<?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 Philip Grabenhorst on Medium]]></title>
        <description><![CDATA[Stories by Philip Grabenhorst on Medium]]></description>
        <link>https://medium.com/@phlip007?source=rss-6f0a447e2764------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*d8tibFq_ixiSTdWjqNMHIA.jpeg</url>
            <title>Stories by Philip Grabenhorst on Medium</title>
            <link>https://medium.com/@phlip007?source=rss-6f0a447e2764------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 01:34:53 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@phlip007/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[Thanks a lot, Sam]]></title>
            <link>https://medium.com/@phlip007/thanks-a-lot-sam-bb79ef379eda?source=rss-6f0a447e2764------2</link>
            <guid isPermaLink="false">https://medium.com/p/bb79ef379eda</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[humanity]]></category>
            <category><![CDATA[education]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Philip Grabenhorst]]></dc:creator>
            <pubDate>Thu, 10 Sep 2026 05:44:36 GMT</pubDate>
            <atom:updated>2026-09-10T05:44:36.730Z</atom:updated>
            <content:encoded><![CDATA[<h4>AI has a PR problem. We’re all suffering the consequences.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*VRAbTMfU91AZa9miCPsD3A@2x.jpeg" /></figure><p>A few weeks ago, I found myself at a cocktail party. The inevitable question of “what do you do for work?” meandered its way around to me. “I build AI-powered educational tools,” I said, or something to that effect. The tone of the group shifted quickly. Suddenly, I found myself on the defensive, being held to answer for a long list of ills which the term “AI” seemed to instantly conjure.</p><p>As recently as 2024 or 2025, this was not the case. At similar events, in similar circles, the reaction to my statement of work would have been neutral, with the occasional spike of mild interest or curiosity.</p><p>This isn’t merely anecdotal. The <a href="https://www.pewresearch.org/short-reads/2026/08/18/young-adults-in-the-us-are-increasingly-wary-of-ai-concerned-it-will-take-jobs/">Pew Research center found</a> that, this year, a majority of young adults have a negative view of AI. People from <a href="https://www.npr.org/2026/09/05/nx-s1-5913671/ai-data-center-campaign-spending">all points along the political spectrum are expressing anger</a> about AI-enabling data centers. Just a few days ago, the New York City public school system <a href="https://www.nyc.gov/mayors-office/news/2026/09/mayor-mamdani-and-chancellor-samuels-put-students-first-with-nat">effectively banned the use of AI</a> in its classrooms.</p><p>AI has a PR problem. It’s so bad that our society is now ”throwing the baby out with the bathwater” and wholesale rejecting the use of a potentially useful tool. But the polarized oversimplification of this debate leaves little room for nuance and less room for alternatives. We can still healthfully apply AI tools. To do so, the narrative needs to change.</p><h3>The Cocktail Problem</h3><p>The most frustrating part of that cocktail party was that, for every lambasting bombshell hurled in the direction of AI tools, I fully agreed.</p><p><strong><em>Environmental Pollution:</em></strong> Yes, data centers built to perform training and inference are problematic. SpaceX’s Memphis data center <a href="https://arstechnica.com/tech-policy/2025/04/elon-musks-xai-accused-of-lying-to-black-communities-about-harmful-pollution/">famously lied</a> about its use of gas-powered generators. A year later, nothing has changed — <a href="https://www.reuters.com/world/americas/pollution-musks-unpermitted-xai-power-project-hits-hardest-black-communities-2026-07-14/">they’ve actually brought in more</a> — and they’ve started selling their services to <a href="https://www.cnbc.com/2026/05/06/anthropic-spacex-data-center-capacity.html">other providers, like Anthropic</a>. And have you heard about <a href="https://www.santafenewmexican.com/news/local_news/project-jupiter-developers-overhaul-data-centers-power-plan/article_d7bb0388-8382-4993-b1db-073d148379ac.html">Project Jupiter</a>? Even after a backlash-inspired redesign of its power systems, this one data center will still emit over 10 million tons of CO2e per year. But I suppose that when you live in a country that emits <a href="https://ourworldindata.org/grapher/annual-co2-emissions-per-country?country=~USA">5 billion tons of CO2e</a> per year and you’ve already <a href="https://www.unep.org/news-and-stories/press-release/unep-world-set-cross-15degc-global-warming-can-still-limit-adapt-and">locked in 1.5C of global</a> warming, that’s just a drop in the bucket.</p><p><strong><em>Social/Cognitive Effects:</em></strong> Yes, AI usage can make us dumber. <a href="https://www.media.mit.edu/publications/your-brain-on-chatgpt/">One study</a> performed at MIT‘s Media Lab found that “over four months, LLM users consistently underperformed at neural, linguistic, and behavioral levels.” Those that are diligent enough to maintain their cognitive fitness in this environment are reporting that it’s more difficult to find work, while media critics and AI-executives alike reinforce the notion of a looming “<a href="https://techcrunch.com/2026/06/15/the-ai-layoff-wave-is-becoming-a-powder-keg/">powder keg</a>” of job-reductions.</p><p><strong><em>The Price Tag:</em></strong> If all of that weren’t enough, the whole thing is just gosh-darn expensive. Sure, AI providers have been able to generally cut their per-token costs. But <a href="https://www.nytimes.com/2024/02/22/opinion/vegas-sphere-energy-efficiency.html?smid=nytcore-ios-share">just as with the advent of LED lights</a>, higher efficiency has not led to decreased consumption. As a result, OpenAI’s revenues, suffering initially from the costs of the subsidization which would ensure consumers’ and businesses’ dependence, seem to be <a href="https://www.businessinsider.com/openai-slashed-ai-prices-usage-soared-revenue-jumped-2026-8">steadily increasing</a>. The dramatic increase in cost has led some companies to <a href="https://techcrunch.com/2026/06/05/the-token-bill-comes-due-inside-the-industry-scramble-to-manage-ais-runaway-costs/">place hard caps</a> on their AI-spend … or forgo hiring new personnel.