Summarizing Long Sequences”⁠, Peter J. Liu, Mohammad Saleh, Etienne Pot, Ben Goodrich, Ryan Sepassi, Lukasz Kaiser, Noam Shazeer (2018-01-30; ; backlinks; similar):

We show that generating English Wikipedia articles can be approached as a multi-document summarization of source documents. We use extractive summarization to coarsely identify salient information and a neural abstractive model to generate the article. For the abstractive model, we introduce a decoder-only architecture that can scalably attend to very long sequences, much longer than typical encoder-decoder architectures used in sequence transduction. We show that this model can generate fluent, coherent multi-sentence paragraphs and even whole Wikipedia articles. When given reference documents, we show it can extract relevant factual information as reflected in perplexity, ROUGE scores and human evaluations.

Revisiting “The Rise and Decline” in a Population of Peer Production Projects


“Learning to Organize Knowledge and Answer Questions With N-Gram Machines”, Yang et al 2017

“Learning to Organize Knowledge and Answer Questions with N-Gram Machines”⁠, Fan Yang, Jiazhong Nie, William W. Cohen, Ni Lao (2017-11-17; ; backlinks; similar):

Though deep neural networks have great success in natural language processing, they are limited at more knowledge intensive AI tasks, such as open-domain Question Answering (QA). Existing end-to-end deep QA models need to process the entire text after observing the question, and therefore their complexity in responding a question is linear in the text size. This is prohibitive for practical tasks such as QA from Wikipedia, a novel, or the Web.

We propose to solve this scalability issue by using symbolic meaning representations, which can be indexed and retrieved efficiently with complexity that is independent of the text size.

We apply our approach, called the N-Gram Machine (NGM), to three representative tasks. First as proof-of-concept, we demonstrate that NGM successfully solves the bAbI tasks of synthetic text. Second, we show that NGM scales to large corpus by experimenting on “life-long bAbI”, a special version of bAbI that contains millions of sentences. Lastly on the WikiMovies dataset, we use NGM to induce latent structure (ie. schema) and answer questions from natural language Wikipedia text, with only QA pairs as weak supervision.

“Seq2SQL: Generating Structured Queries from Natural Language Using Reinforcement Learning”, Zhong et al 2017

“Seq2SQL: Generating Structured Queries from Natural Language using Reinforcement Learning”⁠, Victor Zhong, Caiming Xiong, Richard Socher (2017-08-31; ⁠, ; backlinks; similar):

A significant amount of the world’s knowledge is stored in relational databases. However, the ability for users to retrieve facts from a database is limited due to a lack of understanding of query languages such as SQL. We propose Seq2SQL, a deep neural network for translating natural language questions to corresponding SQL queries. Our model leverages the structure of SQL queries to significantly reduce the output space of generated queries. Moreover, we use rewards from in-the-loop query execution over the database to learn a policy to generate unordered parts of the query, which we show are less suitable for optimization via cross entropy loss. In addition, we will publish WikiSQL, a dataset of 80,654 hand-annotated examples of questions and SQL queries distributed across 24,241 tables from Wikipedia. This dataset is required to train our model and is an order of magnitude larger than comparable datasets. By applying policy-based reinforcement learning with a query execution environment to WikiSQL, our model Seq2SQL outperforms attentional sequence to sequence models, improving execution accuracy from 35.9% to 59.4% and logical form accuracy from 23.4% to 48.3%.

“Does Copyright Affect Reuse? Evidence from Google Books and Wikipedia”, Nagaraj 2017

2017-nagaraj.pdf: “Does Copyright Affect Reuse? Evidence from Google Books and Wikipedia”⁠, Abhishek Nagaraj (2017-07-26; ; similar):

While digitization has greatly increased the reuse of knowledge, this study shows how these benefits might be mitigated by copyright restrictions. I use the digitization of in-copyright and out-of-copyright issues of Baseball Digest magazine by Google Books to measure the impact of copyright on knowledge reuse in Wikipedia. I exploit a feature of the 1909 Copyright Act whereby material published before 1964 has lapsed into the public domain, allowing for the causal estimation of the impact of copyright across this sharp cutoff. I find that, while digitization encourages knowledge reuse, copyright restrictions reduce citations to copyrighted issues of Baseball Digest by up to 135% and affect readership by reducing traffic to affected pages by 20%. These impacts are highly uneven: copyright hurts the reuse of images rather than text and affects Wikipedia pages for less-popular players greater than more-popular ones.

