<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Dennis Leeftink on Medium]]></title>
        <description><![CDATA[Stories by Dennis Leeftink on Medium]]></description>
        <link>https://medium.com/@dennisleeftink?source=rss-e3bc4f24817d------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*13JfA-Dm-0alM2KJ.jpg</url>
            <title>Stories by Dennis Leeftink on Medium</title>
            <link>https://medium.com/@dennisleeftink?source=rss-e3bc4f24817d------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 06:19:51 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@dennisleeftink/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[From produsage to redusage]]></title>
            <link>https://medium.com/@dennisleeftink/from-produsage-to-redusage-bcfa3eb50764?source=rss-e3bc4f24817d------2</link>
            <guid isPermaLink="false">https://medium.com/p/bcfa3eb50764</guid>
            <category><![CDATA[question-answering]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[knowledge-economy]]></category>
            <category><![CDATA[knowledge-management]]></category>
            <category><![CDATA[exponential-growth]]></category>
            <dc:creator><![CDATA[Dennis Leeftink]]></dc:creator>
            <pubDate>Tue, 12 Nov 2024 12:01:20 GMT</pubDate>
            <atom:updated>2025-05-09T07:24:05.975Z</atom:updated>
            <content:encoded><![CDATA[<h4>How automated content generation has flipped user-led creation on its head</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/880/1*UcSSW1ZnqNM_1owcG4RJCA.png" /><figcaption>Illustration from <a href="https://unsplash.com/@nypl?utm_source=medium&amp;utm_medium=referral">The New York Public Library</a></figcaption></figure><p><strong>In the wake of the ‘next web’</strong> <strong>of the aughts</strong>, <strong>a new wave of decentralised content creation and distribution channels was heralded as people taking back the reigns of knowledge production</strong>. Communal wiki’s and other user-led initiatives meant that knowledge production did not have to pass through the traditional channels of curation and review anymore, instead becoming ‘permanently mutable bodies of work’ under endless refinement and construction (Bruns &amp; Schmidt, 2011).</p><p>Originating in the early years of <em>wikification </em>as a play on ‘prosumer’, Axel Bruns introduced the term <strong>produser</strong> to refer to this observed shift of content consumers becoming active <em>users </em>and <em>producers </em>in the content creation and consumption process (Bruns, 2008). A shift recently flipped on its head, as content generation models are beginning to drive down prior <em>produsage </em>towards <em>redusage </em>— a global reduction in novel content creation owing to the increased recycling of our existing stock.</p><h3>A crowded future</h3><p>Let’s wind the clock back a bit: the early days of Web 2.0. As the ‘many hands’ involved in our initial crowdsourced endeavours meant that anyone ranging from knowledgeable to ignorant could contribute as they saw fit, new ways of managing and moderating our collective efforts were in dire need.</p><p>For one, this meant negotiating and developing new norms and technical measures to bring a new mass of contributors up to speed to a community’s content quality guidelines. At the same time, active moderation, quality assessments, reputation systems and community leaderboards could tangibly demonstrate someone’s ‘worth’ or standing in a larger knowledge building community.</p><p>While more people could invest in this ‘user-led’ knowledge economy, hierarchies or rather, <em>heterarchies </em>persisted. According to Bruns &amp; Schmidt (2011),</p><blockquote>… neither a flat structure nor a clear and permanent hierarchy, [w]here junior participants do have the potential to gain status by participating in a consistently constructive fashion.</blockquote><p>Intrinsic to these <em>heterarchical communities </em>was (and remains), that although barriers to entry are lower than their hierarchical counterparts, most users would still likely be beholden to status and other network ordering effects. While ‘flatter’ from the outset, many communities would still coagulate around community leaders and preferred ways of doing things, including inventive moderation procedures and novel organisational structures.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/779/1*-QJgVZu8c7MVrnd4Q-Zxpg.png" /><figcaption>Comparing community structures, adapted from Norman et al. (2011)</figcaption></figure><p>Even when built with egalitarian principles in mind, budding creator communities would inadvertently introduce or reinforce certain power differentials, be it by introducing certain participatory requirements (e.g. terms of use) or other normative constraints such as reputation thresholds for unlocking contribution privileges (e.g. having <em>x</em> amount of approved edits). All to maintain the ‘quality’ of a community’s central body of work — <em>its creative commons</em> — from undue tampering.</p><h3>Non-creative commons</h3><p>In recent years however, the role of <em>produser </em>has been steadily giving way to that of the <em>reduser</em>: a marked shift from active contributors to users with a reduced interest, motivation or need to actively contribute back to the community they had previous vested interests in.