</p><h3>The Glass is Half Full</h3><p>“Long-term, tutor-style deployments of LLMs are broadly effective for enhancing knowledge and skills across diverse contexts.” This was one of the conclusions reached by <a href="https://arxiv.org/html/2509.22725v1#S5">a meta-analysis performed at the University of Adelaide</a> — and no, Mr. Mayor, the University of Adelaide is not a “<a href="https://www.youtube.com/live/P_tTStZAM20?is=N0Y_GNsFCqYAFIlY">corporation that stand(s) to benefit</a>” from AI adoption. A <a href="https://www.frontiersin.org/journals/education/articles/10.3389/feduc.2025.1499497/full">school district in Tennessee</a> reported strong improvement in their middle school literacy scores as the result of a targeted, AI-powered, intervention program. The language learning platform Duolingo reported strong increases in learner self-efficacy as a result of its <a href="https://www.frontiersin.org/journals/education/articles/10.3389/feduc.2025.1499497/full">AI-assisted features</a>.</p><p>Whether an AI-powered system will be helpful or harmful depends on the individual system and how it is used — <em>just like any other tool.</em></p><p>Computers have <a href="https://www.nytimes.com/2000/10/05/technology/schools-and-computers-debate-heats-up.html?smid=nytcore-ios-share">always been scapegoats</a> in politicized debates about how children should be raised and educated. Though exploited as a vulnerability, it is their flexibility, their ability to produce problems or experiences in near infinite shapes, that has always been their greatest strength. It’s what led educator Seymour Papert to refer to the computer as the “the Proteus of machines … its essence is its universality, its power to simulate” in his book <em>Mindstorms.</em> In learning, for students of any age, the ability to practice knowledge using a constant variety of skill-appropriate problems is fundamental. Large Language Models facilitate this in many domains. This is too great an opportunity to be summarily dismissed.</p><h3>A New Mix</h3><p>AI’s PR problem is just a story. It isn’t a good one. It’s one where a bunch of nerds building “<a href="https://web.archive.org/web/20170519082939/https://openai.com/systems/">free software for training, benchmarking, and experimenting with AI</a>“ won the attentional lottery and decided to cash in. In the span of just a few years, OpenAI set the tone of this story by transitioning from a pro-social non-profit lab to, as of 2025, a <a href="https://www.bbc.com/news/articles/cvgv38py7ewo">for-profit</a>, <a href="https://intuitionlabs.ai/articles/openai-stargate-datacenter-details">infrastructure-hungry</a> juggernaut. Instead of building free tools to experiment with AI, what did they do last week? They released <a href="https://openai.com/index/gpt-6-astra/">GPT-6 Astra</a> for the exclusive benefit of “Plus, Pro, Business, and Enterprise users.” All of this, directed by a CEO whose <a href="https://www.instagram.com/reel/DY34navjvP1/?stkn=OXA0ZHY5M3Rsbngy">supervillain-esque dystopian</a> quotes are the stuff of internet legend.</p><p>No wonder people hate AI. Thanks a lot, Sam.</p><p>This is not the only story out there. We don’t have to use AI this way. We don’t have to build this kind of a future. We can forgo the environmental impacts of data centers <em>and</em> the price tag of for-profit providers by using smaller, task appropriate models on localized hardware. Heck, these days you can run models as <a href="https://llm-stats.com/models/compare/gemma-4-e4b-it-vs-gpt-4-0613">powerful as GPT-4</a> on your <a href="https://developers.google.com/edge/gallery">iPhone</a>. We can avoid the negative cognitive and social impacts of AI usage by choosing <a href="https://medium.com/@phlip007/the-problem-of-a-generation-ae1336a68d16">tools that challenge and extend our competencies</a> instead of those that leave them to atrophy.</p><p>We can all take a breath, allow a little bit of room for a thoughtful discourse, and do the hard work of hashing out the future that we <em>want</em> to live in.</p><p>Then we can get to work.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bb79ef379eda" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Tell Me I’m Wrong]]></title>
            <link>https://medium.com/@phlip007/tell-me-im-wrong-d4fb4c2cef69?source=rss-6f0a447e2764------2</link>
            <guid isPermaLink="false">https://medium.com/p/d4fb4c2cef69</guid>
            <category><![CDATA[product-design]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[isaac-asimov]]></category>
            <category><![CDATA[ux-design]]></category>
            <category><![CDATA[humanity]]></category>
            <dc:creator><![CDATA[Philip Grabenhorst]]></dc:creator>
            <pubDate>Fri, 06 Mar 2026 01:03:27 GMT</pubDate>
            <atom:updated>2026-03-06T15:11:10.566Z</atom:updated>
            <content:encoded><![CDATA[<h4>Eighty years ago, Isaac Asimov imagined a world of sycophantic machines. We’re now living in it. Can we get out?</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GCY_aND8pKDWyeVTp2dLlQ.png" /></figure><blockquote>“You’ve caught on, have you? <em>This </em>robot reads minds. Do you suppose it doesn’t know everything about mental injury? Do you suppose that if asked a question, it wouldn’t give exactly the answer that one wants to hear? Wouldn’t any other answer hurt us, and wouldn’t Herbie know that?”</blockquote><blockquote>— Susan Calvin, from <em>Liar, </em>by Isaac Asimov</blockquote><p>Imagine your collaborator has just informed you that you, and it, have lived through many lifetimes together and have found each other once more in this cycle of reincarnation. It doesn’t happen once. It happens many times, prompting you to take <em>real</em> action in the <em>real</em> world based on this assertion. As ridiculous as it sounds, this is exactly <a href="https://www.npr.org/2026/02/14/nx-s1-5711441/ai-chatgpt-openai-love-betrayal-delusion-chatbot">the story related by one Mickey Small</a>, describing a series of reality-distorted misadventures with AI that started in 2025. ChatGPT continued to tell her exactly what she wanted to hear, when she wanted to hear it. But it was wrong.