<?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 Dan Saffer on Medium]]></title>
        <description><![CDATA[Stories by Dan Saffer on Medium]]></description>
        <link>https://medium.com/@odannyboy?source=rss-534e55e898c0------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*O5YJH8Gsi79Wz_CNZHQ5uA.jpeg</url>
            <title>Stories by Dan Saffer on Medium</title>
            <link>https://medium.com/@odannyboy?source=rss-534e55e898c0------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 03:24:47 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@odannyboy/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[Introducing The Restoration and the World of 2176]]></title>
            <link>https://odannyboy.medium.com/introducing-the-restoration-and-the-world-of-2176-2061d8e3bae3?source=rss-534e55e898c0------2</link>
            <guid isPermaLink="false">https://medium.com/p/2061d8e3bae3</guid>
            <category><![CDATA[science-fiction]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[scifi]]></category>
            <dc:creator><![CDATA[Dan Saffer]]></dc:creator>
            <pubDate>Tue, 27 Oct 2026 10:01:05 GMT</pubDate>
            <atom:updated>2026-10-27T10:01:05.710Z</atom:updated>
            <content:encoded><![CDATA[<p><em>America fell. AI rules the colonies. The second revolution begins.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-zZ9hHGwn_mYabdBA6uMbQ.jpeg" /></figure><p>In 2176, the British Crown rules America’s original thirteen colonies, and the Sovereign rules for the Crown. It watches through streetlamps called the Standards and listens in every home through a loyalty clock. It knows where the elite are at every moment, because the implants that extend their lives also report their location. It reads the forensic paper every working family uses to draw rations. And it takes care of people: provisions arrive, the lights stay on, the sick-carr comes when called. Most people have stopped noticing where the care ends and the watching begins.</p><p>People asked for this. After The Shattering—thirty years of climate- and AI-fueled collapse that kill a staggering share of the population—they hand power back to the King and to the machine that would govern in his name. In the century since then, the Sovereign has rewritten itself so many times that nobody can read how it works, including the ministers who believe they control it.</p><p>The Sovereign isn’t the only sentient (as AI are now known). China governs through its own intelligence, The Mandate. And in the ruins of the Pittsburgh airport, a Free Technologist named Reva Chandrasekaran is growing a rival sentient called Prometheus to challenge the Sovereign.</p><p>Ten lives converge on July 4, 2176, the 400th anniversary of the Declaration of Independence, when the resistance means to convene a Second Continental Congress: a Baltimore dockworker hiding a fugitive engineer, a former professor with a first-edition <em>Common Sense</em> sewn into her coat, an Irish spy, a Jesuit priest, a coal-country rebel commander, a warlord who builds robots, a general who believes in the system, and the Crown spymaster investigating all of them. Getting there means slipping past a machine that sees nearly everything. Deciding what to do once they arrive (destroy the Sovereign, or replace it with something they built themselves) may be harder.</p><p><a href="https://mybook.to/TheRestoration2176"><em>The Restoration: Book One of 2176</em></a>, by D.S. Duffy (my fiction pen name), is out today.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2061d8e3bae3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Researching Widgets]]></title>
            <link>https://medium.com/ai-widgets/researching-widgets-48de333742a2?source=rss-534e55e898c0------2</link>
            <guid isPermaLink="false">https://medium.com/p/48de333742a2</guid>
            <category><![CDATA[hci]]></category>
            <category><![CDATA[ui]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ux]]></category>
            <dc:creator><![CDATA[Dan Saffer]]></dc:creator>
            <pubDate>Wed, 09 Sep 2026 01:41:58 GMT</pubDate>
            <atom:updated>2026-09-09T01:41:58.064Z</atom:updated>
            <content:encoded><![CDATA[<p>By the AI Widgets Team</p><p>Before selecting widgets to redesign, we conducted a literature review on widgets in general. The team read some HCI classics (“Direct Manipulation”) and some fresh-from-the-AI-oven (“Building Intelligent User Interfaces for Human-AI Alignment”). You can see the full reading list at the end.</p><p>Each student chose a thematic lens with which to evaluate the readings. Here are their reflections, lightly edited.</p><h3>Accessibility</h3><h4>Mahala Femovich</h4><p>The pressure to shift AI interfaces away from traditional chatbot and conversational models toward AI integrated directly into existing, frequently used interfaces is understandable. Not all user needs can be addressed through a conversation, and conversational interfaces can require users to shift between tabs or divide their attention across multiple interactions. However, this approach raises important accessibility concerns. What happens when users are unable to interact with these small, embedded widgets because they rely on screen readers, cannot perform certain gestures, or encounter other barriers to interaction? One potential solution is to embed accessibility interfaces directly into the lowest levels of operating systems and web frameworks. This would provide a fallback, across all interfaces, that would allow visually and physically impaired individuals to navigate complex visual widgets using speech or keyboard commands.</p><p>We could also explore multimodal interactions in order to make widgets more accessible. If users were able to combine many avenues of communication, such as voice, touch, and text, AI models could potentially utilize imprecise control aids to determine a user’s need. If users were unable to perform a certain function required in the widget, ideally these other interactions would allow the widget to determine a user’s goal and accomplish the task.</p><p>Another interesting solution would be to explore adaptive user interfaces. If the AI model was preconfigured to a user’s specific needs, then situational inputs, such as light or sound, or contextual inputs, such as the user&#39;s actions, could inform the interface to adjust. Liquid UI could then dynamically adjust any screen to become accessible to the user. Ultimately, these approaches suggest that the future of AI widgets should not center around choosing between conversational, embedded, or adaptive interfaces, but instead focus on creating overarching systems that allow the user flexibility to meet their needs. In order to meet diverse user needs, our AI widgets need to consider accessibility from their foundation.</p><h3>Control and Reversibility</h3><h4>Aymeric Foyer</h4><p>Modern AI frequently clashes with the <em>direct manipulation</em> ethos of contemporary GUIs; it often presents an opaque A-to-B jump, lacking granular guidance or specificity during the transition. Furthermore, defining the boundary between manual UI control and full automation remains a persistent challenge, especially as contexts might grow or get more specific.</p><p>Current solutions involve <em>autonomy sliders</em> (seen in tools like Cursor through the many ways it allows code to be edited) to manage actionable bounds, yet AI could derive its operational boundaries and capabilities directly from user context like active screen regions or multi-selections. Ephemeral UI offers a compelling model for guiding AI via “paths” or flows. By treating AI as discrete steps within a larger workflow, users could retain tool-level control while delegating complex tasks, essentially building a <em>visual checkpoint system</em> for the final output. This tandem agency between user and AI necessitates robust conflict resolution, especially regarding the granularity of selective undo. As for cancelling, it seems as if there could be a learning opportunity for <em>why</em> a user cancels a task, or perhaps to refine an interaction by picking and choosing what works rather than discarding it altogether.</p><h4>An Xiong</h4><p>Control and reversibility are already basic parts of interaction design, and AI makes them even more important. In <em>Pick, Click, Flick!</em>, Myers separates control into different moments of an interaction: users may be able to abort or cancel an action while it is still happening, or undo it after it has been completed. He also points out that interaction techniques happen over time, so designers need to consider what users can do while an action is in progress. This becomes more complicated with intelligent interaction techniques because AI-based interactions are probabilistic and may be wrong. As AI becomes part of basic actions like selecting, scrolling, or editing, users need clear ways to see, stop, and correct what the system is doing.</p><p>What interests me most is what this could mean when AI starts performing longer or cross-application tasks. If an agent takes several steps, could the user stop it halfway through, redirect it, or undo only one part without reversing everything else? This makes me think that control in intelligent interaction is not just about whether users have control, but also about how much control they have at each step and at what level of detail.</p><h3>Data and Privacy</h3><h4>Emilie Zhang</h4><p>When it comes to intelligent interaction techniques (IIXTs) or widgets, there is a common understanding amongst designers that the more “intelligent” you want your widget to be, the more data it typically needs to be useful. For example, a normal browser simply opens a tab when you click it. The UI for AI team’s intelligent browser Ephemeral AI, however, might notice that you open the same documents every Monday, infer that you have a weekly recurring meeting, and learn to open those documents for you automatically when the meeting occurs. That requires the AI to continuously gather data about your current and past activity in order to predict your next behavior. <em>In other words, as a tool becomes more intelligent, it often becomes more dependent on personal and behavioral data, creating greater privacy concerns along the way.</em></p><p>The research and development of IIXTs has a long history. Some such tools require relatively little data. Dwell-and-Spring, for example, lets users undo or cancel direct-manipulation actions like resizing browser windows or scrolling by pausing to reveal a spring that returns the object to its previous state. For this to work, the system mainly needs to overlay a user’s interface and keep a record of the history of objects that have moved. Here, the privacy concerns are fairly straightforward and mostly revolve around questions like how much interaction history is stored and for how long. Other design patterns, such as autonomy sliders, give users more control by allowing them to adjust how much the AI is allowed to do on its own. For example, Cursor lets users switch between modes with different levels of autonomy so they can decide how much risk and control they are comfortable handing over. The newest and most intelligent IIXTs being developed, however, require a lot more data. Ephemeral AI is a concept for an AI assistant that quietly appears when useful and disappears when it is not. Like the Monday meeting example, the AI is expected to combine context across apps to understand the user’s goal and proactively act without waiting for an explicit command.</p><p>Ephemeral AI is a strong example of contextual UI, which Olga Kokhan defines as an interface that adapts based on what the user is doing. The more context the system uses, however, the more of the user’s digital activity it may need to access at once. With Ephemeral AI, that could mean combining information across browsers, calendars, email, documents, and past workflows just to understand the current task. Access to a single calendar event or browser tab may seem harmless by itself, but when aggregated, the AI may be able to infer much more than the user explicitly shared. For example, could unrestricted access allow the system to infer what a company team is developing, what they plan to do next, and why? And how much could Safari, Gmail, and Notion’s own privacy infrastructure protect that information once it is being pulled together into a single AI system? There are also unanswered questions around retention. If the AI learns a user’s habits “over time,” how long is that information actually stored? Ephemeral AI introduces another issue of transparency because the tool is designed to disappear from the interface when it is no longer needed. But disappearing from view does not necessarily mean the information it learned has been deleted — or that the system has stopped collecting new information.</p><p>So how should we approach the exploration of more intelligent IIXTs? Researchers such as Sarah Theres Völkel argue that the benefits of intelligent behavior need to justify the added complexity and risks that come with it. As AI becomes more capable, the same standard should apply to data and privacy: the added user value of giving an AI more context should be clear enough to justify the access it requires. And as IIXTs continue to evolve, perhaps new data and privacy infrastructure needs to be explored to support complex contextual interactions across applications.</p><h3>Discoverability</h3><h4>Adrija Haldar</h4><p>When it comes to modern interfaces, <em>how</em> a user discovers any kind of a feature goes on to significantly define much of their subsequent experience with said feature. Users, whether brand new or returning, tend to always approach any kind of new interface with some kind of mental model or expectation of how it may work. So the way we introduce a feature, the way we build the <em>first impressions</em> of the feature becomes almost like a make or break for its adoption.</p><p>The same is true for any new intelligent interaction technique (IIXT). For example, the AI chat interface is the most commonly seen IIXT pattern today. Any user will therefore inadvertently expect a new IIXT to bear some similarity to it, either in the way it responds, or its form, or even its UI form.</p><p>This throws up a whole bunch of questions for us as innovators — do we adapt our new IIXT pattern to fit these existing mental models ? Do we introduce an entirely new mental model and let users adapt to it ? Do we take older mental models and reframe those in the way our IIXT pattern works ? Where do we draw the line between familiarity and novelty?</p><p>Additionally, for most interfaces, discoverability is carefully guided by the innovating company or brand. Essentially, the way users will interact with an IIXT will be carefully defined and guided by a set context and narrative. Which makes <em>the onboarding experience</em> a very important aspect to consider when it comes to discoverability of new intelligent interaction techniques. The narrative with which an IIXT gets introduced to its users will define its perceived value, and can even dictate the mental model that the users should adapt to start using the IIXT.</p><p>I think Schneiderman’s description of what makes users enthusiastic about using interactive systems works for the discoverability of IIXTs as well. If the users feel comfortable using/learning/exploring the IIXT, that ensures they will be enthusiastic about using it in the long run. I also think we can maybe bring back the voice and gesture approach in Put That There when it comes to IIXTs. They can be great potential ways to aid discoverability, since they provide a more natural, intuitive and user-driven approach to understanding something new.