The online appendix is available⁠.

“The Valuation of Unprotected Works: A Case Study of Public Domain Images on Wikipedia”, Heald et al 2015

2015-heald.pdf: “The Valuation of Unprotected Works: A Case Study of Public Domain Images on Wikipedia”⁠, Paul Heald, Kristofer Erickson, Martin Kretschmer (2015-02-06; ; backlinks; similar):

What is the value of works in the public domain?

We study the biographical Wikipedia pages of a large data set of authors, composers, and lyricists to determine whether the public domain status of available images leads to a higher rate of inclusion of illustrated supplementary material and whether such inclusion increases visitorship to individual pages. We attempt to objectively place a value on the body of public domain photographs and illustrations which are used in this global resource.

We find that the most historically remote subjects are more likely to have images on their web pages because their biographical life-spans pre-date the existence of in-copyright imagery. We find that the large majority of photos and illustrations used on subject pages were obtained from the public domain, and we estimate their value in terms of costs saved to Wikipedia page builders and in terms of increased traffic corresponding to the inclusion of an image.

Then, extrapolating from the characteristics of a random sample of a further 300 Wikipedia pages, we estimate a total value of public domain photographs on Wikipedia of between $310.0$246.02015 to $340.2$270.02015 million dollars per year.

[Keywords: public domain, copyright, valuation, econometrics, Wikipedia, photographs, composers, lyricists, value]

…In the absence of established market prices, valuation is always the domain of estimation and proxies. This is especially true of intellectual property in copyrights and patents, where works are original or novel by definition. Nevertheless, the exercise of quantifying the value of legal rights, and the value of the absence of legal rights, illuminates issues for policymakers even when precise numbers cannot be put on consumer surplus and overall social welfare. Our study demonstrates that the value of the public domain can be estimated at least as precisely as the commercial value of copyrights. Even though our estimates make use of several proxies, implications for both copyright term extension and orphan works legislation are substantial. The time has come for the Copyright Office and the U.S. Congress to endorse an evidence-based regime for the federal management of creative works.

“Impact of Wikipedia on Market Information Environment: Evidence on Management Disclosure and Investor Reaction”, Xu & Zhang 2013

2013-xu.pdf: “Impact of Wikipedia on Market Information Environment: Evidence on Management Disclosure and Investor Reaction”⁠, Sean Xin Xu, Xiaoquan Michael Zhang (2013-12-01; ; similar):

In this paper, we seek to determine whether a typical social media platform, Wikipedia, improves the information environment for investors in the financial market. Our theoretical lens leads us to expect that information aggregation about public companies on Wikipedia may influence how management’s voluntary information disclosure reacts to market uncertainty with respect to investors’ information about these companies.

Our empirical analysis is based on an unique data set collected from financial records, management disclosure records, news article coverage, and a Wikipedia modification history of public companies.

On the supply side of information, we find that information aggregation on Wikipedia can moderate the timing of managers’ voluntary disclosure of companies’ earnings disappointments, or bad news. On the demand side of information, we find that Wikipedia’s information aggregation moderates investors’ negative reaction to bad news.

Taken together, these findings support the view that Wikipedia improves the information environment in the financial market and underscore the value of information aggregation through the use of information technology.