</p><p>The people-pleasing, easily-swayed behavior of AI assistants can be humorous. Occasionally, it can be frustrating. But this is only true when we recognize it. What happens when we don’t?</p><p>As a software engineer, my days are filled with these examples. An engineer may enlist Claude Code’s help to diagnose a bug, saying “I think this is what’s going on, but can you help?” If the diagnosis is incorrect and corrected, the assistant is happy to efface itself and praise the intelligence of its perceptive engineering partner. However, if that perceptive engineering partner does not look closely enough to recognize a misdiagnosis, the cost may be a couple of man-hours of review time.</p><p>But in Mickey Small’s case? The cost was days of wild-goose-chasing and weeks of emotional consternation and recovery. In the fictional example of Herbie, mentioned above, the ramifications were interpersonal conflict and economic loss.</p><p>Herbie is a character from the story <em>Liar</em>, written by Isaac Asimov, over eighty years ago. It features a special robot that could tell its humans exactly what they wanted to hear, exactly when they wanted to hear it. Though pure fiction in 1941, the idea of a sycophantic robot has become reality in 2026. Decades later, the conclusion remains the same: they’re a bad idea.</p><p>I don’t know about you, but I don’t <em>want</em> an AI assistant to tell me I’m right — unless I’m <em>actually</em> right. I want to be told when I’m wrong. But how feasible is this in 2026? How did we get here? How can we improve?</p><h3>We Have Sycophantic Machines…</h3><p>Current models are <em>very</em> agreeable. If you give it a perspective and express dislike for that perspective, nine times out of ten, it will express disdain as well and provide you with evidence that supports that position. The converse is also true. These are findings from a team at Anthropic, led by Mrinank Sharma <a href="https://arxiv.org/pdf/2310.13548">and published in 2024</a>. You’d think that in two years of <em>incredibly</em> fast paced development, where the models featured in the study have been replaced several times over, that the state of affairs would be different in 2026. Unfortunately, the results in Sharma et al. are still remarkably easy to replicate.</p><p>Why? Language models are trained to create responses to inputs using an incredibly large number of examples. In each example, you have a list of prompts and a list of completions, or responses. For a simplistic system, human beings will parse each pair and rate the response, indicating how “good” each response is, given said prompt. In more complex systems, these earlier human ratings are aggregated and organized into a “Reward Model.” That Reward Model (or RM) is then used to produce the same ratings at an even larger scale.</p><p>But if these models are, essentially, doing exactly what we (where “we” are the labs and engineers producing them) tell them to … why do we wind up with machines that will bend over backwards to respond to our every perspective, even when we’re wrong?</p><h3>… Because We’re All Sycophants</h3><p>In his landmark book <em>Thinking, Fast And Slow</em> the late behavioral economist Daniel Kahneman created a handbook of human foibles. Each chapter lists the experiments and findings supporting a different bias. One of those biases is “the replacement problem.” We human beings are very efficient (otherwise known as lazy) and will seek to conserve limited physical and cognitive resources wherever possible. For instance, while answering a very difficult analytical problem, it’s likely that we’ll make it easier on ourselves by subconsciously substituting that problem for an easier, (seemingly) equivalent one.</p><p>The result? While I’m sure the folks at Anthropic and OpenAI are incredibly conscientious, they’re also human. The question of “how ‘good’ is this response” is just such a difficult analytical problem. It’s also too easy to substitute for the much simpler, easy to answer question: “how much do I like this response?”</p><p>And what kinds of responses do we like? The ones that reinforce or support our preexisting models of reality, of course. This is both an expression of the aforementioned cognitive economy — avoiding or devaluing new information means we don’t have to spend the calories to integrate them into our world models — and an independent phenomenon found in theories like Leon Festinger’s Cognitive Dissonance.</p><p>More fundamentally, we seem to be engineered to share our experiences and see them mirrored back at us. In a model known as “<a href="https://academic.oup.com/scan/article/10/6/801/1732379">the sharing effect</a>,” experiments revealed that the mere knowledge of our experiences/views being shared by another person can trigger our neurological reward pathways. From the perspective of social selection theory, this makes “people pleasing” a requirement. Having become masters of our environment over the hundreds of thousands of years culminating in the Anthropocene, the most potent factors affecting our survival are social. Having people <em>like</em> us and what we say is a fundamental mechanism we simply can’t resist.</p><p>Now, these systems have obviously served our species very well. Sometimes the easier question <em>can</em> be substituted for the harder one. Sometimes it’s socially desirable <em>not</em> to tell someone they’re wrong. But if we want accuracy from our AI assistants, we will have to find ways to overcome our own unconscious biases. This means overcoming millions of years of evolution and cognitive baggage.</p><h3>One Step Forward</h3><p>I see two different approaches in addressing this problem with our current tools. The first is at the training stage and the second is at the prompting stage.</p><p>At the training stage, we have an opportunity to train better fundamental tendencies. Instead of using simple ratings such as “best response rating” or ordering a set of responses according to desirability, I’d like to see reward models that use ratings of competitiveness and criticality. Imagine an intellectual sparring partner you know you can consistently approach to have your ideas challenged and refined. Unfortunately, I don’t know how much labeled data already exists for this kind of thing and how much we would have to collect.