</p><p>For instance, in the context of collective question-answering sites such as Stack Overflow, automated chat-based systems have become increasingly able to point users to the ‘right’ information while completely going around the communities from where their training data originated (e.g. when trained on <em>The Stack, The Pile</em> or <em>Common Crawl, </em>three go-to datasets). While the site’s activity had been declining before, Chanona et al. (2024) observed a notable drop around early 2023:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*L8Y6QeqJYIgXzpkZ2Si2IQ.jpeg" /><figcaption>A decline in Stack Overflow activity between Dec-2022 and May-2023, from Del Rio-Chanona et al. (2024)</figcaption></figure><p>Compared to prior question-answering scenario’s where users would post questions in open fora for all to see and respond to, now user prompts are sucked into the voids of computing engines and model internals, preventing user communities to discover and respond to queries[1]. This is breaking the existing knowledge chains around which <a href="https://medium.com/automated-decision-making-and-society/burning-the-hooks-what-happens-when-we-lose-our-subreddits-apis-and-exchanges-1b72bd3aa49f">question-answering communities have long organised</a>, turning prior public question-answering practices into an increasingly private affair.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*skwQEI_dhtcOU2k7uD1jeQ.png" /><figcaption>Dataset release timeline, adapted from Liu et al. (2024)</figcaption></figure><p>At the tail end of half a century of information retrieval developments, can we not only witness the disposal of tried and tested knowledge production practices in favour of the latest knowledge management fad or ‘dataset refinement’ (= extracting more/diverse community content), but a general transformation of individual search preferences trading <em>collaboration </em>and <em>observability </em>for retrieval <em>speed </em>and <em>convenience</em>.</p><p>It is not difficult to see why, as immediate answer gratification compared to waiting for days for others to respond, allows question-askers to explore new problem spaces much more readily — even when the immediate answer may seem less comprehensive or accurate.</p><h3>The great (re)cycling</h3><p>Thus, compared to <strong>produsage</strong> which was characterised by “the collaborative and continuous building and extending of existing content in pursuit of further improvement” (Bruns 2008:21), the recent wave of automatic content generators is rather characterised by creative <strong>redusage:<em> </em>a reduction in open collaboration and increased recycling of existing user content.</strong> And we are well set to recycle, well <em>everything. </em>Three hundred trillion tokens worth to be exact, possibly before the end of 2032:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2DSs6_IL0pxyBVBZTaxEgQ.png" /><figcaption>Forecast of the ‘full use of our data stock’ by Villalobos et al. (2024)</figcaption></figure><p>Herein lies the paradox of redusage: as we increasingly eat into our creative commons to train automated content generation systems, the material generated by these very systems drives down the incentive for people to contribute back, as new material based on the sum of all works is only a ‘prompt away’. Instead of going through open channels to refine and request information (and leaving traces for the public to respond to), the flow of content is increasingly directed through private instances, making it more difficult to extend or ‘replenish’ our existing data stock.</p><blockquote><em>Note that redusage is </em><strong><em>not </em></strong><em>readily observed across the board. In a pre-print on Wikipedia, Reeves et al. (2024) note:</em> “where evidence from Stack Exchange found a rapid and significant drop in engagement following the launch of ChatGPT, we find no such drop … the largest opportunity for future work, however is the question of whether these shorter-term trends will hold” <em>(p.9)</em></blockquote><h3>A step back</h3><p>This raises some interesting questions and dilemmas. Was the phenomenon of <em>produsage </em>ever new at all, or rather newly named, only visible at different timescales? Is the <em>produsage / redusage </em>flip common to all <a href="https://doctorow.medium.com/my-mcluhan-lecture-on-enshittification-ea343342b9bc">late-stage innovations</a>? Is there a limit to how large a communal body of works can grow? Or do innovations ‘naturally’ pass through phases of honing and experimentation, after which most technical use-cases are settled and the need for user contributions and negotiation decreases? What other side-effects can we observe when markets are increasingly saturated?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0A4qhtrn8atPL1Y8aapQtg.png" /><figcaption>Increasingly steep adoption curves, <a href="https://hbr.org/2013/11/the-pace-of-technology-adoption-is-speeding-up">HBR</a>.