[Keywords: Social media, Wikipedia, information environment, financial market, management disclosure, information aggregation]

“Predicting Google Closures”, Branwen 2013

Google-shutdowns: “Predicting Google closures”⁠, Gwern Branwen (2013-03-28; ⁠, ⁠, ⁠, ⁠, ; backlinks; similar):

Analyzing predictors of Google abandoning products; predicting future shutdowns

Prompted by the shutdown of Google Reader⁠, I ponder the evanescence of online services and wonder what is the risk of them disappearing. I collect data on 350 Google products launched before March 2013, looking for variables predictive of mortality (web hits, service vs software, commercial vs free, FLOSS, social networking, and internal vs acquired). Shutdowns are unevenly distributed over the calendar year or Google’s history. I use logistic regression & survival analysis (which can deal with right-censorship) to model the risk of shutdown over time and examine correlates. The logistic regression indicates socialness, acquisitions, and lack of web hits predict being shut down, but the results may not be right. The survival analysis finds a median lifespan of 2824 days with a roughly Type III survival curve (high early-life mortality); a Cox regression finds similar results as the logistic - socialness, free, acquisition, and long life predict lower mortality. Using the best model, I make predictions about probability of shutdown of the most risky and least risky services in the next 5 years (up to March 2018). (All data & R source code is provided.)

“Aesthetics of Destruction: Music and the Worldview of Shinji Ikari in Neon Genesis Evangelion”, Hoffer 2012

2012-hoffer.pdf: “Aesthetics of Destruction: Music and the Worldview of Shinji Ikari in Neon Genesis Evangelion⁠, Heike Hoffer (2012; ⁠, ; backlinks; similar):

Director Anno Hideaki’s series Neon Genesis Evangelion caused a sensation when it first aired on TV Tokyo in 1995 and has become one of the most influential anime ever made. Since its premiere, fans across the globe have debated the possible interpretations of the complex plot, but little has been said about how composer Sagisu Shiro’s score might contribute to understanding the series. Anno’s rehabilitation in a Jungian clinic and subsequent personal study of human psychology plays heavily into understanding the main character Ikari Shinji, and music has much to contribute to appreciating Shinji’s view of the world. Shinji is an impressionable fourteen-year old boy, so his musical interpretations of the people and things around him do not always match reality. Sagisu’s music gives the viewers welcome insight into Shinji’s thoughts and feelings as he matures throughout the series.

“Wikimedia UK Board Meeting, London”, Gardner 2011

“Wikimedia UK Board Meeting, London”⁠, Sue Gardner (2011-11-19; ; similar):

It’s getting harder for new people to join our projects. Newbies are making up a smaller percentage of editors overall than ever before, and the absolute number of newbies is dropping as well. Wikimedia needs to attract and retain more new and diverse editors, and to retain our experienced editors. A stable editing community is critical to the long-term sustainability and quality of both our current projects and our movement. We consider meeting this challenge our top priority.

“Circadian Patterns of Wikipedia Editorial Activity: A Demographic Analysis”, Yasseri et al 2011

“Circadian patterns of Wikipedia editorial activity: A demographic analysis”⁠, Taha Yasseri, Róbert Sumi, János Kertész (2011-09-08; ; backlinks; similar):

Wikipedia (WP) as a collaborative, dynamical system of humans is an appropriate subject of social studies. Each single action of the members of this society, i.e. editors, is well recorded and accessible. Using the cumulative data of 34 Wikipedias in different languages, we try to characterize and find the universalities and differences in temporal activity patterns of editors. Based on this data, we estimate the geographical distribution of editors for each WP in the globe. Furthermore we also clarify the differences among different groups of WPs, which originate in the variance of cultural and social features of the communities of editors.