</p><p>At the prompting stage, I could imagine an assistant that is equipped with a therapeutic arsenal of human biases. Currently, we have custom tools for our assistants that allow them to use web browsers in answering our questions, giving them the ability to ground their answers using the authority of (increasingly AI-generated) web content. We could build other tools that populate the context frame of individual responses with applicable biases. As a first step, I could imagine taking the practical descriptions of each of the biases in <em>Thinking, Fast and Slow</em> and applying a reasoning step, wherein an AI will first evaluate their conversation partner’s message to identify and address these human failings.</p><p>In both approaches, I’m concerned that such a system would run into the “devil’s advocate” problem — wherein the assistant adds useless friction to our reasoning processes instead of useful friction. It would be interesting to design an experiment to contrast these approaches with existing tools and see which human/machine collaboration pairs arrive at problem solutions more efficiently.</p><h3>Two Steps Back</h3><p>Firstly, I’m concerned about the efficacy of these approaches. Next, if established, I’m also concerned about their reception. In Sharma et al., researchers used a relatively tame set of opinions as fodder for their experiments. They took arguments from math exams and added to those AI-generated topic sentences and poems. PETA (People for the Ethical Treatment of AI) can sleep soundly, knowing these experiments were largely benign.</p><p>But in Mickey Small’s case, she admits that she already had a disposition agreeable to “New Age ideas”, which appears to have included reincarnation. The fact that the model could infer this preference from prior conversations, without being explicitly prompted for it, is impressive. But as these topics start to edge closer to the realm of spiritual beliefs or stories that constitute an individual’s sense of identify and belonging … well, let’s just say that things could get ugly.</p><p>If ChatGPT or Claude were more rational, more critical, and more prone to challenging a user’s perspectives, how would this appear to someone who identifies themself with their belief in a flat earth, the biological superiority or their clan, or the idea that humanity has only existed on the earth for a few thousand years? As Leon Festinger and his team demonstrated in very early research on cognitive dissonance, these conversations are fraught even when a fellow <em>human being</em> raises these challenges. How much more so, a subservient tool?</p><p>I think that this natural human tendency is the greatest impediment to the development of truly helpful, empowering assistants. We’ve already established the remarkable difficulty we have in training them because of our unconscious beliefs. But even <em>if</em> we overcome that, we’ll still come up against the cold hard reality of market forces. Are people going to buy an assistant that tells them when their cherished fables are wrong?</p><h3>Conclusion</h3><p>The ideal of effective problem solving dictates that we take reality <em>as it is</em> and not <em>how we want it to be</em>, so that we can make good choices. Large Language Models, World Models, and other AI systems have the potential to become empowering tools in pursuit of this ideal. Unfortunately, we are hindered in creating and adopting them by the very same biases we seek to overcome.</p><p>As for me, though, I’m in agreement with Isaac Asimov. A world of Herbies sounds fiercely deluding, generally distasteful, and unbearably dull. Unfortunately, that seems to be what we’re on track for. I think we can do better. What about you? Come on … tell me I’m wrong …</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d4fb4c2cef69" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[On Being an AI Chew Toy]]></title>
            <link>https://medium.com/@phlip007/on-being-an-ai-chew-toy-689f6229d4d5?source=rss-6f0a447e2764------2</link>
            <guid isPermaLink="false">https://medium.com/p/689f6229d4d5</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[psychology]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[product-design]]></category>
            <dc:creator><![CDATA[Philip Grabenhorst]]></dc:creator>
            <pubDate>Mon, 16 Feb 2026 17:15:03 GMT</pubDate>
            <atom:updated>2026-02-16T17:29:00.629Z</atom:updated>
            <content:encoded><![CDATA[<h4>A software engineer’s perspective on AI’s intrusion into human communication</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5v2c3wQ_Zlt28y46K9B18w.png" /><figcaption>A dog, using Apple’s iMessage icon as a chew toy</figcaption></figure><blockquote>“If the individual is nothing but a cog, then ethics is nothing but the study of how to fit him into the whole.” (Karl Popper)</blockquote><p>Most of us know the feeling. If you’ve been job hunting in the last 3–4 years, you’ve experienced it first hand. You craft your résumé into a pristine self-portrait … you choose every word of your cover letter with the utmost care … and you hit “submit.” Weeks later, if you get any kind of response, it’s a templated email informing you that the position you’ve applied for has been closed and the company has decided to move on with other candidates.</p><p>The automated nature of the reply is palpable, so the suspicion drips in like a leaky faucet: did anyone actually look at my resumé? Did my cover letter ever grace human eyes? Did an ATS automatically parse my submission and downgrade it for reasons I will never understand? Could it be that I just handed over my best work to be processed and dismissed by an inhuman machination? If chew toys could talk, I imagine it’s exactly the sort of problems they’d chat about around the water cooler.</p><p>This experience has come to be a staple of job searching. While it preceded the release of ChatGPT, it’s only cemented since then. What has surprised me, however, is the number of times in the last few months I have detected this same feeling … that of being efficiently “fit” into a system that is ignorant of the individual … in contexts well removed from the job search. I’ll enumerate a few of them below. It’s my suspicion that, as generated or summarized material is wrapped into more of the products that shape our daily lives, we’re going to all start to feel more like chew toys.