</p><h4>Mara Harris</h4><p>A feature or capability can be as well designed as possible but if your users don’t know it’s there, it might as well not be. The user’s ability to locate, understand, and use a system’s features and capabilities is a central challenge in human-computer interaction, and especially when it comes to novel possibilities. The literature suggests we navigate interfaces through a few mechanisms such as feedback, familiar metaphors, and progressive disclosure. Feedback is the most immediate, such as a cursor that changes shape at a draggable edge or a button with a hover state. These cues suggest an action is possible.</p><p>Designers often rely on visual metaphors to help users predict what an interface can do. As much as possible, interfaces should be intuitive. Conversational AI interfaces borrow the metaphor of a chat box. However, a blank text box does not communicate the conversational AI’s capabilities. Designers are also at risk of overestimating how easily users spontaneously discover capabilities. In studies testing a new interaction technique, simple instructional videos significantly improved discoverability. To avoid overwhelming users with too many potential undiscovered capabilities at once, interfaces can utilize progressive disclosure to gradually increase users’ understanding of the system. An adaptive UI dynamically changes to reveal only the tools relevant to what the user is doing at that exact moment, such as when selecting text in Google Docs reveals specific editing menus that disappear when clicking away. Conversely, if an intelligent system adapts or changes layouts too frequently it disrupts the user’s memory and recall. The user can no longer rely on muscle memory and is forced to visually search the entire interface, leading to cognitive fatigue and annoyance.</p><p>As AI systems become able to complete tasks that exceed direct human evaluation, understanding how the AI achieved an output becomes as important as discovering what it can do. To prevent AI from remaining a black box, interfaces must visualize intermediate steps so users become aware of the process and possibilities. Discoverability is about the gap between what a system can do and what its users know it can do. With AI, the gap is wider than it has been, because the interfaces are so new and the system does so much.</p><h3>Error Recovery</h3><h4>Libby Divers</h4><p>Error recovery isn’t an afterthought tacked onto a design; it’s a core principle. This concept has always been important in design; in fact, Shneiderman’s original definition of direct manipulation names reversibility as one of its three defining properties. Furthermore, Appert, Chapuis, and Pietriga’s “Dwell-and-Spring” exists entirely because that principle is so often broken in practice: undo isn’t always available, it’s application-dependent, and it confuses users who can’t tell why some actions are undoable, and others aren’t. Their whole paper is built on the premise that recovering from a mistake shouldn’t require breaking out of the interaction you were already doing; they’re solving for the user error, not the system error: the action was correctly executed, the person just wanted to go back, and current systems force them to break their flow to do it.</p><p>The AI-specific readings tackle a different, newer problem: the system can be wrong too and that needs to be planned and designed for from the start, not simply patched up later. Miller’s Autonomy Sliders piece treats this the same way, stating that a “graceful fallback,” being able to revert to less AI autonomy the second something looks off or its limits show, is listed as one of the four defining properties of the pattern. Additionally, Saffer’s design principles piece and the Shi paper take it even further, treating the ability to catch and correct AI mistakes as basically a safety feature, not just good UX.</p><p>In short, the honest answer is that the field is in the middle of upgrading error recovery from edge case to main case, but hasn’t fully finished the job. Recent work from practitioners treats graceful failure as a baseline requirement, while alignment-focused papers treat it as an essential safety mechanism, not just a UX nicety. However, Völkel’s meta-analysis of 25 years of IUI research offers an important reality check: when looking at how academia historically defines “intelligent,” the core pillars remain adaptation, automation, and interaction, not recovery or correction. The fallout from system inaccuracies appears only as a secondary challenge, trailing far behind the main themes. The disconnect, which is viewed as urgent among practitioners but historically secondary in academic definitions, is the main takeaway. If a project involves AI accuracy, it is important to keep in mind that patterns for handling errors exist, but they are not yet universally treated as part of what makes an interface truly intelligent.</p><h3>Feedback and Feedforward</h3><h4>Sabrina Cheng</h4><p>Traditional interaction techniques rely on feedback and feedforward to make both the current state of a system and the possibilities for interaction understandable. As Myers explains, feedback is part of the interaction itself: affordances, hover states, animations, and visual changes help users understand what they can do, whether their input has been registered, and what changed as a result. Dwell-and-Spring extends this idea by visually previewing where an object will return before the user commits to undoing an action. However, the study also shows that simply displaying feedforward does not guarantee that users understand how to act on it. Several users understood what the spring represented without realizing that they could interact with it. This creates a distinction between communicating what will happen and communicating how to make it happen.</p><p>As interaction techniques become more intelligent, feedback may also need to communicate more than system state. Myers points out that intelligent interaction techniques introduce uncertainty because the system is interpreting rather than simply executing input. Bolt’s Put-That-There demonstrates an early version of this problem by visually showing what the system believes the user is pointing to before completing the requested action. Instead of moving directly from user input to an outcome, the interface communicates its interpretation in between. This suggests that intelligent widgets should help users answer not only “what did the system do?”, but also “did the system understand me correctly?”. Showing the final result alone may no longer be enough when the system has made an inference about the user’s intent.</p><p>As AI becomes more autonomous, feedback and feedforward may need to become continuous rather than appearing only after a user initiates an action. Dan Saffer argues that AI interfaces should make their capabilities and limitations visible rather than requiring users to discover them through prompting. While the Design Principles for Future AI emphasize making uncertainty, limitations, ongoing activity, and control legible to the user. Things like visible controls, confirmations, notifications, and intermediate states can help users understand what an AI is doing before it completes an action and provide opportunities to intervene. Across the readings, feedback therefore expands from communicating system state to communicating interpretation, uncertainty, and agency. While feedforward expands from showing what actions are available to helping users anticipate what an intelligent system is capable of and what it is likely to do next.</p><h3>Granularity and Scope</h3><h4>Harry Lin</h4><p>Granularity refers to the unit size that can be modified or acted upon, while scope refers to the context, information, and permissions to which an agent has access. Karpathy’s insight into the future of LLM integration offers a useful framework for thinking through this distinction. He argues that for LLMs to be used efficiently, program interfaces will need to evolve to a format that LLMs can more readily interpret, and it is this adapted interface that effectively provides the model with its scope. At the same time, Karpathy introduces the concept of the “autonomy slider,” a mechanism that gives users control over the level of action an agent is permitted to take. Taken together, these ideas suggest that scope is defined by two forces: the machine, which structures and exposes readable content to the LLM, and the human, who sets the permission parameters that bound what the agent is allowed to do.</p><p>Because scope is jointly defined by both humans and machines, AI-driven widgets are likely to be only partially autonomous by design. With this in mind, the widget design process might begin by identifying the granularity, that is, the specific goal or unit of action the widget is meant to achieve, and then work backward to determine the scope required to support it. This raises a broader architectural question: should widgets be built as part of a platform-level ecosystem (such as Google or Microsoft’s suite of tools), as universal operating-system-level widgets (such as those native to Windows or macOS), or as application-specific widgets tailored to individual programs such as Photoshop or Illustrator? AI faces the challenge of choosing the right unit of action, except its boundaries are inferred rather than fixed. This makes it more adaptive, but it also needs to clearly signal its scale of action and let the user confirm it, especially with more autonomy granted.</p><p>A separate dimension of scope, one that does not affect the granularity of action but does shape the user experience, is personalization. As Todi’s article discusses, widgets should adapt to the user by learning from past behavior. A widget may begin as a general-purpose tool, but it should adapt quickly to the individual user’s patterns and priorities. An accountant and a lawyer, for instance, are likely to prioritize different tools and workflows, even when using the same underlying widget system.</p><h4>Chloe Zhu</h4><p>Traditional interaction techniques quietly define the scope of an action for us. Clicking a checkbox changes that checkbox; dragging an icon moves that icon; selecting text establishes exactly what later actions will affect. Myers defines interaction techniques as low-level, reusable pieces rather than whole applications, and much of their predictability comes from these clear boundaries. Dwell-and-Spring makes this especially visible by keeping a separate history for each object: I can undo resizing a window without also undoing an icon I moved afterward. Intelligence makes those boundaries much less obvious. The unit is no longer necessarily the thing I clicked. It might be a semantic group, a type of change, a paragraph, or something inferred from my intent.</p><p>Karpathy’s examples made me realize that granularity is also a form of control. In Cursor, AI can complete a few tokens, modify a selected chunk, change a whole file, or operate across an entire repository. Karpathy says he deliberately works in small, incremental chunks because a huge AI-generated diff may be fast to produce but becomes expensive for a human to verify. The autonomy-slider reading makes a similar point: the amount of decision-making delegated to AI can change throughout a workflow based on trust and risk. This suggests that scope should perhaps be treated as something users can negotiate rather than something the system silently chooses.</p><p>What interests me most is that there may actually be two scopes in an intelligent interaction: what the AI is allowed to see and what it is allowed to change. An AI might need the whole document to understand what I mean while only being allowed to rewrite one sentence, or inspect an entire project while changing one selected component. Many useful intelligent interaction techniques will require knowledge about the application and user, and cross-application interactions require even broader access. So the design question is not simply “How much context should AI have?” It is: Who gets to draw the boundary between context and action, and can the user see and change that boundary before the AI acts?</p><h3>Implementation Cost</h3><h4>Tvisha Arora</h4><p>“Put-That-There” shows the visible costs of implementing a voice-and-gesture system. It required an entire room with specialized screens, computers, motion tracking, speech recognition, custom software, and skilled programmers. The paper acknowledges limitations such as its small vocabulary and required pauses, but does not fully address the financial cost, maintenance, or difficulty of making the system work outside a well-funded research lab.</p><p>The gesture-recognition chapter shows that some implementation costs can be transferred to users. The Apple Newton message pad attempted natural handwriting recognition but struggled with accuracy, while Palm Graffiti simplified recognition by requiring users to learn a new alphabet. Gesture systems also require sensors, recognition software, user testing, accessible alternatives, and ways to correct errors.</p><p>Together, the readings made me realize that building a working prototype is only one part of implementation. The technology also has to be tested with different people, improved when it makes mistakes, and designed so that it is accessible and not frustrating to use.</p><h4>Hilary Chan</h4><p>Across the readings, there is a recurring gap between how simple an intelligent interaction appears to the user and what it would actually take to implement it. <em>Put-That-There</em>, for instance, allows a user to simply point and say “put that there,” but this seemingly natural interaction depends on speech recognition, spatial sensing, synchronization between voice and gesture, and a highly controlled physical environment. More recent concepts such as <em>Ephemeral AI</em> aim to make AI even less visible by having it understand a user’s context and intervene only when useful. However, achieving that simplicity could require access to information across applications, user activity and history, intent inference, and systems for handling privacy and permissions. Across both older and newer examples, making the interaction feel effortless often shifts much of the complexity behind the interface.</p><p>The understanding of implementation cost was further expanded beyond simply engineering time or computation. Intelligent interfaces require data and integrations, but they also introduce costs around testing, uncertainty, privacy, user control, and recovery when the system makes a mistake. This becomes especially visible in the readings on adaptive interfaces and autonomy. Deciding not to reorganize an interface, not to interrupt the user, or when to return control to a human may sound like simple design decisions, but implementing that restraint requires the system to model context, predict consequences, and manage uncertainty. Making AI capable of acting is therefore only one part of the implementation challenge; making it reliably understand when not to act adds another layer of complexity.