“Emergence of Good Conduct, Scaling and Zipf Laws in Human Behavioral Sequences in an Online World”, Thurner et al 2011

“Emergence of good conduct, scaling and Zipf laws in human behavioral sequences in an online world”⁠, Stefan Thurner, Michel Szell, Roberta Sinatra (2011-07-02; backlinks; similar):

We study behavioral action sequences of players in a massive multiplayer online game. In their virtual life players use eight basic actions which allow them to interact with each other. These actions are communication, trade, establishing or breaking friendships and enmities, attack, and punishment. We measure the probabilities for these actions conditional on previous taken and received actions and find a dramatic increase of negative behavior immediately after receiving negative actions. Similarly, positive behavior is intensified by receiving positive actions. We observe a tendency towards anti-persistence in communication sequences. Classifying actions as positive (good) and negative (bad) allows us to define binary ‘world lines’ of lives of individuals. Positive and negative actions are persistent and occur in clusters, indicated by large scaling exponents alpha 0.87 of the mean square displacement of the world lines. For all eight action types we find strong signs for high levels of repetitiveness, especially for negative actions. We partition behavioral sequences into segments of length n (behavioral ‘words’ and ‘motifs’) and study their statistical properties. We find two approximate power laws in the word ranking distribution, one with an exponent of kappa-1 for the ranks up to 100, and another with a lower exponent for higher ranks. The Shannon n-tuple redundancy yields large values and increases in terms of word length, further underscoring the non-trivial statistical properties of behavioral sequences. On the collective, societal level the timeseries of particular actions per day can be understood by a simple mean-reverting log-normal model.

“Nootropics”, Branwen 2010

Nootropics: “Nootropics”⁠, Gwern Branwen (2010-01-02; ⁠, ⁠, ⁠, ⁠, ⁠, ⁠, ⁠, ⁠, ⁠, ⁠, ⁠, ⁠, ⁠, ⁠, ⁠, ⁠, ; backlinks; similar)

“Writing a Wikipedia RSS Link Archive Bot”, Branwen 2009

Wikipedia-RSS-Archive-Bot: “Writing a Wikipedia RSS Link Archive Bot”⁠, Gwern Branwen (2009-11-02; ⁠, ⁠, ; backlinks; similar):

Archiving using Wikipedia Recent Changes RSS feed (obsolete).

Continuation of the 2009 Haskell Wikipedia link archiving bot tutorial, extending it from operating on a pre-specified list of articles to instead archiving links live by using TagSoup parsing Wikipedia Recent Changes for newly-added external links which can be archived using WebCite in parallel. (Note: these tutorials are obsolete. WebCite is largely defunct, doing archiving this way is not advised, and WP link archiving is currently handled by Internet Archive-specific plugins by the WMF. For a more general approach suitable for personal use, see the writeup of archiver-bot in Archiving URLs⁠.)

“Beware Trivial Inconveniences”, Alexander 2009

“Beware Trivial Inconveniences”⁠, Scott Alexander (2009-05-06; ⁠, ; backlinks; similar):

The Great Firewall of China. A massive system of centralized censorship purging the Chinese version of the Internet of all potentially subversive content. Generally agreed to be a great technical achievement and political success even by the vast majority of people who find it morally abhorrent. I spent a few days in China. I got around it at the Internet cafe by using a free online proxy. Actual Chinese people have dozens of ways of getting around it with a minimum of technical knowledge or just the ability to read some instructions.

The Chinese government isn’t losing any sleep over this (although they also don’t lose any sleep over murdering political dissidents, so maybe they’re just very sound sleepers). Their theory is that by making it a little inconvenient and time-consuming to view subversive sites, they will discourage casual exploration. No one will bother to circumvent it unless they already seriously distrust the Chinese government and are specifically looking for foreign websites, and these people probably know what the foreign websites are going to say anyway.

Think about this for a second. The human longing for freedom of information is a terrible and wonderful thing. It delineates a pivotal difference between mental emancipation and slavery. It has launched protests, rebellions, and revolutions. Thousands have devoted their lives to it, thousands of others have even died for it. And it can be stopped dead in its tracks by requiring people to search for “how to set up proxy” before viewing their anti-government website.

…But these trivial inconveniences have major policy implications. Countries like China that want to oppress their citizens are already using “soft” oppression to make it annoyingly difficult to access subversive information. But there are also benefits for governments that want to help their citizens.