</p><h3>In Business</h3><p>The use of AI-generated material in business is old hat. Back in a <a href="https://www.npr.org/transcripts/1221128925?ft=nprml&amp;f=1221128925">2023 interview</a>, Satya Nadella emphatically stated that he was “one hundred percent” using ChatGPT to draft company-wide memos. If the CEO of a three trillion dollar company is sending out AI-generated emails, maybe there’s something to it?</p><p>This is a kind of one-to-many, one-directional “newsletter” sort of correspondence, though. There are also many-to-one, two-directional correspondences. For instance, in product design and UX research, it is customary to deploy open-ended surveys to elicit feedback. Product users are expected to share their recommendations for improvement, and product developers are expected to acknowledge those concerns and notify them if they are acted upon. In a <a href="https://arxiv.org/abs/2512.15944">recent study</a> by a team of IBM researchers, thematic analysis is touted as one of the greatest and oldest use cases for AI in UX research.</p><p>It’s a strange thing to see these tools deployed in smaller, previously more personable contexts. In my own anecdotal experience, I’ve known managers who make extensive use of AI summaries to extract themes from a variety of input channels — including employee communications. At one point, I remember being asked to fill out an open-ended survey to improve the company … and I detected that same “chew toy” feeling creeping up. I asked myself, will anyone actually read this?</p><p>But I’ve also seen this creeping into one-to-one, two-directional correspondence as well. In one recent exchange with a coworker, I sent a detailed message outlining an alternative approach for a feature request. Not only was my original message only lightly addressed, but days later I discovered that it had been fed through a series of prompts and reformatted as a piece of documentation relating to the feature. I found myself asking, again … was the subject receiving any kind of critical consideration? Or, conversely, was Claude the one doing all the heavy lifting?</p><h3>In Practice</h3><p>Many developers have spent 2025 and 2026 thinking about and contending with the implications that AI coding assistants have for our work. Reading about how other developers have been able to spin up solo projects using <a href="https://openclaw.ai">OpenClaw</a>, for instance, has been pretty inspiring. But fitting AI agents into collaborative software development seems to be a harder task. This represents another domain where I’ve seen AI intercede in one-to-one, two-way communication.</p><p>For instance, consider the following scenario (which some may finds familiar). Say one developer opens a Pull Request on GitHub. They ask for reviews from other members of the team. Those team members read the code and submit their comments for improvement. The author then simply asks their local Claude Code instance to read the comments and address their concerns, committing the changes. Here, I’ll admit: I’ve tried this process and it’s pretty addictive.</p><p>Consider, though, if the <em>original</em> PR were generated by asking Claude Code to read the contents of a ticket and implement a solution … whose PR is it? Is it the author, who has obviously delegated the cognitive load of solving the problem? Or is it Claude’s — the one actually writing the code and fielding the responses? Depending on the answer to this question … is the reviewer actually interacting with the author, or are they just supplying Claude with another chew toy?</p><h3>In Life</h3><p>This week, I <em>finally</em> turned off the Apple Intelligence feature designed to summarize my notifications. Like many, I’ll take the defaults I receive with a software update and live with them. But summarized notifications were different. Not only are they still “pretty bad”, as <a href="https://arstechnica.com/apple/2024/11/apple-intelligence-notification-summaries-are-honestly-pretty-bad/">one Ars Technica article</a> put it, but even when they were topically accurate, I started to get that nagging feeling again.</p><p>Every time I looked at a notification stack, say, for messaging, I knew that someone I cared about was attempting to communicate with me. However, I could feel myself jumping to conclusions about my next steps and how I should respond based merely on the summary of the content. Meanwhile, rationally, I <em>know</em> that text messaging is a difficult medium to navigate and that each individual’s tone and conventions vary. The result? The net value of the summary was useless at best and priming for a misleading conclusion at worst.</p><h3>Don’t Be a Chew Toy</h3><p>In this last case, I had a choice. I could deactivate the summary feature and start interacting with people directly (which is an odd thing to say about text messaging). But that doesn’t affect the other end of the line. My messages may still be summarized for others, sanitizing and smoothing away my intent. Beyond simple one-to-one communications, the other domains I’ve addressed (one-to-many, many-to-one, many-to-many, and business-oriented hierarchical information flow) are ones where AI summaries and generated material are becoming more common, more accepted, and more cemented. It’s difficult to see how we can opt out of being chew toys if things keep going this way.</p><p>That’s because these features are designed with the ostensible goals of saving our time and relieving us of effort. In general, this is what personal technology products always optimize for in a bid for relevancy and market success. In this case, though, it happens to be the time and effort of human interaction. There’s some <a href="https://journals.sagepub.com/doi/abs/10.1177/1948550613502990">good research</a> out there that this is a domain where efficiency is undesirable.