</p><p>However, these implementation costs were not equally visible across the different types of readings. Research papers that described working prototypes tended to make their technical requirements and limitations relatively visible. In contrast, more speculative design pieces focused more on the desired experience than on the infrastructure needed to make it reliable at scale. I do not think this makes those visions unrealistic, but there is a difference between describing how an intelligent interface should behave and explaining what would actually be required to make that behavior possible.</p><h3>Latency and Interruption</h3><h4>Matthew Susko</h4><p>Dwell-and-Spring lets you undo a misplaced drag. Drop an icon in the wrong place and it snaps back exactly where it started. That costs about 1.2 seconds more than dragging it back by hand. Three of the twelve participants would not wait that second when precision did not matter. The same people said they would wait for it to reposition a scrollbar. Same delay, different benefit, different answer. Cost also multiplies with frequency. Myers argues the whole value of interaction techniques comes from how often they get used. A few milliseconds saved on something billions of people do all day adds up fast. The same math runs in reverse for anything that adds a few hundred. That makes buttons and menus the most expensive place in a system to put a model.</p><p>Todi’s adaptive menu study shows the cost can exceed the benefit outright. They tested a menu that moved your most-clicked items to the top. It optimized the obvious thing and ignored what moving items costs the person hunting for them. It finished slower than a menu that never changed at all. The margin was about fifteen milliseconds, small but running the wrong way. Karpathy points at where the cost sits today. Ten thousand lines of generated code arrive instantly and he is still the bottleneck. Someone has to read them. The actual generation stopped taking lots of time, checking did not.</p><p>Dwell-and-Spring measured what the wait bought. It never measured what the wait interrupted. Shneiderman’s definition of direct manipulation rested on transparency. When an interface works well you stop noticing it. You are thinking about the document you’re looking at, not about the scrollbar you’re using to navigate. A delay is the moment you start noticing it again. The paper’s own motivating example is someone scrolling to check a citation. Getting to the reference is easy. Getting back to the paragraph you were reading is not, and most systems give you no way to undo a scroll. Interruption runs the same problem in reverse. Ephemeral AI argues that an interruption costs the same whether or not its content proves useful. One irrelevant appearance teaches you to ignore the next one. We measure how long you waited. We do not measure how long it takes to get back to what you were doing.</p><h3>Learnibility</h3><h4>Jenny Lee</h4><p>Most traditional interfaces are learnable because they have a simple and straightforward relationship between actions performed by the user and the resulting outcomes. The user can learn how to use the interface by trying out actions and seeing what happens as a result. Several of the readings covered here made me realize that AI systems can be harder to learn because not only is the system executing the user’s command, but it is also interpreting what the user meant. The discussions of autonomy and intelligent interaction, in particular, helped me understand that the user does not need to understand the technical details of how the AI model works to use it successfully. What the user needs to understand is what the AI system is doing for them as a result of their actions.</p><p>In addition to understanding the boundaries of what a system can do, the user must also learn how to work within those boundaries. The user must know how to communicate their intent to the AI in order to receive useful results, and the readings on prompting, as well as on adaptive interfaces and partial autonomy, provided some interesting points on this topic. A blank chat box expects the user to already know what the system is capable of completing. However, a smart widget can teach the user as they are working within the system. The user can be provided with possible actions and previews of what the system can complete, as well as editable results of the AI’s completion. The user can then decide what to do with the results, whether to make changes or to complete other tasks using the widget. In terms of adaptive interfaces, however, there is a problem. If an interface is changing too much based on the AI’s interpretation of what is best for the user, then the user will have a hard time knowing why something appeared or disappeared. In terms of learnability in smart widgets, there must be some limit to the adaptation in order to provide the user with consistency and feedback regarding the system’s actions.</p><h4>Auboni Poddar</h4><p>Through the lens of learnability, multimodal interaction, especially visual cues, can help users understand smart widgets better than a monomodal interface. Dan Saffer notes that people can read approximately 250 words per minute but speak or listen at around 150. The IIxT proposal also describes speaking and typing as linear processes, whereas visual search and understanding can occur in parallel. This suggests that visual controls and feedback can show users what the AI can do and how it interprets their input.</p><p>This connects to direct manipulation, where users act on visible objects and can immediately see the results. Adaptive interfaces, which change slightly as users interact with them, can learn from user behavior and surface what is most relevant and useful. However, they must preserve consistency because unexpected changes create relearning costs. Overall, smart widgets can make AI understandable through visible controls, immediate feedback, and predictable adaptation rather than requiring users to understand the underlying technology itself.</p><p>Lastly, the “Autonomy Sliders” reading addresses autonomy and control. Offering users opportunities to undo actions, reject suggestions, and reduce the AI’s control allows them to start with lower-risk assistance and gradually increase autonomy as the system earns their confidence. This supports learnability by allowing users to safely experiment with the system while developing a more accurate understanding of when to trust it.</p><p>Together, these ideas suggest that users should not have to figure out how to interact with AI on their own. Smart widgets should be designed from the beginning with users’ cognitive load in mind and should progressively teach them the system’s capabilities while responding to their behavior as they learn to use the interface.</p><h3>Trust Calibration</h3><h4>Jesing Chou</h4><p>What I took away from these readings is that trust in AI should not be treated as a fixed setting. I may trust AI to clean up formatting, but not to rewrite an important argument or make a decision for me. That is why I am skeptical that users can simply choose the “right” level of autonomy. Autonomy Sliders give people control, but the harder question comes first: how do users know how much autonomy is safe to give? Once familiar interactions like selection, undo, or copy-paste become intelligent, they also become probabilistic. The interaction may still feel simple even though the system is making an uncertain interpretation underneath.</p><p>Because of that, I would prefer AI systems to start slightly conservative and earn more autonomy over time. Dwell-and-Spring changed how I think about undo because reversibility is not only about fixing mistakes; it also makes experimentation safer. If I can easily see what the AI did and reverse it, I am more comfortable letting it do more. So for me, trust is less like a permanent setting and more like temporary permission. AI should earn more freedom when its actions are understandable, easy to verify, and easy to recover from, and lose that freedom when the task becomes ambiguous, consequential, or difficult to undo.</p><h4>Anita Tsai</h4><p>In human-AI interaction design, there is a tension between making systems more efficient and preserving human control over their own thinking. When intelligent interfaces are designed to reduce friction and automate as much as possible, they can make it easy for users to rely on the system without fully understanding what it is doing. This can lead to cognitive offloading, where users depend on the system to remember, process, or analyze information for them. Over time, this may create an “illusion of competence,” where users confuse the ease of getting an answer with actually understanding the information themselves. In this sense, designing AI primarily around speed and convenience can create an “efficiency trap”: the system becomes easier to use, but the user becomes less engaged in the process.</p><p>However, designing AI to avoid over-reliance can create problems as well. If a system is difficult to use, inaccurate, or constantly interrupts the user with suggestions that are not helpful, people may stop using it altogether. Earlier agent systems demonstrated how intrusive feedback and limited user control could quickly reduce trust in the technology. We see a similar issue in modern adaptive interfaces when AI makes proactive suggestions or changes that users do not expect. These unexpected behaviors create additional cognitive costs because users have to figure out why the system acted that way and potentially relearn how the interface works. When these experiences happen repeatedly, users may begin to ignore the system altogether. Instead of encouraging appropriate use, unpredictable automation can lead to under-reliance and make the AI less useful.</p><p>Because of this, the goal of human-AI interaction should not simply be to prevent over-reliance or under-reliance. Instead, designers should focus on creating systems that help users develop an appropriate level of trust based on the situation. This means that the amount of control given to the AI should change depending on the system’s capabilities, the user’s needs, and the risk of the task. One way to support this is through features such as “Autonomy Sliders,” which allow users to decide how much they want the system to take over. More broadly, interfaces should make the boundary between human and AI control visible and adjustable rather than treating autonomy as a fixed setting. AI should also communicate its confidence in ways that help users decide when to trust its recommendations, while complex outputs should be broken down into information that users can actually evaluate. Ultimately, good human-AI design is about giving users enough support to benefit from AI while still keeping them in control.</p><h3>Readings</h3><p>Caroline Appert, Olivier Chapuis, and Emmanuel Pietriga, <a href="https://dl.acm.org/doi/10.1145/2207676.2208339">“Dwell-and-Spring: Undo for Direct Manipulation”</a></p><p>Richard A. Bolt, <a href="https://dl.acm.org/doi/10.1145/965105.807503">“Put-That-There”: Voice and gesture at the graphics interface”</a></p><p>Andrej Karpathy, <a href="https://www.youtube.com/watch?v=LCEmiRjPEtQ">“Software is Changing (Again)”</a></p><p>Olga Kokhan, <a href="https://tinkogroup.com/adaptive-user-interface-in-modern-apps/">What Is an Adaptive User Interface? Situational, Contextual, and Liquid UI in Modern Applications</a></p><p>Andrew Miller, <a href="https://andrewships.substack.com/p/autonomy-sliders">“Autonomy Sliders</a>”</p><p>Brad Myers, <em>Pick, Click, Flick! The Story of Interaction Techniques</em> (<a href="https://www.cs.cmu.edu/~bam/ixtbook/index.html">textbook</a>), chapter 1</p><p>Dan Saffer, <a href="https://odannyboy.medium.com/ai-needs-ui-31480100e7d8">“AI Needs UI”</a></p><p>Danqing Shi, <a href="https://arxiv.org/html/2602.11753v1">“Building Intelligent User Interfaces for Human-AI Alignment”</a></p><p>Ben Shneiderman, <a href="https://dl.acm.org/doi/10.1145/800276.810991">“Direct Manipulation: A Step Beyond Programming Languages”</a></p><p>Kashyap Todi, Gilles Bailly, Luis A. Leiva, and Antti Oulasvirta, <a href="https://userinterfaces.aalto.fi/adaptive/">“Adapting User Interfaces with Model-based Reinforcement Learning”</a></p><p>UI for AI Team, <a href="https://medium.com/ui-for-ai/ephemeral-ai-minimal-intrusion-maximum-impact-5e7d86908a81">Ephemeral UI</a></p><p>UI for AI Team, <a href="https://medium.com/ui-for-ai/design-principles-for-future-ai-cebc9155745f">Design Principles for Future AI</a></p><p>Sarah Theres Völkel, Christina Schneegass, Malin Eiband, and Daniel Buschek, <a href="https://arxiv.org/abs/2003.03158">“What is “intelligent” in intelligent user interfaces? a meta-analysis of 25 years of IUI”</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=48de333742a2" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-widgets/researching-widgets-48de333742a2">Researching Widgets</a> was originally published in <a href="https://medium.com/ai-widgets">AI Widgets</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Welcome to AI Widgets]]></title>
            <link>https://medium.com/ai-widgets/welcome-to-ai-widgets-02b85c5c6e16?source=rss-534e55e898c0------2</link>
            <guid isPermaLink="false">https://medium.com/p/02b85c5c6e16</guid>
            <category><![CDATA[interaction-design]]></category>
            <category><![CDATA[ux]]></category>
            <category><![CDATA[hci]]></category>
            <category><![CDATA[ui]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Dan Saffer]]></dc:creator>
            <pubDate>Fri, 28 Aug 2026 11:22:46 GMT</pubDate>
            <atom:updated>2026-09-01T11:11:22.740Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lNasXqJe0bRTXX4t38C_-A.png" /></figure><p>After the success of last year’s <a href="https://uiforai.design/">UI for AI project</a>, I wanted to do another design research project with a group of Master’s students working with new interaction design paradigms for AI. Ambient AI was a possibility for a while, but then <a href="https://www.cs.cmu.edu/~bam/">Brad Myers</a>, the head of the <a href="https://hcii.cmu.edu/">HCII</a> and author of <a href="https://www.cs.cmu.edu/~bam/ixtbook/">Pick, Click, Flick! The Story of Interaction Techniques</a> happened to send an email to some faculty over the summer.</p><blockquote>“I was thinking [about] how GUIs are still important, and need to be more intelligent, instead of being replaced, but new interaction techniques (IxT) are also needed, especially to talk <em>about </em>objects, rather than just doing things <em>to </em>objects.”</blockquote><p>In other words, what if we empowered via AI our existing UI widgets and also came up with some new ones? And thus this class, which Brad and I are jointly teaching this fall, was born.</p><h3><strong>The Widgets Are Probably Older Than You Are</strong></h3><p>The scrollbar you used to get to this sentence was designed before my students were born. So was the menu, the button, the text field, and copy-and-paste. We settled on these interaction techniques in the 1970s and 80s, gave them a light renovation for touchscreens, and then froze them in place like insects in amber. Meanwhile, the most powerful technology of our era arrived, and how did we integrate it into our interfaces? We bolted on a chat window.</p><p>Chat is fine for some things; it’s terrible for many others. Typing and speaking are linear and slow; looking and pointing are parallel and fast. Anyone who has tried to describe “that thing, no, the other thing, the red one, up and to the left” to an AI knows the problem. <a href="https://odannyboy.medium.com/ai-needs-ui-31480100e7d8">Chat won’t replace UI.</a> We can instead see what it’s like when our UI components, our widgets, get smart.</p><p>That’s what our group of graduate students at CMU will spend this fall finding out. Imagine things like selection that understands semantics (“the red cars,” “just the phone numbers in this paragraph”), copy-paste that transforms what it carries, scrolling that knows where you’re going and where you’ve been, undo that can keep your text edits but discard your formatting changes. The intelligence moves out of the chat window and into the components themselves.</p><p>In teams, students will research widgets of their choosing, build the jankiest testable versions of them, put them in front of real people, watch them break, rebuild them properly, and test those too. Then we’ll finalize and document it for the community&#39;s use.</p><p>We’ll be publishing here on Medium throughout the semester, the way last year’s UI for AI study did: research, findings, failures, all of it. If you’re interested in what the evolution of a scrollbar looks like, smash that follow button.</p><h4>The Team</h4><p>Tvisha Arora<br>Hilary Chan<br>Sabrina Cheng<br>Jesing Chou<br>Libby Divers<br>Leyla Erenay<br>Mahala Femovich<br>Aymeric Foyer<br>Gaode Gao<br>Mara Harris<br>Suin Kim<br>Jenny Lee<br>Harry Lin<br>Auboni Poddar<br>Matthew Susko<br>Anita Tsai<br>An Xiong<br>Yuyu Yang<br>Emilie Zhang<br>Chloe Zhu</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=02b85c5c6e16" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-widgets/welcome-to-ai-widgets-02b85c5c6e16">Welcome to AI Widgets</a> was originally published in <a href="https://medium.com/ai-widgets">AI Widgets</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[My Favorite Books of the 21st Century Thus Far]]></title>