</figcaption></figure><p>Bruns reminds us how the ‘fixed and static’ conception of industrial-era production hardly applies to its digital counterpart, as digital (knowledge) products are often under constant construction or in a state of perpetual beta. We are also reminded <strong>how digital content creators can increasingly switch back and forth between consumptive and productive modes. </strong>To take this further:</p><ul><li>If we assume the rate of oscillation between consumption and production to be the distinguishing factor of <em>produsage</em>,</li><li>and remind that the global rate of innovations to be lower in the past[2],</li><li>we can surmise that past creative practices would have likely involved some form of <em>produsage</em>, albeit at a lower rate of oscillation between usage and production.</li></ul><p>Think how the capacity to switch between content consumption and production would have been ‘capped’ by the general availability and access to other users, technologies and innovations[3]. While in the past, it may have been more difficult for individuals to switch from consuming to producing (say to read, write and publish a manuscript), as the global pool of innovations and distribution channels increased, so too would anyone’s ability to switch between accessing published works and adding their own to the pool.</p><p>Just as with the printing press, we should be careful to assign Gutenberg the sole role of ‘inventor’ or ‘producer’ and delegate the increasing pool of readers to that of the ‘consumer’. Rather, the press allowed common knowledge to be more easily recorded and (re)distributed, lowering the barriers for more people to contribute to expanding this knowledge— the back-and-forth between reading, writing, typesetting and publication just unfolding at a slower pace compared to contemporary <em>produsage.</em></p><h3>Just a phase?</h3><p>While I am asking about more things than ever before (from the trivial to the absurd and complex), current automatic answering tools leave little public trace of my curiosity. What will remain of our collective commons if even our questions are increasingly withdrawn from public view? While the phenomenon of <em>produsage </em>may have been around longer than is often credited, what sets the past decade’s knowledge building efforts apart from prior and later efforts may have exactly been the relative ease in switching between ‘user’ and ‘producer’ roles that Web 2.0 communities afforded.</p><p><strong>But as we stray evermore away from a user-led web</strong>, <strong>the ‘<em>redusage flip’</em> now warrants equal or more attention</strong>.</p><p>It can teach us about the inflection points after which once user-led content creation becomes subsumed by automatic regeneration, a muddled stockpile now at capacity, whether due to market saturation or the next startup lapping up our <a href="https://epoch.ai/blog/will-we-run-out-of-data-limits-of-llm-scaling-based-on-human-generated-data">collective data stock</a>. Better yet, it can teach us about the knowledge production and moderation strategies that <em>did work</em>, and that by taking humans out the loop, we are not only losing valuable co-production, but decrease the pool of questions that would have otherwise been openly asked, refined, and collaborated on, putting the days of the wise and helpful crowd truly behind us.</p><h4>Notes</h4><p>[1]: User conversations are generally only visible to system owners and operators and used as <a href="https://gregoreite.com/drilling-down-details-on-the-ai-training-datasets/">additional training data</a>. For an openly accessible dataset by Zheng et al. (2023) see: <a href="https://arxiv.org/abs/2309.11998">https://arxiv.org/abs/2309.11998</a></p><p>[2]: Compare for instance, annual patent grants per year: <a href="https://patentlyo.com/patent/2023/08/utility-patents-granted-calendar.html">https://patentlyo.com/patent/2023/08/utility-patents-granted-calendar.html</a></p><p>[3]: Not unlike the <em>The Law of Accelerating Returns </em>(Kurzweil, 2001) where each new innovation becomes a stepping stone for future inventions, and that as the pool of available inventions increases, so does rate of innovation rate, exponentially.</p><h3>References</h3><ul><li>Bruns, A. (2008). <em>Blogs, Wikipedia, Second life, and beyond: From production to produsage</em> (Vol. 418). Peter Lang.</li><li>Bruns, A., &amp; Schmidt, J. H. (2011). Produsage: A closer look at continuing developments. <em>New Review of Hypermedia and Multimedia</em>, <em>17</em>(1), 3–7.</li><li>Del Rio-Chanona, R. M., Laurentsyeva, N., &amp; Wachs, J. (2024). Large language models reduce public knowledge sharing on online Q&amp;A platforms. <em>PNAS nexus</em>, <em>3</em>(9), pgae400.</li><li>Felton, N. (2008). Consumption spreads faster today. <em>New York Times</em>, <em>10</em>.</li><li>Kurzweil, R. (2001). The law of accelerating returns. In <em>Alan Turing: Life and legacy of a great thinker</em> (pp. 381–416). Berlin, Heidelberg: Springer Berlin Heidelberg.</li><li>Liu, Y., Cao, J., Liu, C., Ding, K., &amp; Jin, L. (2024). Datasets for large language models: A comprehensive survey. <em>arXiv preprint.</em></li><li>Norman, G.J., Catherine J. Norris, Jackie Gollan, Tiffany A. Ito, Louise C. Hawkley, Jeff T. Larsen, John T. Cacioppo, and Gary G. Berntson. “Current emotion research in psychophysiology: The neurobiology of evaluative bivalence.” <em>Emotion Review</em> 3, no. 3 (2011): 349–359.