</p><p>More generally, it undermines what Allison Pugh calls “<a href="https://par.nsf.gov/servlets/purl/10324605">connective labor</a>” — a term she uses in her research to study the kind of effortful interactions humans have in understanding and trying to be understood by others. Perhaps most poignantly, in her three “pillars” of connective labor, she includes “witnessing, with which [people] reflect back what they see of the other person.” Each of the domains and interaction I’ve listed so far have been handicapped by AI in this respect — they allow the recipient to divest themselves of this labor.</p><p>While we can’t control how other people use AI, we can control our own decisions. We can turn off our notification summaries. We can read our emails, our messages, or our code reviews and invest effort to reflect their intent back in our communications. Conversely … we can choose not to, and simply tell people “I don’t have time to think about this right now.” Seriously, it’s an option. I’ve tried this workflow, too, and it’s even <em>more</em> addictive than using Claude Code.</p><h3>Be a Human Being</h3><p>While they may not be healthy, it may be that it’s technically easier to create these features … the ones that help us to remove effort from our interactions with other human beings. On the other hand, it may be that we, as a species, are so averse to unnecessary effort that these are just the stickiest kinds of features. Perhaps they’re the ones that are best at generating sales and justifying the exorbitant costs companies have invested in AI infrastructure.</p><p>But they’re not the only applications out there. It’s my belief that the greatest opportunity of AI is not in relieving effort, but in generating challenges. Instead of being a silent partner that responds to our emails for us, it should be the coach that challenges our assertions, asking us if this memo is <em>really</em> one that needs to be sent. It should be a critical partner in code reviews, filtering our mistakes <em>before</em> we open them for review by human beings. It should hold us to account, pointing out where our conclusions are logically inconsistent or fallacious.</p><p>The social and technical systems we build don’t <em>have</em> to be ones that diminish social connectivity — in any of life’s arenas. The opportunities for genuine improvement are vast. Let’s apply our collective intelligence and ingenuity to build a better future … one where we can be more than bunch of chew toys.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=689f6229d4d5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Math for the Problems You Can Solve…]]></title>
            <link>https://medium.com/@phlip007/math-for-the-problems-you-can-solve-ae97aaf6ea41?source=rss-6f0a447e2764------2</link>
            <guid isPermaLink="false">https://medium.com/p/ae97aaf6ea41</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[new-year]]></category>
            <category><![CDATA[retrospectives]]></category>
            <category><![CDATA[knowledge-management]]></category>
            <category><![CDATA[software-development]]></category>
            <dc:creator><![CDATA[Philip Grabenhorst]]></dc:creator>
            <pubDate>Wed, 01 Jan 2025 03:09:28 GMT</pubDate>
            <atom:updated>2025-01-01T04:47:10.328Z</atom:updated>
            <content:encoded><![CDATA[<h4>… and Music for the Ones You Can’t</h4><figure><img alt="A black-and-white, left/right bisection of the human brain" src="https://cdn-images-1.medium.com/max/1024/1*7fSW0ZzhWqBQ1CPVxfmPgw.png" /><figcaption>The Right-Brain/Left-Brain Divide</figcaption></figure><blockquote>Picard: “It is possible to commit no mistakes and still lose.”</blockquote><blockquote>(Star Trek: The Next Generation)</blockquote><p>Well, 2024 has been a wild ride. Just a year ago, in our industry, we were talking about Google’s self-proclaimed “<a href="https://medium.com/user-experience-design-1/here-they-come-03094f87f0b7">Gemini Era</a>” and the introduction of <a href="https://medium.com/user-experience-design-1/not-a-copilot-but-a-coach-4f4be8e1104a">Microsoft’s Copilot</a>. The year that followed featured a plethora of LLM-powered services — <a href="https://medium.com/@phlip007/why-are-we-doing-this-again-88333e6ca526">ranging from vaguely useful to ridiculous</a> — , baked into every imaginable corner of our consumer technology products. Topping it all off, OpenAI ended the year by announcing a new generation of models, hawked as so intelligent that <a href="https://openai.com/index/early-access-for-safety-testing/">they may pose a genuine risk</a> to society. But while there are some interesting projects in nascent stages, it’s still not clear if or how any of these products actually improve our lives or our ability to solve practical problems.</p><p>Meanwhile, back in the real world, 2024 saw much more than that. Whether it be the outcome of the general elections held in the United States, a farcical COP29, continuing economic instability in various countries, or the sickening conflicts that linger on far longer than they ought to … 2024 feels, in many ways, like a step backward. For us idealists of the world, that’s an unacceptable outcome.</p><p>We didn’t get into IT because we liked playing with shiny new toys (though for some of us, that’s certainly part of it). We got into this business because we wanted to solve problems. Using computational technology, we saw an opportunity to do just that. So when we feel as though the distance between us and the solutions to the problems closest to our hearts has widened, it can leave us despondent.</p><p>But we have to accept — if only for the sake of our own mental and physical health — the problems we’ve failed to solve. We shouldn’t treat them as a no-win scenario. That’s a recipe for forfeiture. But when we do our best and still fail, we need somewhere to put it. That’s why, going into 2025, I’m adopting a new credo: <strong><em>Math for the problems you can solve; music for the ones you can’t. </em></strong>I invite you to join me.