“Generalizing From One Example”, Alexander 2009

“Generalizing From One Example”⁠, Scott Alexander (2009-04-28; ⁠, ⁠, ⁠, ; backlinks; similar):

[Alexander defines the “typical mind fallacy”: everyone reasons about their mental experiences as if they are universal. People with vivid visual imagery assume everyone can see things in “the mind’s eye” while ‘aphantasics’ assume that this is simply a poetic metaphor; people with color-blindness wonder why other people get so worked up about various shades of gray, and people with anosmia are puzzled by the focus on flowers etc. Further examples include maladaptive daydreaming, pain insensitivity, the prevalence of visual & auditory hallucinations in mentally-healthy individuals like ‘scintillating scotoma’, misophonia⁠, hearing voices⁠, inner monologues, facial self-awareness, trypophobia⁠, Severely Deficient Autobiographical Memory⁠, hypermnesia, ASMR⁠, face blindness/​prosopagnosia⁠, musical anhedonia⁠, ‘the call of the void’/​intrusive thoughts⁠, hypnagogia⁠, the nasal dilation cycle

This phenomenon for visual imagery was discovered only recently by Francis Galton⁠, who asked if the interminable debate between philosophers/​psychologists like Berkeley or Behaviorists like Skinner, where neither could accept that there was (or was not) visual imagery, was because both were right—some people have extremely vivid mental imagery, while others have none at all. He simply circulated a survey and asked. Turned out, most people do but some don’t.

The typical mind fallacy may explain many interpersonal conflicts and differences in advice: we underappreciate the sheer cognitive diversity of mankind, because we only have access to our limited personal anecdote, and people typically do not discuss all their differences because they don’t realize they exist nor have a vocabulary/​name.]

“Dual N-Back FAQ”, Branwen 2009

DNB-FAQ: “Dual n-Back FAQ”⁠, Gwern Branwen (2009-03-25; ⁠, ⁠, ⁠, ⁠, ⁠, ⁠, ⁠, ⁠, ⁠, ⁠, ⁠, ; backlinks; similar):

A compendium of DNB, WM⁠, IQ information up to 2015.

Between 2008 and 2011, I collected a number of anecdotal reports about the effects of n-backing; there are many other anecdotes out there, but the following are a good representation—for what they’re worth.

“In Defense of Inclusionism”, Branwen 2009

In-Defense-Of-Inclusionism: “In Defense of Inclusionism”⁠, Gwern Branwen (2009-01-15; ⁠, ⁠, ⁠, ⁠, ⁠, ⁠, ⁠, ⁠, ; backlinks; similar):

Iron Law of Bureaucracy: the downwards deletionism spiral discourages contribution and is how Wikipedia will die.

English Wikipedia is in decline. As a long-time editor & former admin, I was deeply dismayed by the process. Here, I discuss UI principles, changes in Wikipedian culture, the large-scale statistical evidence of decline, run small-scale experiments demonstrating the harm, and conclude with parting thoughts.

“Writing a Wikipedia Link Archive Bot”, Branwen 2008

Wikipedia-Archive-Bot: “Writing a Wikipedia Link Archive Bot”⁠, Gwern Branwen (2008-09-26; ⁠, ⁠, ; backlinks; similar):

Haskell: tutorial on writing a daemon to archive links in Wikipedia articles with TagSoup and WebCite; obsolete.

This is a 2008 tutorial demonstrating how to write a Haskell program to automatically archive Internet links into WebCite & Internet Archive to avoid linkrot, by parsing WP dumps, downloading & parsing WP articles for external links with the TagSoup HTML parsing library, using the WebCite/​IA APIs to archive them, and optimizing runtime. This approach is suitable for one-off crawls but not for live archiving using the RSS feed; for the next step, see Wikipedia RSS Archive Bot for a demonstration of how one could write a RSS-oriented daemon.