</li><li>Reeves, N., Yin, W., Simperl, E., &amp; Redi, M. (2024). “ The Death of Wikipedia?” — Exploring the Impact of ChatGPT on Wikipedia Engagement. <em>arXiv preprint</em>.</li><li>Villalobos, G., Ho, A., Sevilla, J., Besiroglu, T., Heim, L., &amp; Hobbhahn, M. (2024). Will We Run Out of Data to Train Large Language Models? [Blog]. <em>Epoch</em>. Available at <a href="https://epoch.ai/blog/will-we-run-out-of-data-limits-of-llm-scaling-based-on-human-generated-data">https://epoch.ai/blog/will-we-run-out-of-data-limits-of-llm-scaling-based-on-human-generated-data</a></li><li>Zheng, L., Chiang, W. L., Sheng, Y., Li, T., Zhuang, S., Wu, Z., … &amp; Zhang, H. (2023). Lmsys-chat-1m: A large-scale real-world llm conversation dataset. <em>arXiv preprint.</em></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bcfa3eb50764" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[We need more failure]]></title>
            <link>https://medium.com/@dennisleeftink/we-need-more-failure-3fd7064f3c47?source=rss-e3bc4f24817d------2</link>
            <guid isPermaLink="false">https://medium.com/p/3fd7064f3c47</guid>
            <category><![CDATA[learning-and-development]]></category>
            <category><![CDATA[academia]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[crowdstrike]]></category>
            <dc:creator><![CDATA[Dennis Leeftink]]></dc:creator>
            <pubDate>Sun, 21 Jul 2024 15:20:24 GMT</pubDate>
            <atom:updated>2025-01-15T16:22:49.193Z</atom:updated>
            <content:encoded><![CDATA[<h4>How to mend massively broken systems</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*guhkAIi83M8U2saq" /><figcaption>Photo by <a href="https://unsplash.com/@jachan_devol?utm_source=medium&amp;utm_medium=referral">Jachan DeVol</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>A technical outage so massive, unprecedented, impactful — why is it we only tend to care about our technical dependencies <em>after</em> the fail-safes have been broken? Leaving our house without a phone, being offline in a new city, losing access to an important account... these things we generally try to avoid. Even the smallest hiccup in our technical routines makes us not only deeply uncomfortable, but increasingly, incapable. Why then, are we not more concerned about all the flimsy technical layers that envelop our lives <em>before</em> they imminently crack?</p><p>While the Crowdstrike + Azure security whammy unfolding over the weekend has everyone up its arms, on any normal day our technical concerns rarely start before all system foundations have given out. After a security patch pushed by Crowdstrike on Thursday did just that — rendering millions of Windows devices inoperable in the process — cost projections are now in the billions, with almost as many lives and occupations afflicted and no less concerned sysadmin faces. A wake-up call, if you or your devices had been fed the unduly update.</p><p>As specialists scramble to a solution (complicated by how affected devices are ‘soft-bricked’ and difficult to debug remotely), folks on the ground have been met with a harsh reality of all the critical computational work we have since outsourced. We haven’t completely forgotten how to operate a pen and paper just yet, but we are close. I don’t lament this apparent loss of skill but I am worried about the finer print of our technical contract: <em>we are blind to how technologies fail us, all the time</em>.</p><h4>Being system illiterate</h4><p>Okay, maybe not all the time, but we are blind. Blind to a minute file synchronisation error of some note-taking app we use to jot down our thoughts. Blind to that severely outdated BIOS driver that stopped receiving updates years ago. Blind to the development shortcuts someone has taken to meet a release deadline. Blind to the hundreds of vendors and development teams we rely on daily to run our cloud based work. We aren’t blinded because we don’t want to see but because we <em>cannot</em> easily see behind the thick walls of our software production lines.</p><p>Besides the obvious bug or error, most development steps behind our software can only be retraced to a certain degree. Even if you do, the difficulty remains in understanding these technical pedigrees holistically. For instance, the <a href="https://distrowatch.com/dwres.php?resource=family-tree"><em>Genealogy of Linux</em></a> may be well documented, but a <em>true understanding</em> of how each individual release relate to its predecessor is a feat only reserved to the dedicated few. And <em>even if</em> you were to possess a full knowledge of an OS’s history, what can you or any single person, entity or vendor realistically do to relay such complexity to the end-user, before mindlessly clicking &#39;Update Now&#39; to get on with our day?</p><p>The reality is, we simply cannot read and parse all our intricate software and hardware dependencies. Apart from a cohort of a few apt engineers, you and I are systems illiterate. We may have a hunch as to to how our tech ‘talks’ and we might ‘speak’ their programming to some degree, but most of us are nowhere near able to follow along the more complex conversations technical systems are increasingly having on our behalf. If you are proficient, kudos, but as Thursday’s incident shows, even the most technically skilled among us may not be fluent enough when it comes to the complex grammars of our new global software interlocutors.