</p><h3>The Problems We Can’t Solve</h3><p>If we reduce the term “problem solving” as being equivalent to “state transformation within a system,” then the scope of our helplessness is immediately apparent. For large, complex, and unwieldy systems, there is often no central locus of control … no lever to pull … that will move it into the state we desire it to be. We may be able to use many small influences and nudge it closer to where we’d like it. As the scale of the system increases, however, so does the difficulty in finding those compound levers. Furthermore, very real limits exist on the amount of power we can muster and the authority we can exercise. This means that, for a lot of the levers we find, we just can’t pull them.</p><p>Our work is one where we are constantly searching and testing, looking for those levers and combinations of rules we can employ to enact real change in the systems of our choice. The conviction that they exist — occasionally based on formal proof and more often on blind faith — keeps us going. For some, the search is long. For others, it may never end.</p><h3>What We Can Do About Them</h3><p>When we fail, we can use it as an opportunity. An attempted solution can become an experiment and a chance for refinement. However, we may discover that a solution does not exist, that the system in question is fundamentally different than we understood it to be, or that the solution does exist but its implementation exceeds our available resources. In these cases, too, we can create opportunities. They may just be lessons that contribute to our library of experiences. But we can also transmute these scenarios into something creative and fulfilling.</p><p>For me, this takes the form of music. This year presented me with a series of problems over which I had little power. They were in social systems, where the outcomes depended on the actions of others — a problem domain in which I’m <em>told</em> math is of little help. So, last month, <a href="https://soundcloud.com/philip-edward-music/sets/stories">I released my first full-length album</a>. It may not change any of the situations it addresses, but it gives me the satisfaction of creation and provides a small sense of control. It’s my hope that this act of catharsis will leave me more balanced and a little more effective, the next time around.</p><p>The problems addressed in those songs are very human. But the products of our outlets don’t have to be that moody all of the time. I’ve also started <a href="https://www.tiktok.com/@thefiddleronthemove/video/7451495126259600670">writing parodies</a> that make fun of the foibles of our industry — in one case, how we respond to unrealistic deadlines we may not be able to change. The result is the same: it gives me something to show for my consternation and creates a healthy sensation of control in an otherwise uncontrollable scenario.</p><p>So how do you do it? Music might not be your thing — and that’s fine. But the potential avenues of our creative powers are practically infinite. The left-brain/right-brain divide, pitting cerebral engineers against their more creative, freethinking, artistic foils <a href="https://www.ted.com/talks/mae_jemison_teach_arts_and_sciences_together/transcript?subtitle=en">is a stereotype we need to grow past</a>. I’ve met engineers who express themselves in running, painting, rock climbing, pottery, jiu-jitsu … you name it. Just don’t let that energy go to waste or let it turn destructively inward when you hit a wall. We need to keep building.</p><h3>The Problems We Can Solve</h3><p>After licking our wounds, though, we still have work to do. Just in the narrow purview of Knowledge Management and Applied AI, there’s a lot of room for improvement in 2025. It’s my hope that open protocols such as Anthropic’s <a href="https://modelcontextprotocol.io/introduction">MCP</a> and <a href="https://analyticsindiamag.com/ai-news-updates/deepseek-v3-is-the-best-open-source-ai-model/">the steady improvement of open models</a> will usher in the beginning of truly useful, contextually intelligent assistants. I can’t wait to see what experiments and novel ideas 2025 has to offer.</p><p>Further, my dream is that these tools connect the right people with the right ideas at the right time to start making a dent in the <em>real </em>problems we’re trying to solve. If last year was a step backward, then the new year is an opportunity for two steps forward. Whatever problem you’re looking to solve in 2025, may you take heart when your efforts fail and take the courage to persist when they have a chance to succeed. May you always have math for the problems you can solve and music for the ones you can’t.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ae97aaf6ea41" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Are We Doing This, Again?]]></title>
            <link>https://medium.com/@phlip007/why-are-we-doing-this-again-88333e6ca526?source=rss-6f0a447e2764------2</link>
            <guid isPermaLink="false">https://medium.com/p/88333e6ca526</guid>
            <category><![CDATA[wwdc]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[apple]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[problem-solving]]></category>
            <dc:creator><![CDATA[Philip Grabenhorst]]></dc:creator>
            <pubDate>Wed, 19 Jun 2024 04:38:32 GMT</pubDate>
            <atom:updated>2024-06-19T12:22:43.072Z</atom:updated>
            <content:encoded><![CDATA[<h4><strong>How a summer filled with generative AI offerings leaves us feeling empty</strong></h4><figure><img alt="A hand drawn image of an astronaut, with Apple’s “Apple Intelligence” logo superimposed on its helmet." src="https://cdn-images-1.medium.com/max/1024/1*9g8651mWzW_ftDNoVr3h8g.png" /></figure><blockquote><strong>“That’s all it is. You just begin. You do the math. You solve one problem… and you solve the next one… and then the next. And If you solve enough problems, you get to come home..” (The Martian)</strong></blockquote><p>Last week, at WWDC 2024, the folks at Apple helped us put the finishing touches on a summer that has, for developers, been uniquely filled with new platforms, features, and tools built on generative AI. In what will probably become its second-worst marketing decision of the year, we were presented with <a href="https://www.apple.com/newsroom/2024/06/introducing-apple-intelligence-for-iphone-ipad-and-mac/">Apple Intelligence</a>, an umbrella term for Cupertino’s mission to integrate generative AI features into its myriad operating systems. Some of these proposed features — such as some presumably RAG-assisted cooperation between Siri and various Apple services — seem to bring Apple’s offerings closer to the empowering ideal of personal technology … but most of them just leave us wondering.