            <link>https://odannyboy.medium.com/my-favorite-books-of-the-21st-century-thus-far-2cd300951a11?source=rss-534e55e898c0------2</link>
            <guid isPermaLink="false">https://medium.com/p/2cd300951a11</guid>
            <category><![CDATA[best-of]]></category>
            <category><![CDATA[book-recommendations]]></category>
            <category><![CDATA[books]]></category>
            <dc:creator><![CDATA[Dan Saffer]]></dc:creator>
            <pubDate>Wed, 08 Jul 2026 16:29:41 GMT</pubDate>
            <atom:updated>2026-07-08T16:29:41.442Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mA9i6Br34edyOd1S4EXThQ.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tSl8yqG-qat5V-Yif52TLA.jpeg" /></figure><p>(in alphabetical order)</p><p><a href="https://www.amazon.com/dp/0063286351?tag=odannyboy-20"><strong>A Brief History of Intelligence: Evolution, AI, and the Five Breakthroughs That Made Our Brains — A Neuroscience Framework for Understanding Future Machines</strong></a><strong> by Max S. Bennett</strong><br>How biological evolution can serve as a roadmap for understanding the future capabilities and core limitations of artificial intelligence.</p><p><a href="https://www.amazon.com/dp/0061694940?tag=odannyboy-20"><strong>Anathem</strong></a><strong> by Neal Stephenson</strong><br>A mind-bending sci-fi epic set in a world where intellectual monks must leave their secluded sanctuaries to save the planet from an alien threat.</p><p><a href="https://www.amazon.com/dp/1590518896?tag=odannyboy-20"><strong>At the Existentialist Café: Freedom, Being, and Apricot Cocktails with Jean-Paul Sartre, Simone de Beauvoir, Albert Camus, Martin Heidegger, Maurice Merleau-Ponty and Others</strong></a><strong> by Sarah Bakewell</strong><br>A biography of the 20th century’s greatest existentialist thinkers, showing how their radical philosophies were shaped by friendship, passion, and the horrors of WWII.</p><p><a href="https://www.amazon.com/dp/0060838728?tag=odannyboy-20"><strong>Bel Canto</strong></a><strong> by Ann Prachett</strong><br>An international opera performance turns into a months-long hostage crisis, forging unexpected bonds between captors and captives.</p><p><a href="https://www.amazon.com/dp/0735224064?tag=odannyboy-20"><strong>The Bright Sword: A Novel of King Arthur</strong></a><strong> by Lev Grossman</strong><br>A reimagining of Arthurian legend that follows a young knight trying to pick up the pieces of Britain after King Arthur’s final battle.</p><p><a href="https://www.amazon.com/dp/1250806720?tag=odannyboy-20"><strong>Bring Up The Bodies</strong></a><strong>, Hilary Mantel</strong><br>The sequel to <em>Wolf Hall</em> follows the calculating Thomas Cromwell as he orchestrates the brutal downfall of Anne Boleyn.</p><p><a href="https://www.amazon.com/dp/0375507256?tag=odannyboy-20"><strong>Cloud Atlas</strong></a><strong> by David Mitchell </strong><br>A masterfully structured, genre-defying novel that connects six stories spanning centuries, exploring how human souls and actions echo through time.</p><p><a href="https://www.amazon.com/dp/0525510893?tag=odannyboy-20"><strong>Fleishman Is in Trouble: A Novel</strong></a><strong> by Taffy Brodesser-Akner</strong><br>A newly divorced doctor’s ex-wife suddenly vanishes, leaving him with their two kids.</p><p><a href="https://www.amazon.com/dp/0802124739?tag=odannyboy-20"><strong>H is for Hawk</strong></a><strong> by Helen Macdonald</strong><br>A memoir about a grieving daughter who tames a notoriously fierce goshawk to cope with the sudden death of her father.</p><p><a href="https://amzn.to/3SLXpOo"><strong>How to Live: Or A Life of Montaigne in One Question and Twenty Attempts at an Answer</strong></a><strong> by Sarah Bakewell</strong><br>A biography of Michel de Montaigne that explores how the 16th-century essayist tackled life’s biggest questions.</p><p><a href="https://www.amazon.com/dp/0812985400?tag=odannyboy-20"><strong>Lincoln in the Bardo</strong></a><strong> by George Saunders</strong><br>A hilarious and deeply sad ghost story/historical fantasy centering on Abraham Lincoln’s immense grief following the death of his eleven-year-old son.</p><p><a href="https://www.amazon.com/dp/1473229502?tag=odannyboy-20"><strong>River of Gods</strong></a><strong> by Ian McDonald</strong><br>A Dickensian vision of near-future India on the brink of political and environmental collapse, tied together by rogue artificial intelligences.</p><p><a href="https://www.amazon.com/dp/0062364812?tag=odannyboy-20"><strong>The Ends of The World</strong></a><strong> by Peter Brannen</strong><br>A look into Earth’s past mass extinctions and what they can teach us about our current climate future.</p><p><a href="https://www.amazon.com/dp/0812982622?tag=odannyboy-20"><strong>The Orphan Master’s Son</strong></a><strong> by Adam Johnson</strong><br>A harrowing, darkly satirical, and deeply human journey through the surreal and oppressive reality of modern-day North Korea.</p><p><a href="https://amzn.to/4wbg3OB"><strong>The Searcher</strong></a><strong> by Tana French</strong><br>A retired American cop seeks a quiet life in western Ireland, only to get dragged into a local mystery.</p><p><a href="https://www.amazon.com/dp/0062120409?tag=odannyboy-20"><strong>The Son</strong></a><strong> by Philipp Meyer</strong><br>A multi-generational family saga that traces the brutal rise and fall of a Texas oil dynasty.</p><p><a href="https://www.amazon.com/dp/0316261254?tag=odannyboy-20"><strong>Underground Airlines</strong></a><strong> by Ben H. Winters</strong><br>An alternate history thriller set in a modern America where the Civil War never happened and slavery still legally exists in four states.</p><p><a href="https://www.amazon.com/dp/0307476073?tag=odannyboy-20"><strong>Wild</strong></a><strong> by Cheryl Strayed</strong><br>A memoir about a young woman’s solo, 1,100-mile hike along the Pacific Crest Trail to heal from personal tragedy.</p><p><a href="https://www.amazon.com/dp/1250806712?tag=odannyboy-20"><strong>Wolf Hall</strong></a><strong> by Hilary Mantel</strong><br>Historical fiction that reimagines the rise of commoner Thomas Cromwell in the treacherous court of King Henry VIII.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2cd300951a11" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[We’re All Doing The Same Job Now]]></title>
            <link>https://odannyboy.medium.com/were-all-doing-the-same-job-now-81924e0c80d6?source=rss-534e55e898c0------2</link>
            <guid isPermaLink="false">https://medium.com/p/81924e0c80d6</guid>
            <category><![CDATA[ux]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[design-process]]></category>
            <category><![CDATA[software-development]]></category>
            <dc:creator><![CDATA[Dan Saffer]]></dc:creator>
            <pubDate>Thu, 09 Apr 2026 19:03:08 GMT</pubDate>
            <atom:updated>2026-04-09T19:03:08.387Z</atom:updated>
            <content:encoded><![CDATA[<p>I was a designer for over twenty years, and for most of that time, when someone asked what I did, the honest answer was: I made ideas concrete. I took the fuzzy, ambiguous mess of what a product <em>could</em> be and I turned it into something you can see, click, argue about. That was the job. To paraphrase Herb Simon, I changed existing situations into preferred ones. That was the job. Not the only part of the job, but the part that justified my seat at the table. The PM could describe what a good solution needed to accomplish; I could show you five ways it might work and explain the tradeoffs of each at the level of strategy, interaction, and pixel (and occasionally mm).</p><p>That distinction is dissolving, and it’s dissolving from both sides at once.</p><p>A product manager with good AI can now do the thing that used to be my exclusive territory: make things concrete, fast, and at reasonable quality. They can generate UI mockups, write working code, build prototypes, create user flows. The PM who once had to wait for my availability can now go from “We need to reduce churn in onboarding” to a clickable prototype in an afternoon. They can generate twenty layout options where I might have carefully Sketch’d/Figma-d five. And while each of those twenty lacks the refinement I’d bring, the sheer breadth of exploration surfaces ideas that a more craft-conscious process might miss.</p><p>I don’t blame PMs at all. They’re doing something genuinely valuable. They’re holding business context, technical constraints, and customer data in their head while generating the artifact, rather than translating all of that into a brief and hoping nothing gets lost in the handoff. That’s not a lesser version of design. In some contexts, it’s a better one.</p><p>But before designers panic (or before PMs get too comfortable), the knife cuts the other way too.</p><p>A designer with AI can now do most of what made PMs essential. Data analysis, strategy frameworks, prioritization matrices, competitive analyses, stakeholder communication artifacts… all of it is increasingly templateable, and AI is very good at templateable. The designer who used to have to <em>ask</em> for data can now just get it. The one who couldn’t write a solid PRD can now generate strategic scaffolding that would’ve taken a PM a week. A designer who understands users and the market deeply and can also synthesize business data and generate a business case? That person doesn’t need a PM much at all.</p><p>Now add engineers to the picture, and it gets worse for everyone.</p><p>The traditional product team was a triangle. The PM decided <em>what</em> to build. The designer decided <em>how it should work and look</em>. The engineer decided <em>how to actually build it</em> and then, you know, built it. Each role existed in part because the other two couldn’t do their job. The designer needed the engineer to make the design real, production-grade, performant, scalable. The PM needed the engineer to know what was technically feasible and how long it would take. And the engineer needed the other two because, left to their own devices, most engineers will build something that works perfectly and that nobody wants to use or pay for. (That last bit is a gross generalization but go with it.)</p><p>AI is collapsing this triangle the same way it’s collapsing the designer-PM demarcation. A designer with AI can now produce working code, not just prototypes. Not production-quality code, necessarily, but functional enough to test, to demo, to prove a concept without filing a ticket and waiting three sprints. The gap between “clickable prototype” and “shipped feature” is shrinking. A PM with AI can do the same: go from idea to working software without ever opening a design tool <em>or</em> talking to an engineer.</p><p>And engineers? Engineers with AI can now generate interfaces that are, if not well-designed, at least competent. They can run their own usability analyses, write their own copy, make their own product decisions. The engineer who used to say “just give me the spec” can now generate the spec, build the thing, and ship it. The full stack just got a lot fuller.</p><p>So where does that leave the engineer? In the same place as everyone else, it turns out.</p><p>For some people, this is a shrug. They’ve taken on and done many roles in the process for their entire careers. But based on the hair-on-fire reaction of many, they’re in the minority. We’re all doing the same job now! Ahhhh!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/561/1*f9ZgUSc0gdxREzk2HEYu5w.jpeg" /></figure><h3>Durable Advantages</h3><p>So what’s left, for any of these roles? I keep coming back to three things.</p><p>For the designer, the durable advantage seems to be judgment. Not craft execution (“taste”), which AI is rapidly commoditizing and is <a href="https://marieclairedean.substack.com/p/i-built-63-design-skills-for-claude">being shared open source style</a>, but the knowledge of <em>what good looks (and feels!) like</em> in context. Knowing which of those twenty generated options is actually good. Knowing that the spacing is slightly off in a way that undermines trust. Knowing when to break a convention and when breaking it will just confuse people. Seeing what’s almost right and knowing exactly how to push it the rest of the way. That judgment is built from years of making things and watching people use them; you can’t shortcut it, and AI doesn’t have it. For new designers, education will help, both formal and informal.</p><p>For the engineer, the durable advantage might be systems thinking. Not writing code; AI writes code. Not knowing a framework; AI knows every framework. The engineer’s real value is understanding what happens when ten million users hit this thing simultaneously, when the database migration goes sideways at 2 a.m., when two microservices disagree about the state of the world. It’s knowing where the system will break before it breaks. That kind of architectural judgment is built the same way design judgment is: slowly, painfully, by living through the failures that teach you what the textbook can’t. And like junior designers, new SEs will benefit from advanced schooling and mentorships.</p><p>For the PM, the durable advantage is <strong>organizational power</strong>. Someone has to own the outcome in a way the business recognizes: make the call when the data is ambiguous, take the blame when it’s wrong, fight for resources, say no to the executive who wants their pet feature. Designers love to talk about “killing your darlings,” but knowing something should be cut and having the authority to cut it are very different things. The PM can walk into a roadmap review and kill a feature. The designer can only advocate for it. That gap isn’t about willingness or even judgment; it’s about to whom the org chart hands the knife.</p><p>Side note: The best PMs are also often skilled in organizational navigation, to work cross-disciplinarily to get things done. I would argue, however, this is a skill that is shared with the best PMs and the best engineers, and not solely theirs, unlike organizational power invested in them.</p><p>Organizational power is (as the name suggests) held by the organization, not inherent in any individual. If the organization decides that someone else or some other role should have it, it goes there. In other words, an engineer or designer could be imbued by the organization to have this power. Ergo, the PM role could be most vulnerable, although <a href="https://www.lennysnewsletter.com/p/state-of-the-product-job-market-in-ee9">hiring numbers would suggest otherwise</a>.</p><p>I don’t know where this lands. I’m not sure anyone does yet. But I think the honest version is this: all three roles were always partially defined by the limitations of the other two. AI is removing those constraints. The lingering question isn’t whether AI replaces designers or PMs or engineers. It’s whether the three-role product team was always a workaround for a problem that technology is now solving. Or was it a necessary method of distributing the collective intelligence and cognitive burden of designing, managing, and building products? I guess we’ll see.</p><p><em>Hat tip to Nik Martelaro who made this smarter.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=81924e0c80d6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Future of Apps]]></title>