</p><h4>Give error some room</h4><p>We have dedicated whole industries solely focused on foolproofing software. Slick infoboxes greet us when installing apps, user profiles come fully preconfigured and if an error occurs, troubleshooting guides are only clicks away. Onboarding new users should be as seamless as possible and without harm, while updates are aggressively pushed with ‘sensible’ defaults. Your software bedding has been made, don&#39;t worry about the monster beneath. But as most things in life, shielding your dependants from any hardship and promising lives without nightmares rarely has the intended effect.</p><p><em>We need room to fail in order to learn.</em> And sometimes that room will be massive. Dissuading failure on a system’s level is doing a disservice to how people learn, acquire new technical skills, hone their vocabularies and become proficient. Mind, I am not advocating everyone to become as fluent as a full-time engineer, make our apps more error prone or have developers mindlessly ‘push to prod’ and see what happens. I want to draw attention to how creating systems that prevent us from making any mistake <em>at all</em> can lead to situations where we don’t even know how technically mute we are until we are suddenly put to the test.</p><p>If your system had been idly purring away for some years, firing up the bootloader may just as well have been arcane knowledge (the unusual keycombo had me on edge for a minute). Most of our technical struggles however, reach far beyond remedying the untimely blue-screen. In fact, even <em>admitting</em> to being stalled on some minor issue, we are often ashamed of our apparent illiteracy. No wonder that in a heavily platformed, multi-app environment where some are carrying 20+ years of personal <em>and </em>occupational data, our attention for the latest and greatest tools while managing our existing ones is spread all too thin.</p><h4>Fail proficient</h4><p>By now, it is no secret that leaning into one software ecosystem makes switching to another all the costlier as time goes on. At the same time, there is no guarantee that whatever ecosystem we choose now won’t be shelved tomorrow. In my experience of building academic research tools, this budding ‘time creep’ is mostly an afterthought or rarely considered at all during the initial phases of a project.</p><p>Maybe worse, when faced with time and resource constraints, adopting new software is almost regarded as an error in itself. Just think about the difficulty of having your colleagues find the time (and energy) to learn and navigate that slightly obtuse interface of your latest shiny app that will definitely solve their nightmares. Yet when faced by similar situations myself, I am no stranger to sticking to ‘what works’ over something unfamiliar and seemingly less resilient.</p><p>Besides maybe those at the very start of their careers, we are simply too constrained to rebuild our toolshed from the ground up — even though we may very well want to. Could we embrace work environments more accommodating to our messy technical work, incompatibilities and (in)proficiencies? Provide room for explicit failure and experimentation, not just catering to the technical regimen of a brilliant few? It may be too late in moments of crisis. We need to fail more in the open.</p><p>We just need systems that allow us to.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3fd7064f3c47" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Burning the Hooks]]></title>
            <link>https://medium.com/@dennisleeftink/burning-the-hooks-1c87b05d4132?source=rss-e3bc4f24817d------2</link>
            <guid isPermaLink="false">https://medium.com/p/1c87b05d4132</guid>
            <category><![CDATA[stackexchange]]></category>
            <category><![CDATA[reddit]]></category>
            <category><![CDATA[api]]></category>
            <category><![CDATA[platform-governance]]></category>
            <category><![CDATA[large-language-models]]></category>
            <dc:creator><![CDATA[Dennis Leeftink]]></dc:creator>
            <pubDate>Tue, 13 Jun 2023 14:17:04 GMT</pubDate>
            <atom:updated>2024-11-12T12:09:26.974Z</atom:updated>
            <content:encoded><![CDATA[<h3>Burning the (API) hooks: What happens when we lose our subs, APIs and exchanges?</h3><p>Reflecting on the current StackExchange and Reddit fallout. Co-authored with fellow PhD <a href="https://medium.com/u/510d19e16e69">Dominique Carlon</a> for the Queensland University of Technology’s ADMs node. Full version at:</p><ul><li><a href="https://medium.com/@dominique.m.carlon/burning-the-hooks-what-happens-when-we-lose-our-subreddits-apis-and-exchanges-1b72bd3aa49f">https://medium.com/@dominique.m.carlon/burning-the-hooks-what-happens-when-we-lose-our-subreddits-apis-and-exchanges-1b72bd3aa49f</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1c87b05d4132" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>