“A Group Is Its Own Worst Enemy”, Shirky 2005

2005-shirky-agroupisitsownworstenemy.pdf: “A Group is Its Own Worst Enemy”⁠, Clay Shirky (2005; ⁠, ; backlinks; similar):

…We had new applications like the Web, email, instant messaging, and bulletin boards, all of which were about humans communicating with one another through software. Now, suddenly, when you create software, it isn’t sufficient to think about making it possible to communicate; you have to think about making communication socially successful. In the age of usability, technical design decisions had to be taken to make software easier for a mass audience to use; in the age of social software, design decisions must be taken to make social groups survive and thrive and meet the goals of the group even when they contradict the goals of the individual. A discussion group designed by an usability expert might be optimized to make it easy to post spam about Viagra. But in social software design it’s pretty obvious that the goal is to make certain things harder, not easier, and if you can make it downright impossible to post spam, you’ve done your job. Features need to be designed to make the group successful, not the individual.

Today, hardly anybody really studies how to design software for human-to-human interaction. The field of social software design is in its infancy. In fact, we’re not even at the point yet where the software developers developing social software realize that they need to think about the sociology and the anthropology of the group that will be using their software, so many of them just throw things together and allow themselves to be surprised by the social interactions that develop around their software. Clay Shirky has been a pioneer in this field, and his talk “A Group Is Its Own Worst Enemy” will be remembered as a watershed in the widespread realization that in this new era, sociology and anthropology are just as crucial to software design as usability was in the last. —Joel Spolsky

People who work on social software are closer in spirit to economists and political scientists than they are to people making compilers. They both look like programming, but when you’re dealing with groups of people as one of your run-time phenomena, that is an incredibly different practice. In the political realm, we would call these kinds of crises a constitutional crisis. It’s what happens when the tension between the individual and the group, and the rights and responsibilities of individuals and groups, gets so serious that something has to be done. And the worst crisis is the first crisis, because it’s not just “We need to have some rules.” It’s also “We need to have some rules for making some rules.” And this is what we see over and over again in large and long-lived social software systems. Constitutions are a necessary component of large, long-lived, heterogeneous groups. “The likelihood that any unmoderated group will eventually get into a flame-war about whether or not to have a moderator approaches one as time increases.” As a group commits to its existence as a group, and begins to think that the group is good or important, the chance that they will begin to call for additional structure, in order to defend themselves from themselves, gets very, very high.

  1. You cannot completely separate technical and social issues
  2. Members [power users] are different than users.
  3. The core group has rights that trump individual rights in some situations.

…if you don’t accept them upfront, they’ll happen to you anyway. And then you’ll end up writing one of those documents that says “Oh, we launched this and we tried it, and then the users came along and did all these weird things. And now we’re documenting it so future ages won’t make this mistake.”

  1. …If you were going to build a piece of social software to support large and long-lived groups, what would you design for? The first thing you would design for is handles the user can invest in.
  2. you have to design a way for there to be members in good standing. Have to design some way in which good works get recognized. The minimal way is, posts appear with identity. You can do more sophisticated things like having formal karma or “member since.”
  3. Three, you need barriers to participation. This is one of the things that killed Usenet⁠. You have to have some cost to either join or participate, if not at the lowest level, then at higher levels. There needs to be some kind of segmentation of capabilities.
  4. And, finally, you have to find a way to spare the group from scale. Scale alone kills conversations, because conversations require dense two-way conversations.

“Do Incentive Contracts Crowd Out Voluntary Cooperation?”, Fehr & Gächter 2001

2001-fehr.pdf: “Do Incentive Contracts Crowd Out Voluntary Cooperation?”⁠, Ernst Fehr, Simon Gächter (2001-11-05; ; backlinks; similar):

In this paper we provide experimental evidence indicating that incentive contracts may cause a strong crowding out of voluntary cooperation.

This crowding-out effect constitutes costs of incentive provision that have been largely neglected by economists. In our experiments the crowding-out effect is so strong that the incentive contracts are less efficient than contracts without any incentives.

Principals, nonetheless, prefer the incentive contracts because they allow them to appropriate a much larger share of the (smaller) total surplus and are, hence, more profitable for them.