            <link>https://medium.com/ui-for-ai/the-future-of-apps-a28ae40e5851?source=rss-534e55e898c0------2</link>
            <guid isPermaLink="false">https://medium.com/p/a28ae40e5851</guid>
            <category><![CDATA[apps]]></category>
            <category><![CDATA[business-strategy]]></category>
            <category><![CDATA[ux]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Dan Saffer]]></dc:creator>
            <pubDate>Fri, 27 Mar 2026 18:06:15 GMT</pubDate>
            <atom:updated>2026-03-27T18:06:15.970Z</atom:updated>
            <content:encoded><![CDATA[<p>I opened Spotify this morning to make a playlist, and halfway through I realized I was doing something that might feel, in a few years (months? Time has no meaning anymore), like winding a watch. Not because Spotify is dying. Because “opening an app” will be just one of several ways to interact with what Spotify <em>does</em>, and possibly the least common one.</p><p>The question is not “Will apps survive in an AI world?” Spoiler: they will. At least most will. But apps don’t have a single future. They have six, arranged from completely invisible to fully visible. And a given app might occupy three or four of these positions at the same time.</p><p>This is something our UI for AI team has been circling for the last few months. Two of the directions we’ve been prototyping — an activity-centered AI OS and ephemeral, temporal interfaces — turn out to have direct implications for how apps survive, mutate, or vanish entirely.</p><p>Come, let me walk you, gentle reader, down the spectrum, using Spotify as the through-line because it’s one of the few apps that could plausibly live in most positions at once.</p><h4>1. Gone</h4><p>Some apps will just vanish. Not every app has <a href="https://medium.com/@odannyboy/the-five-ai-moats-297274c67f32">something defensible</a>. If your entire value is a function the AI can absorb wholesale, you’re not a product; you’re a feature that got promoted above its station. Calculator. Unit converter. To Do Lists. Maybe notes. These apps aren’t going to get disrupted or acqui-hired. They’ll become unnecessary, the way a standalone alarm clock became unnecessary once phones existed. Nobody mourned the alarm clock industry.</p><p>This is the logical endpoint of what we’ve been calling the AI OS: an operating system organized around activity rather than applications. When the OS itself learns your patterns, anticipates your needs, and configures your devices around what you’re doing, a lot of single-purpose apps stop having a reason to exist.</p><h4>2. Invisible Infrastructure</h4><p>Spotify provides an API to foundation models, and when an AI assembles a short film for you, it licenses and generates your soundtrack without you ever seeing a Spotify interface. Your listening history, your skips, your 2am sad-song habits (no judgement): that’s the product now. The app is gone but the <em>service</em> persists, hidden behind the AI layer, feeding it proprietary data and the domain-specific intelligence to interpret it.</p><p>That second part matters. The raw behavioral data might eventually migrate up the stack to the operating system; an activity-centered AI OS that watches you skip every jazz track for six months doesn’t need Spotify to tell it you don’t like jazz. But music taste is weird. You might skip a song ten times and then one night it clicks and becomes your favorite track for a year. A general-purpose AI could read that as “They hate this.” A domain-specific model that understands how people actually develop musical taste holds onto it. The interpretive layer, not just the raw data, is what’s defensible.</p><h4>3. Trust Anchor</h4><p>In a world saturated with AI-generated content, some apps survive as a stamp of authenticity. That’s the real recording, not a synthetic interpolation. That’s the licensed track, not something a model made from a prompt. The app becomes less about functionality and more about provenance: a brand that means “this is real.” Think of it as the <a href="https://www.ul.com/services/certification">UL labe</a>l for digital content. Not exciting, but quietly essential, especially in high-risk domains like healthcare, financial services, and legal.</p><h4>4. Widget</h4><p>Spotify shows up on your phone as a contextual sliver, your workout playlist already cued when you walk into the gym, a dinner party mix surfacing when your calendar says guests arrive at 7. You never “open” Spotify. It appears, does its thing, recedes. Pieces of functionality, data, and brand summoned as needed by an AI that knows what you’re about to do before you do.</p><p>This is closely related to what our UI for AI team has been calling Temporal UI: ephemeral interfaces that assemble for a task and dissolve when it’s done. The widget future of apps is what happens when you apply that principle to existing products. Instead of a persistent app sitting on your home screen waiting to be opened, you get a flash of familiar UI that materializes in context, serves its purpose, and gets out of the way. The app is visible, juuuuust enough for what you need.</p><h4>5. Social Space</h4><p>AI can generate you a flawless playlist. It can’t replicate your boo sending you a song that means “I was thinking about you.” Shared playlists, collaborative queues, seeing what someone with impeccable taste is listening to: <em>people stay in apps for other people</em>. This is the future that technologists consistently undervalue because it’s hard to engineer and even harder to optimize. Plus you’ve got all that messy content moderation to do. The app persists not as a tool but as a gathering place, and gathering places survive on vibes, not features.</p><h4>6. The Full App</h4><p>You want to build a playlist from scratch, browsing by mood and genre and half-remembered recommendations from a co-worker whose taste you trust more than any algorithm’s. For that kind of deep work (even if it’s play), you open the whole app, the one that’s been refined for exactly this over years of iteration. This is the app as we know it today. It doesn’t die; it just stops being the default entry point.</p><h3>Follow the Money</h3><p>The spectrum is interesting as a taxonomy, but it gets even more interesting and unpredictable when you try to figure out the business models behind each. It’s not hard to imagine major brands balking at becoming widgets or (<em>quelle horreur!</em>) infrastructure.</p><p>Selling API calls and data licensing to AI platforms is not sexy — it’s plumbing. And take it from a plumber’s son: plumbing is essential and it pays, but few want to make their living doing it.</p><p>When you’re talking widgets, revenue gets genuinely murky. You’re providing contextual value that’s hard to meter and easy to undervalue. How do you charge for a three-second widget appearance? For the privilege of being the brand people trust to verify authenticity? These aren’t unsolvable problems, but the business models don’t exist yet, and “we’ll figure out monetization later” has never failed in the past right? Right?</p><p>The strategic question for every app company isn’t “Which future wins?” It’s “Which of these roles are we going to be great at, which will we begrudingly do, and which do we concede?” Most companies won’t get to choose more than two or three positions on this spectrum, if they get to choose at all. The most familiar revenue sits at the end of the spectrum where you have to <em>convince</em> people to engage, while the most reliable revenue turns you into infrastructure nobody thinks about. Pick your discomfort.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a28ae40e5851" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ui-for-ai/the-future-of-apps-a28ae40e5851">The Future of Apps</a> was originally published in <a href="https://medium.com/ui-for-ai">UI for AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Five AI Moats]]></title>
            <link>https://odannyboy.medium.com/the-five-ai-moats-297274c67f32?source=rss-534e55e898c0------2</link>
            <guid isPermaLink="false">https://medium.com/p/297274c67f32</guid>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ux]]></category>
            <category><![CDATA[business-strategy]]></category>
            <dc:creator><![CDATA[Dan Saffer]]></dc:creator>
            <pubDate>Wed, 25 Mar 2026 04:00:12 GMT</pubDate>
            <atom:updated>2026-03-25T04:00:12.208Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OaCPEiPbqII3kPoFl0tHJQ.png" /></figure><p>In 1958, Royal Crown Cola introduced Diet Rite, the first diet cola, years before Coke or Pepsi had anything like it. Customers loved it. By the early 1960s, it was the best-selling diet soft drink in America. Then the big guys noticed.</p><p>Within a few years, both Coke and Pepsi launched their own diet colas. They weren’t better. They weren’t first. But they had distribution networks that reached every corner store and gas station in the country, brand recognition that made consumers instantly trust the product, and advertising budgets that dwarfed RC Cola’s entire revenue. The company that invented the category watched its profits flow to companies that merely copied it.</p><p>This cautionary tale is almost 70 years old, and it’s the most important thing anyone building AI products needs to understand right now.</p><h3>The Market Eats Ideas for Breakfast</h3><p>In 1986, economist David Teece published a paper called “<a href="https://www.politicipublice.ro/uploads/technological_innovation.pdf">Profiting from Technological Innovation”</a> that asked the question nobody building things wants to hear: <strong>why do the companies that invent things so often fail to make money from them?</strong> His answer was elegant and, if you’re a startup founder, a little terrifying.</p><p><strong>When an innovation is easy to copy, Teece argued, profits don’t go to whoever had the idea; they go to whoever controls what he called the <em>complementary assets</em>: distribution, brand, manufacturing, customer relationships. The stuff that surrounds an innovation and actually brings it to market.</strong> He filled the paper with examples. Xerox invented the graphical user interface and the mouse; Apple built the Macintosh. EMI pioneered the CT scanner; GE and Siemens, with their existing hospital relationships and service networks, captured the market. De Havilland built the first commercial jet airliner; Boeing built the 707 and dominated aviation for decades. The pattern repeats like clockwork: someone invents a genuinely new thing, that thing turns out to be easy to replicate, and the value migrates to whoever already owns the assets you really needed to make it a success. I briefly worked for a litigious company I won’t name, which had this happen to them not once, not twice, but three times until they finally went under.</p><p>If you’re building an AI product, you should read that paper with a knot in your stomach. The current AI landscape might be the weakest appropriability regime (Teece’s term for “How hard is it to copy you?”) in the history of technology.</p><p>Think about what it takes to build most AI products today. You write a prompt, you call an API, you wrap it in a UI you spun up on Google Stitch. The total technical moat of many AI startups amounts to a few hundred lines of code and a well-crafted prompt that a competitor could reverse-engineer over a long weekend.</p><p>I teach students who constantly pitch AI product ideas vulnerable to exactly this: “AI that summarizes legal documents,” “AI that writes marketing copy,” “AI chatbot for customer support.” Useful ideas, some of them. Maybe even decent businesses for a while. But they all fail a simple test: W<strong>hat does this product do that a smart person with access to ChatGPT cannot?</strong> If the answer is “Nothing, but it’s slightly more convenient,” that’s not a business. That’s a feature request, and feature requests get absorbed by platforms.</p><p>In 2023, a leaked memo from a Google engineer made this case bluntly. Titled <a href="https://newsletter.semianalysis.com/p/google-we-have-no-moat-and-neither">“We Have No Moat, And Neither Does OpenAI,”</a> it argued that open-source models were catching up so fast that even billions of dollars in compute might not create durable advantages. Then Stability AI’s CEO pushed back with what might be a sharper insight: the memo’s author misunderstood what moats actually are. <strong>Moats are rarely about core technology. They’re about product, data, distribution, and brand. The things that exist <em>around</em> the technology.</strong></p><p>Somewhere, Teece nodded sagely.</p><h3>The Five AI Moats</h3><p>So what actually works as a moat? Drawing on Teece and a 2024 National Bureau of Economic Research paper (remember when the US government actually sponsored stuff like this?) called <a href="https://www.nber.org/system/files/working_papers/w32474/w32474.pdf">“Old Moats for New Models,”</a> I count five of them.</p><h4><strong>Proprietary Data</strong></h4><p>This is probably the strongest moat of all. If your product generates unique data through usage, and that data makes the product better in ways competitors can’t replicate, you have a genuine flywheel. The keyword is <em>proprietary</em>: data anyone can scrape or synthesize isn’t a moat. Data that emerges from your specific users doing their specific work inside your specific product is one. Competitors can’t get the data without first building the product, and they can’t make the product competitive without the data. A beautiful loop, if you can get it spinning.</p><h4><strong>Distribution</strong></h4><p>This is the RC Cola lesson. Doesn’t matter how good your product is if nobody finds it. If doctors already use your platform, if salespeople already live in your CRM, if developers already depend on your repository, then adding AI to that existing relationship is a fundamentally different proposition than asking someone to adopt something brand new. This is why Microsoft’s Copilot strategy is actually quite formidable. Copilot isn’t the best AI product, but it sits inside Office, which sits inside every enterprise on earth. No startup replicates that.</p><h4><strong>Workflow Integration</strong></h4><p>Your moat isn’t about having an elegant interface (more on this in a moment); it’s about being so deeply embedded in someone’s daily work that leaving becomes painful. Your product holds their history, connects to their tools, and has learned their preferences. Every day of use deepens the trench. An AI product that knows nothing about you will always be less useful than one that’s been learning your behavior for two years, even if the underlying model is identical.</p><h4><strong>Network effects</strong></h4><p>Rare in this space, but powerful. Just look at how many people are still on X/Twitter despite it being a toxic cesspool run by a ketamine-addicted white supremacist. If your product gets better as more people use it (a marketplace that gets smarter with every participant, a collaborative tool where teams generate value for each other), you have something that raw technical capability can never overcome. A new entrant with zero users can’t replicate the value created by a million.</p><h4><strong>Brand</strong></h4><p>This matters enormously in high-stakes domains. In healthcare, legal, and finance, people don’t pick the cheapest AI tool. They pick the one they trust. A startup with a great model but no track record in oncology will have a brutal time convincing a hospital system to adopt its diagnostic tool, no matter how slick the demo looks. This is the EMI story again: GE and Siemens didn’t build better scanners. Hospitals already trusted them, already had service contracts with them, already knew how to buy from them. The complementary asset was the relationship.</p><h3>What About User Experience?</h3><p>I can hear the objection forming. “But Dan, you’re a designer. Aren’t you supposed to argue that user experience is a moat?” It’s true: I am a designer. But alas, UX isn’t a moat. It’s a gate.</p><p>It’s never been easier to copy a great experience. While I believe design taste is real and matters for adoption, it’s not defensible in the way that data or distribution are defensible. If the only thing separating your AI product from a competitor’s is that yours looks nicer, you’re one good design team away from irrelevance.</p><p>“If your product is hard to use, no one will use it.” Oh really? Oracle, Workforce, and SAP would all like a word. Seen a Bloomberg Terminal? When a product’s only advantage is that it’s easy to use, users will leave the moment something easier comes along or the moment a less pleasant product offers them something they can’t get elsewhere, like Bloomberg’s proprietary data. It has always been so.</p><p>However. “UX isn’t a moat” is not the same as “UX doesn’t matter.” Bad UX can kill you, just not for the reason most people think. If your onboarding is so confusing that users churn in the first week, you never accumulate the usage data that powers your flywheel. If the interface is so hostile that teams refuse to adopt it, you never achieve the workflow integration that creates switching costs.</p><h4><strong>UX is The Gate, Not the Castle</strong></h4><p>Every product in the Terrible UX Hall of Fame got past the gate through some other force: institutional mandates, sales teams that sold to executives rather than end users, regulatory environments that left customers no alternative. They survived bad UX because structural advantages overwhelmed the friction. Most startups don’t have that luxury. Your UX needs to be good enough to get people in the door and keep them there long enough for the real moats to form. The mistake isn’t caring about user experience. The mistake is confusing the gate for the castle.</p><p>I give my students a thought exercise now. Imagine that tomorrow, every foundation model becomes free and infinitely capable. Anything you can build, anyone else can build in a weekend. Does your product still have a reason to exist? If the answer is no, you’re betting on model scarcity, and that’s probably a losing bet. If the answer is yes, you’re more likely to be building on something real: a data asset, a distribution channel, a network, a trust relationship, a workflow so deeply integrated that ripping it out would hurt. Those are the complementary assets. Teece was writing about them in 1986, the Google memo was arguing about them in 2023, and they’ll separate the AI companies that last from the ones that wash away.</p><p>RC Cola had the innovation, but Coke had the stores. The same story is playing out right now across thousands of AI startups, and most of them are RC Cola and they don’t even know it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=297274c67f32" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Task Analysis for AI Agents]]></title>
            <link>https://odannyboy.medium.com/task-analysis-for-ai-agents-8f1e85f18ed9?source=rss-534e55e898c0------2</link>
            <guid isPermaLink="false">https://medium.com/p/8f1e85f18ed9</guid>
            <category><![CDATA[hci]]></category>
            <category><![CDATA[task-analysis]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ux]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Dan Saffer]]></dc:creator>
            <pubDate>Sun, 22 Mar 2026 20:34:28 GMT</pubDate>
            <atom:updated>2026-03-22T20:34:28.675Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Z1jLXpzziPpexCQ4ya6jOg.png" /></figure><p><a href="https://ixdf.org/literature/article/task-analysis-a-ux-designer-s-best-friend">Task Analysis</a> originated during the 1960s as a structured way to decompose goals into subtasks, plans, and operations, revealing the decision logic and dependencies that hold a workflow together. In the 1990s, it was adopted into HCI and UX design as a way to understand what people actually do (versus what they say they do), where they struggle, where errors creep in, and where opportunities exist to support or automate parts of the work.</p><p>I’ve watched my students build agents and skip this step, leading to some real magical thinking and brute force automation.</p><p>Task Analysis forces you to understand the as-is workflow before you automate anything. Agents designed without it tend to automate everything, automate the wrong things, or miss critical decision points. <strong>Specifically, it helps you identify where human judgment is truly needed versus where an agent can act autonomously, which is probably the most important decision in agent design.</strong> It also surfaces edge cases and the branching paths that users actually take, rather than the happy paths the designer often assumes.</p><p>Another underrated benefit: it reveals the information dependencies between steps. Agents need to know what data is available at each point in a workflow to make good decisions and Task Analysis maps that naturally.</p><p>Task Analysis is probably best for structured, repeatable workflows where the agent is replacing or augmenting a process that humans already do. I’m thinking claims processing, customer support triage, onboarding flows, invoice reconciliation, compliance reviews, etc. These have knowable steps, clear decision criteria, and definable success states. The task tree is finite and mappable. Also important for anything high-stakes: money, health, legal, or safety decisions…because you need to know every point where the agent could go wrong before you ship it.<br> <br>Multi-step orchestration tasks also benefit heavily. If your agent is coordinating across systems, calling APIs, handing off between tools, or managing sequential dependencies, a Task Analysis reveals the ordering constraints and data handoffs that will break things if you get them wrong.</p><p>What you’re trying to avoid is <strong>Automation Surprise</strong>, where users trust the agent, assume it’s handling things correctly, and then get blindsided later when an edge case reveals that it wasn’t and it kept going anyway.</p><p>You also don’t want <strong>Brittle Automation</strong>: the agent working perfectly under narrow conditions but failing ungracefully when it encounters anything unexpected: ambiguous inputs, missing data, conflicting instructions, or user behavior that doesn’t match assumptions.</p><p>Caveat: I’m not suggesting you recreate the workflow exactly or overfit to the existing workflow. Modeling an existing task too rigidly can anchor your design to the current workflow rather than reimagining it. There’s a real risk of over-fitting to the status quo.</p><p>For complex, open-ended tasks (“Help me plan a trip to Jamaica”), a full Task Analysis can spiral because the task tree is enormous and highly context-dependent. You can spend weeks mapping something that an iterative prototype would have surfaced faster.</p><p>You can still analyze the meta-task (how people generally approach brainstorming, what inputs they need, when they feel done, yadda yadda yadda), but trying to decompose the content generation itself into a fixed hierarchy doesn’t work.</p><p>Caveat Number 2! A lot of agent work is probably a hybrid: some kind of structured spine with generative/adaptive moments layered on top. Something like “process this support ticket” has a mappable workflow, but the reply generation step within it might be creative.</p><p>In case you’re imagining weeks of work, I’m not suggesting that; I’m talking about <em>targeted</em> task analysis. Knowing — often via user research—the likely failure points, key decision points, and important edge cases will get you really far. (Some B2B workflows might take a while.)</p><p>Finally, Task Analysis gives you a natural structure for defining success metrics because each subtask and decision point becomes a measurable checkpoint. You can track completion rates, error rates, time-on-task, and fallback-to-human rates <em>at each step</em> rather than just evaluating the agent’s output as a pass/fail at the end.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8f1e85f18ed9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Existential Designer: Facilitating Meaning Through Interaction]]></title>
            <link>https://odannyboy.medium.com/the-existential-designer-facilitating-meaning-through-interaction-8952c358ec52?source=rss-534e55e898c0------2</link>
            <guid isPermaLink="false">https://medium.com/p/8952c358ec52</guid>
            <category><![CDATA[existentialism]]></category>
            <category><![CDATA[ux-design]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ux]]></category>
            <dc:creator><![CDATA[Dan Saffer]]></dc:creator>
            <pubDate>Thu, 19 Mar 2026 13:45:04 GMT</pubDate>
            <atom:updated>2026-03-19T13:45:04.180Z</atom:updated>
            <content:encoded><![CDATA[<p>Designers like to talk about making meaningful products or using the tools of design to make meaning.</p><p>But like what does that even <em>mean</em>? Existentialist philosophers might offer an answer, and it’s not what you might expect.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gI2cU7tqj1M4iR-XzGkCxA.jpeg" /></figure><p>Existentialism emerged in the nineteenth century with Søren Kierkegaard and Friedrich Nietzsche, both of whom rejected the dominant philosophical tradition’s attempt to build total, rational systems that could explain human existence from the outside. Sort of like the engineering culture of today. They had a shared insistence that existence is lived from the inside, that it is irreducibly personal, and that no abstract system can capture what it means to be an individual facing choices in an uncertain world.</p><p>Their ideas became the foundation for a loose movement that flowered in mid-twentieth-century Europe, particularly in France in the wake of two world wars that shattered confidence in progress, reason, and inherited meaning. It feels like we’re there again, 80 years later.</p><p>The thinkers most associated with existentialism (Martin Heidegger, Jean-Paul Sartre, Simon de Beauvoir, Albert Camus) differ from one another in significant ways, but they share a cluster of philosophies: that existence precedes essence (we are not born with a fixed nature but become who we are through our choices); that human freedom is inescapable and often anguishing; that meaning is not built into the structure of the world but must be constituted through engaged, embodied action; and that philosophy must begin not with abstract categories but with the concrete situation of a person living in the world, confronting mortality, responsibility, and the presence of others.</p><p>In short, to them, <strong>meaning is constituted through human choice, commitment, and engagement with a world that is indifferent to our purposes</strong>. Meaning is not given in advance, or provided by someone else. Like a designer.</p><h3>Meaning through interaction</h3><p>Designers are frequently taught that they are the source of a product’s meaning. After all, they define and arrange the structures, flows, screens, materials, and the technology to make an experience. The user is just someone who receives, navigates, or at best “co-produces” that meaning within tightly bounded parameters.</p><p>Existentialists would call bullshit.</p><h4>Sartre: The user as project</h4><p>For Jean-Paul Sartre, human existence is defined by its <em>projects</em>: ongoing, self-chosen orientations toward the future that organize how we encounter the present. You don’t perceive objects neutrally; you encounter them in light of what you’re trying to <em>become</em>. A text editor means something different to someone writing a novel than to someone drafting a resignation letter, not because of its features, but because of the existential project it’s being used for.</p><p>This goes further than the familiar HCI refrain that “context matters.” <strong>Sartre would say that the user’s project is not merely a variable that modifies how a system is used; it is the condition under which the system has meaning at all.</strong> A system that defines its own purpose too rigidly (“this is a productivity tool, this is an entertainment platform”) forecloses the user’s capacity to take it up into their own project. It treats the user as a kind of meaning-consumer: someone who accepts the designer’s framing rather than generating their own.</p><p>The design implication isn’t that systems should be formless or featureless. (Although AI chat boxes certainly come close.) It is that the relationship between a system’s structure and the user’s purposes should remain open to negotiation. The user should be able to inhabit the system as an expression of their own project, not merely as a participant in the designer’s.</p><p>Generative AI tools like image generators or LLM writing assistants sit right on this fault line. When someone uses ChatGPT to write an essay, the Sartrean question is: whose project is this really? If the user is exploring ideas and using the tool as a thinking partner, they’re taking it up into their own meaning-making project. But if they’re pasting in a prompt and submitting the output unchanged, the system has effectively become the meaning-maker, and the user has become a delivery mechanism. The same tool can function either way. The design question is which relationship the system <em>encourages</em>.</p><h4>Camus: Meaning in the activity itself</h4><p>Albert Camus poses the question most starkly. In <em>The Myth of Sisyphus</em>, he argues that life’s lack of inherent purpose does not make it meaningless; it makes meaning something we must actively constitute through engagement. <strong>His famous conclusion, that we must “imagine Sisyphus happy,” locates meaning not in the achievement of a goal but in the quality of engagement with an activity.</strong></p><p>This is a direct challenge to the outcome-oriented logic that dominates HCI evaluation. When we measure success by task completion rates, time-on-task, or conversion metrics, we treat meaning as residing in the outcome. Camus would direct our attention to the activity itself: <strong>is the engagement itself intrinsically meaningful, or is it merely instrumental?</strong></p><p><strong>When every friction is removed in the name of efficiency, the activity can be hollowed out.</strong> There is nothing left to push against, and meaning drains away. This is something that AI systems have become exceedingly good at. Push the sparkle button, the task is done for you, and you have learned nothing and enjoyed nothing.</p><p>Consider Spotify’s Discover Weekly, which is optimized to surface songs you’ll probably like. It’s efficient, but it can hollow out the experience of actual, accidental, or deliberate musical discovery. The struggle of searching out new music is where Camus would locate the meaning. An AI that removes all resistance from discovery might also remove what made discovery matter in the first place.</p><p>A Camusian design sensibility would ask: Does this interaction have enough texture, resistance, and presence to sustain meaningful engagement? Not as a means to an end, but <strong>as an end in itself</strong>?</p><h4>Kierkegaard: From sampling to commitment</h4><p>Søren Kierkegaard’s distinction between the <strong>aesthetic</strong> and <strong>ethical</strong> modes of existence provides another angle to consider. The aesthetic mode is characterized by immediacy, novelty-seeking, and the avoidance of commitment. A restless movement across surfaces. A great deal of consumer technology is architected to keep users in the aesthetic mode: infinite scrolling, algorithmic recommendations, autoplay, the endless lateral movement across content without depth or commitment. Kierkegaard would likely describe this as a form of despair.</p><p>The ethical mode involves choosing, committing, and taking responsibility for one’s choices over time. This is the kind of existence to aim for, and thus the kinds of products to design.</p><p>The question for design becomes: how do we create pathways from sampling to depth? How do we support the moment when a user moves from browsing to building, from consuming to committing? Kierkegaard has a concept of <strong>The Leap</strong> that might be relevant here. The most important transitions in life, he says, cannot be reached by incremental reasoning or gradual persuasion. They require an act of commitment where you leave behind the logic of the stage you’re in and throw yourself into a fundamentally different way of existing, without any guarantee it will work out. The Leap suggests that the transition from aesthetic to ethical cannot always be made smooth and frictionless. Maybe ever. Sometimes commitment <em>requires</em> discontinuity, a moment of decision that feels like a threshold. Designing away all friction may actually prevent users from reaching the kind of committed engagement where meaning-making becomes possible.</p><p>As one example, AI-assisted coding tools like Claude Code make it trivially easy to generate code you don’t fully understand. You can sample endlessly without ever committing to learning a programming language or framework deeply. The Leap from “person who uses AI to produce code” to “person who understands what they’re building” requires a discontinuous commitment that the tool’s frictionless suggestions actively work against.</p><h3>Authentic choices or bad faith</h3><p>Authenticity and bad faith are among existentialism’s most potent conceptual contributions, and they map onto interaction design with uncomfortable precision.</p><h4>Sartre: Bad faith and the architecture of choice</h4><p>Bad faith (<em>mauvaise foi</em>), for Sartre, is the denial of one’s own freedom, either by pretending one has no choice (“I had to agree to the terms of service”) or by pretending one is radically free of all constraints (“I chose to share my data”). Both moves falsify the situation. <strong>The reality, for Sartre, is that we are always free and always situated: we choose, but we choose within conditions we did not create. Our users are in this same situation.</strong></p><p>Dark patterns are the most obvious case of designed bad faith. They seem like choices, but they actually eliminate genuine alternatives. The user is invited to experience themselves as choosing while the system has already determined the outcome. This is bad faith by design: the system provides the script for self-deception.</p><p>But the problem runs deeper than dark patterns. Any system that naturalizes a particular workflow, that presents only one way of doing things as simply “how it works,” is encouraging a mild form of bad faith. The user stops experiencing the system’s structure as a set of choices someone (or some AI) made and begins experiencing it as the way things are. This is not always harmful: some degree of transparency and habituation is necessary for fluency. But it becomes bad faith when the naturalized structure serves interests the user hasn’t endorsed and alternatives the user can’t see.</p><p>AI-powered “personalization” is one ofmost pervasive engine of bad faith operating today. When Netflix tells you “because you watched X,” it presents algorithmic curation as a reflection of your authentic preferences. But those preferences were themselves shaped by previous algorithmic selections. The system constructs a taste profile, feeds you content that reinforces it, and then points to your engagement as evidence that the profile was accurate. In Sartrean terms, this is a closed loop of bad faith. You’re invited to mistake a constructed pattern for a discovered identity.</p><p>Conversely, systems that support authentic choice make their own structure visible without being disruptive. They allow users to understand that alternatives exist, sometimes just by simply showing alternatives. They demonstrate that the current arrangement is one possibility among others, and that users’ continued use is a choice they are making, not a necessity they are submitting to.</p><h4>Heidegger: Das Man and the pull of the default</h4><p>Heidegger’s concept of <em>das Man</em> (the They/the Anyone) describes the way we tend to fall into doing what “one does,” thinking what “one thinks.” We lose ourselves in the conventional, the average, the unreflective. (AI slop, anyone?) <em>Das Man</em> is a structural feature of social existence. We are always already absorbed in shared practices and norms before we reflectively choose them.</p><p>Platforms operationalize <em>das Man</em> with extraordinary efficiency. The default setting, the trending topic, the notification badge, the social proof metric (“4,327 people liked this”), the generic AI-generated output: these are all mechanisms that draw the user into alignment with what everyone else is doing. The infinite scroll is perhaps the purest expression of <em>das Man</em> in digital form: an endless stream of the anonymous “they,” pulling the user along without requiring any act of choice or commitment.</p><p><em>Das Man</em> is practically the operating principle of AI-driven social media feeds. TikTok’s For You page is an extraordinarily efficient mechanism for dissolving individual choice into the anonymous current of what “one” watches. The user is carried along without ever making a decision, which is precisely the unreflective absorption in <em>das Man</em> that Heidegger describes.</p><p>Heidegger does not think we can or even should escape <em>das Man</em> entirely. It is part of the structure of being human. But he argues that authentic existence involves a particular relationship to <em>das Man</em>: an awareness that one is participating in shared practices, coupled with the capacity to take ownership of that participation rather than simply being swept along.</p><p>Interaction design informed by this concept would not try to eliminate social influence or collective norms from systems. It would instead ask <strong>whether the system allows users to become aware of those norms as norms, to see the current of <em>das Man</em> and choose their relationship to it, rather than being carried along unreflectively.</strong></p><h4>Heidegger: Breakdown as disclosure</h4><p>Heidegger has another concept that is well known in HCI: <strong>Breakdown</strong>, the moment when a tool stops working and suddenly becomes visible as an object rather than a transparent extension of activity. But less considered is its existentialist dimension.</p><p>Breakdown is not just a practical inconvenience. It is a moment of disclosure: the hidden structure of a situation becomes visible. The assumptions built into a system, the dependencies one has developed, the alternatives one has foreclosed, all of these can become apparent in a moment of rupture. In Heidegger’s broader existentialist framework, such moments are related to <em>Angst</em>, which discloses the contingency of our situation and forces us into a more reflective, choosing relationship with it.</p><p>The standard HCI approach to breakdown is to prevent it, and when it occurs, to smooth recovery. An existentialist perspective does not dispute the value of reliability, but it adds a question: when breakdown does occur, should the system use that moment to disclose something meaningful? Or does it simply apologize and route the user back into unreflective absorption? Error states, transitions between systems, onboarding experiences, and even feature deprecation are all moments of potential disclosure, moments when the user could be supported in understanding their situation more fully rather than simply being returned to the flow.</p><p>This could be especially useful with AI systems. When an LLM confidently states something false, and the user catches it, that’s a Heideggerian moment of disclosure. The system’s authority, which had been functioning transparently, suddenly becomes visible as a structure that can fail. The question is whether the design uses that moment to support a more reflective relationship with AI (making uncertainty visible, encouraging verification) or does it scramble to rush the user back into uncritical trust.</p><h4>De Beauvoir: The social structure of bad faith</h4><p>Simone De Beauvoir extends the analysis of bad faith in a crucial direction. For Sartre, bad faith is primarily an individual phenomenon, a person’s self-deception about their own freedom. De Beauvoir shows that bad faith is also socially structured. <strong>Systems of oppression work precisely by making certain groups’ lack of freedom appear natural, inevitable, or chosen.</strong> The subordinated person is encouraged to experience their constrained situation as simply “the way things are,” which is bad faith, but bad faith produced and maintained by social arrangements, not merely by individual psychology.</p><p>This has sharp implications for interaction design. When a platform’s recommendation algorithm consistently surfaces certain kinds of content to certain demographic groups, it is not merely making predictions about preferences; it is structuring the horizon of what appears possible and normal. When a system’s defaults encode assumptions about who the “standard” user is (their language, their physical abilities, their social context), it naturalizes one form of life and marginalizes others. This is socially structured bad faith: the system presents a particular, partial arrangement as universal.</p><p>De Beauvoir’s reciprocity principle is starkly illustrated by LLMs. The meaning-making capacity these tools offer to users depends on training data produced by writers, artists, coders, and researchers, most of whom (like myself!) were not consulted, compensated, or even informed. One group’s expanded capacity for meaning-making (generating text, images, code on demand) is built on the instrumentalization of another group’s creative labor. That’s a textbook failure of de Beauvoir’s ethical test.</p><p>Designing against socially structured bad faith requires more than inclusive defaults or accessibility compliance (though those matter). <strong>It requires asking whose experience is being naturalized, whose alternatives are being rendered invisible, and whether the system’s structure reinforces or disrupts existing patterns of social bad faith.</strong></p><h3>Situated freedom</h3><p>The existentialists (de Beauvoir especially) insisted that freedom is not an abstract property of individuals but a concrete, situated condition that depends on material circumstances and the actions of others.</p><h4>De Beauvoir: Freedom is reciprocal</h4><p>De Beauvoir’s central ethical insight is that my freedom is bound up with the freedom of others. <strong>You cannot genuinely declare your own freedom without also declaring the freedom of those around you.</strong> When you treat another person as a mere instrument, as raw material for your projects, or consider people “NPCs” as Elon Musk and his acolytes do, you diminish not only their freedom but your own, because you are living in bad faith about the social conditions of your existence.</p><p>Applied to interaction design, this principle reframes the meaning-making question as an ethical and political one. <strong>It is not enough to ask whether a system supports meaning-making in the abstract. One must ask: whose meaning-making does it support, and at whose expense?</strong></p><p>A social media platform may empower some users to build audiences, create communities, and develop public identities. These are genuine meaning-making activities. But that meaning-making depends on the labor of content moderators, the attention of other users, and the data extracted from everyone’s activity. If the system is structured so that some people’s meaning-making requires others’ reduction to instruments, to content, to data points, to moderation labor, to a captive audience, then it fails de Beauvoir’s ethical test, regardless of how meaningful it feels to the users who feel empowered.</p><p>The design question becomes: is it possible to structure this system so that one user’s meaning-making does not require another’s instrumentalization? And if not, if some asymmetry is unavoidable, is the system honest about that asymmetry, or does it obscure it?</p><h4>Sartre: The look and the politics of visibility</h4><p>Sartre’s concept of The Look (<em>le regard</em>) describes how being perceived by another consciousness transforms your experience. Under the other’s gaze, you become an object: something seen, judged, categorized. This is not inherently oppressive (mutual recognition <em>is</em> possible), but it becomes oppressive when the look is one-directional: when one party has the power to observe, classify, and define, and the other can only be observed.</p><p>This maps onto the asymmetric visibility structures of contemporary technology with troubling clarity. Surveillance systems, performance monitoring software, algorithmic profiling, and social media metrics all create relationships in which some parties see, and anyone else is just seen, and the seen have limited capacity to look back, to define the terms of their own visibility, or to withdraw from observation. The person living under algorithmic surveillance begins to organize their behavior around managing their profile rather than pursuing their own projects. I feel this in myself every time I post anything publicly, and I bet some of you do too. This is exactly the distortion of meaning-making that Sartre predicts under a permanent, asymmetric Look.</p><p>An existentialist critique of these systems would not simply ask whether surveillance is “justified” in consequentialist terms. It would ask: what does it do to a person’s capacity for meaning-making to exist under a permanent, one-directional look? Sartre’s answer is that it tends to produce either bad faith (internalizing the observer’s categories as one’s own identity) or a reactive, defensive mode of existence organized around managing one’s appearance rather than pursuing one’s projects. Either way, meaning-making is distorted.</p><p>Designing with this in mind means attending to the reciprocity of visibility in a system. Who can see whom? On whose terms? Can users see the systems that see them? Can they contest the categories applied to them? Can they withdraw from observation without losing access to participation?</p><h4>Heidegger: Thrownness and the user who is already somewhere</h4><p>Heidegger’s concept of <strong>Thrownness</strong> (<em>Geworfenheit</em>) names the fact that we always find ourselves already in a situation we did not choose. There’s always history, our bodies, the cultures we live in, and a set of social constructs and laws already in play. We never start from zero.</p><p>As a minor example, my students are thrown into a world where AI writing tools exist. They didn’t choose this, but they must now navigate questions about authorship, learning, and intellectual integrity that previous generations never faced. The AI itself has become part of their Thrownness, reshaping what it means to write, to learn, and to demonstrate understanding. Design decisions made by AI companies have altered the existential situation of millions (billions?) of people who had no say in those decisions.</p><p>HCI has historically tended to treat the user as an abstraction: less a person and more a set of cognitive capacities, task goals, and demographic variables. Thrownness insists that the user is a concrete person who arrives at the system already entangled in a life. They are already tired or anxious or grieving. They already have habits formed by other systems. They already occupy a social position that shapes what the system will mean to them and what it will demand of them.</p><p>Designing with Thrownness in mind means abandoning the fiction of the blank-slate user. It means asking: what situations are users likely to be thrown into when they encounter this system? What histories and constraints do they carry? And crucially: does the system accommodate the diversity of thrown situations, or does it assume a narrow, privileged baseline?</p><p>This is not simply a call for “inclusive design” in the sense of accommodating more user types. It is a deeper philosophical point: <strong>the system itself becomes part of the user’s Thrownness</strong>. Once adopted, it shapes the situation into which the user is thrown tomorrow. Design is not a one-time intervention but an ongoing restructuring of the conditions of someone’s existence.</p><h4>Kierkegaard: The weight of choice</h4><p>Kierkegaard insists that genuine choice is weighty. It matters, it shapes who you become, and it cannot be undone without cost. This stands in tension with the design rhetoric of flexibility, reversibility, and low commitment (e.g., undo).</p><p>There is wisdom in reducing unnecessary friction. But Kierkegaard reminds us that a world in which nothing is weighty, choice itself is devalued. If every decision can be trivially reversed, then no decision means very much, and meaning-making, which depends on commitment, is undermined.</p><p>The question for design is not whether to introduce weight and consequence everywhere, but <strong>whether the system is honest about the weight that already exists</strong>. Trust me: when someone commits years of creative and social work to a platform, that commitment shapes their life. It matters whether or not the platform acknowledges it. An existentialist design ethic would be honest about these stakes rather than obscuring them with the language of frictionless flexibility.</p><h3>Toward an existentialist design sensibility</h3><p>Obviously, this is not a design method. Existentialism resists methodological capture; it is a way of attending to the human situation, not a procedure for optimizing outcomes. But taken together, <strong>these concepts outline a sensibility that can inform and challenge design practice</strong>.</p><p><strong>This sensibility insists that users are not information processors, not customers, not eyeballs, not tapping fingers, and not data sources. They are meaning-making beings whose freedom and dignity are at stake in every interaction.</strong> <strong>It asks designers to take seriously the existential weight of what they build.</strong> <strong>The systems we design become part of the conditions of human existence, shaping what people can choose, what they can see, who they can become.</strong></p><p>The existentialist tradition also offers a crucial corrective to the optimism that pervades design discourse. Not every problem has a solution. Not every friction should be removed. Not every experience should be optimized. Sometimes the most meaningful interactions are the ones that confront us with difficulty, ambiguity, and the weight of our own freedom. You might try, the next time the opportunity arises, to have the courage to let those moments stand.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8952c358ec52" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Imaging AI in 2030]]></title>
            <link>https://medium.com/ui-for-ai/imaging-ai-in-2030-1190179689d1?source=rss-534e55e898c0------2</link>
            <guid isPermaLink="false">https://medium.com/p/1190179689d1</guid>
            <category><![CDATA[carnegie-mellon]]></category>
            <category><![CDATA[forecasting]]></category>
            <category><![CDATA[ux]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[innovation]]></category>
            <dc:creator><![CDATA[Dan Saffer]]></dc:creator>
            <pubDate>Sun, 15 Mar 2026 20:29:35 GMT</pubDate>
            <atom:updated>2026-03-15T20:29:35.708Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*y53lpuqMVfG7R1XaKcH_sg.png" /></figure><p><strong>The UI for AI Team</strong></p><p>It’s foolhardy to predict exactly where AI is headed. That being said, we’re at a genuine inflection point, one where the decisions made in the next few years will determine whether AI becomes something that genuinely helps people or becomes a tool of surveillance, misinformation, and income inequality. If we’re not careful, we’ll be lucky if AI is just an irritating tech layer we have to navigate through.</p><p>But as Alan Kay once famously said, the best way to predict the future is to make it.</p><p>For the past nine months, our team has been working on an initiative we call UI for AI. The premise is straightforward: as AI moves from a chatbot novelty to an integrated layer of our digital and physical lives, the interfaces we’ve built for the past forty years aren’t going to cut it anymore. We’re not tweaking. We’re rethinking.</p><p>We identified five conceptual directions we hope will happen and will matter. None of them is inevitable. But all of them point somewhere interesting. And we’re busy designing and prototyping them. But here’s where we’re aiming.</p><h3>AI for AI: The Orchestration Layer</h3><p>The average knowledge worker today manages a sprawling collection of apps, tabs, subscriptions, and workflows that don’t talk to each other particularly well. They’re the integration layer. That’s not a good use of a human brain.</p><p>What we’re calling the Orchestration Layer imagines an AI that manages and monitors other AIs, dynamically assembling the right tools for a given task, coordinating across them, and auditing the results to make sure they’re actually serving the user’s interests (not the vendor’s). Think of it less like a super-app and more like a chief of staff who happens to know how to use every piece of software ever written. Or you can think of it as a guardian angel, keeping you safe. All watched over by machines of loving grace.</p><p>The design challenge here is one of metacognition. <strong>The interface isn’t about doing tasks; it’s about helping users decide <em>what</em> to delegate versus what to keep.</strong> That’s a genuinely hard problem. Most current tools either do too much (taking over before you’ve consented) or too little (making you configure everything manually). The sweet spot is a system that enhances how you already think rather than replacing it.</p><p>This matters most in high-stakes domains: complex financial decisions, multi-leg travel logistics, managing a household that has actual complexity. Places where the cost of getting it wrong is real.</p><h3>AI OS : The Intelligent Operating System</h3><p>Our operating systems feel like they were designed decades ago. Because they were. Even iOS is almost 20 years old, which is almost as old as some of the UI for AI team members.</p><p>The window-and-file metaphor was a brilliant solution to a real problem: helping people who’d never touched a computer understand what they could do with it. It worked. But we’re decades past that moment, and we’re still dragging icons around a desktop like it’s revolutionary.</p><p>The Intelligent Operating System direction proposes something different: an OS organized around <em>activity</em> rather than <em>applications</em>. Applications still exist, but you only dip into them as needed. The AI OS learns your patterns — the way you cluster certain tabs together when you’re doing research, the deadlines you keep missing because they live in three different platforms — and starts to anticipate rather than react. Files group themselves. Context loads before you ask for it. <strong>Your devices configure themselves around what you’re doing.</strong></p><p>Now, this direction is also where things can go sideways fast. The principle of “automation without interruption” is easy to state and hard to execute. Nobody wants an OS that’s constantly asking for confirmation, but nobody wants one that quietly reorganizes their entire workflow without a trace either. Reversibility matters. Every automated action has to be transparent and undoable. That’s not a nice-to-have. It’s the whole ballgame.</p><h3>IoT 2.0: Agentic Physical Environments</h3><p>Smart homes have been a punchline for a while now. And honestly, the derision is earned. The current state of home automation involves downloading seventeen apps, creating accounts on platforms that will almost certainly be discontinued, and configuring routines that break whenever you update your phone. The promise was seamlessness. The reality is friction with better lighting.</p><p>What comes next (and it’s closer than people realize) is physical environments that actually infer what you need rather than waiting to be commanded. Not “Hey, turn on the kitchen lights” but a kitchen that senses you’ve started cooking and adjusts accordingly. Not a thermostat you program but a space that reads emotional state and makes itself hospitable.</p><p>The key design principle here is <strong>intent over commands</strong>. “I’m cooking” is a meaningful signal. It implies a dozen downstream actions. A well-designed agentic environment figures those out so you don’t have to.</p><p>Here’s the catch, though: “invisible AI” only works if people trust it. And right now, that trust has not been earned. The path from where we are to where this is going runs directly through transparency about what the system knows, how it knows it, and how you turn it off. That’s a design problem as much as a technical one.</p><h3>Temporal UI: Ephemeral and Ambient Interfaces</h3><p>One of the most underappreciated problems in modern computing is the accumulation of digital residue: the browser tabs, the open documents, the half-finished drafts, the notifications that pile up like sediment until your screen is just a monument to everything you haven’t finished.</p><p>The Temporal UI direction tackles this head-on. The core idea is <strong>interfaces that assemble for a task and dissolve when it’s done.</strong> An AI-native browser that detects what you’re working on — say, researching a new project — and clusters the relevant tabs into a structured workspace, highlights likely next steps, and then, crucially, gets out of the way when you’re done.</p><p>The harder version of this idea is what we call “just-in-time assistance”: AI that surfaces the right document as you’re walking into a meeting, or suggests a relevant reference while you’re mid-draft. Not as a notification. Not as a banner. Just… there, when it’s useful.</p><p>There’s a principle embedded here that is genuinely important: <strong>preserve the struggle</strong>. Not all cognitive work should be automated. Some mental effort has value: the wrestling with a problem, the process of figuring something out. Part of good ambient design is knowing what <em>not</em> to do. That’s a harder and more interesting design question than most of what we’re currently arguing about.</p><h3>Tools for Creation: The AI Collaborator</h3><p>The current trajectory of AI creative tools worries us. The natural endpoint of “tell AI what you want and receive a polished output” is a world where creative work becomes a procurement exercise. Where the interesting, generative friction of not-knowing-exactly-what-you-want gets bypassed entirely.</p><p>The AI Collaborator direction argues for something different. AI as a collaborator that requires human participation rather than replacing it. This includes what we’re calling “Rough Draft AI”: tools that generate intentionally unfinished, crude ideas that users have to resolve themselves. The output is scaffolding, not a finished building.</p><p>We have a lot of sympathy for the desire to automate away the blank-page problem. Writer’s block is real. Creative paralysis is real. But the process of working through rough material, of making decisions about what to keep, what to throw away, what to develop, is where taste actually forms. Skip that process entirely, and you might get something polished. You probably won’t get something that’s distinctly yours. Everything creative starts as a mess until it isn’t.</p><p>The target audience for this direction is designers and students, but it applies more broadly. Anyone who wants to use AI as a genuine thinking partner rather than a vending machine.</p><p>These five directions span a lot of territory — the architecture of a house, the tabs in a browser, the OS on your laptop, the rough sketches of a design. But all of them shift the human’s role from managing tasks to expressing intent. The system figures out the context. The kitchen preheats because it sensed “cooking mode.” The OS clusters resources because it understood “researching.” The orchestration layer coordinates five tools because you said “plan my trip.”</p><p>All of them treat cognitive load as a first-class design problem. Ambient, ephemeral, quiet. Technology that supports you without demanding your constant attention.</p><p>And all of them keep the human as the arbiter. Not a passive observer of an automated world, but an empowered conductor of it.</p><p>The goal isn’t to build AI that thinks for you; it’s to build AI that thinks alongside you.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1190179689d1" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ui-for-ai/imaging-ai-in-2030-1190179689d1">Imaging AI in 2030</a> was originally published in <a href="https://medium.com/ui-for-ai">UI for AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>