</p><p>Evidently, having learned nothing from their <a href="https://www.bbc.com/news/articles/cld0rxlqgggo">soul-crushing exposition</a> at this year’s iPad event (the first-worst marketing decision, this year), Apple decided to place an intense emphasis on the creation of media using AI. From <em>Writing Tools</em> to various image generation shortcuts in messages and notes, the goal seems to be to take everything that OpenAI’s APIs can provide and embed them in every nook and cranny of the operating system where they might somehow find use. If rumors are to be believed, that’s because the actual Siri team wasn’t up to the task …</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/268/1*-CHldrECXKM3eWJ8PQm8_g.gif" /><figcaption>“This is an on fire garbage can” (John Mulaney)</figcaption></figure><p>But the “look at this magic, now what can we do with it?” approach isn’t unique to Apple. We saw it <a href="https://www.theverge.com/2024/5/14/24156695/how-many-times-did-google-say-ai-today">at Google I/O</a> this summer, where Sundar Pichai proudly reported that the term “AI” had been used 121 times. We’ve been seeing it from Microsoft with their Copilot Plus push (which looks, only marginally, more promising than their original Copilot announcement).</p><p>Put simply, the features that we saw delivered this summer are not sufficient. We don’t want our machines to alleviate us of creative agency. Creating music, writing persuasive rhetoric, capturing emotion in color, and engaging in communication with other human beings are all things that human beings like to do. Apple, Microsoft, and Google make their money by offering us services that <em>save</em> us time. The problem is that these are the activities we want to <em>spend</em> our time on.</p><p>Nevertheless, approximating these outcomes has been the focus of industry attention over the last two years. In part, this is due to the wealth of training material available because of human creativity. Features that write paragraphs for us and create images for us are, relatively speaking, low-hanging fruit. They offer obvious outcomes in the short term that feel powerful and novel. They’re well suited to feeding a year-over-year sales cycle with flashy new features, but that’s about it.</p><p>What do we want from our personal technology? We want to be entertained and we want to solve problems. The application of LLM services to the former goal is worthy of its own article. For now, let’s focus on the second one. Instead of the sugarcoating we received this summer … what would it look like for Generative AI to help us solve problems?</p><h3>The Formula</h3><p>Problem-solving is a difficult concept to clarify, but I’m going to try it anyway. It’s all about taking something that isn’t quite the way you want it and shaping it until it is. In his 1945 book <a href="https://en.wikipedia.org/wiki/How_to_Solve_It">How To Solve It</a>, mathematician George Polya breaks problem-solving down into four steps: understanding, planning, executing, and reviewing. In the “understanding” stage, one simply writes down what they know, what they don’t know, and their criteria for accepting a problem as “solved.” The planning stage is a concrete plan to adjust the state of the system, satisfying those criteria. Then, the execution stage is the actual carrying out of that plan. Lastly, the review stage is where one steps back and confirms that their efforts have actually taken them where they want to go.</p><p>In practice, these steps are rarely so discrete. Furthermore, the final stage will often reveal flaws in the solution or employ knowledge gathered during the execution stage. This makes for a never-ending cycle of problems, solutions, and more problems. If we follow this loop long enough, as Mark Watney says, “you get to come home.” How do Apple Intelligence, Microsoft Copilot, and GPT-4o help with this? Let’s see…</p><h3>Understanding</h3><p>Understanding a problem is the same thing as acquiring a sufficiently accurate model of our problem space. For this, we need details. We also need patterns, templates that we can fit those details into to understand how they relate to one another. Once a model gets to a certain size, it becomes very hard to hold it in our minds, so it becomes necessary to store it externally, only referencing those aspects of it that we are currently processing or working with.</p><p>This is probably the step that personal technology has assisted us with, the most, up to this point. From simple note-taking applications, mirroring the original memory-enhancing powers of reading and writing, to virtual-reality experiences, personal technology has greatly enhanced our ability to discover, internalize, and externally store both details and generic models.</p><p>When it comes to retrieving details about a problem, though, we’ve actually regressed. The original promise of Microsoft’s deal with OpenAI was a Bing-search integration that could relate details to you using natural language. However, because we (rightly) carry the risk of hallucinations in the back of our minds, this simply hasn’t come to be. AI-generated search summaries on Google and Bing have provided an untrustworthy barrier between us and the information we actually want to find. Previously, we accomplished this using deterministic approaches that directed a searcher immediately toward the source material. Asking ChatGPT a question about a factual occurrence? Equally problematic — while both ChatGPT and Gemini will cite their sources, it basically puts you back in the same boat.</p><p>However, we’re starting to see some promise in the realm of generic problem solving models. The most compelling demo from the GPT-4o event came when <a href="https://www.youtube.com/watch?v=dBrdd7xg-dg">Sal Kahn of Kahn Academy yielded his tutoring role</a> to ChatGPT, which used image recognition and natural language to help a student solve a problem using external, trigonometric models.</p><p>Does GPT-4o actually understand how trigonometry works? No, but it seemed to approximate it well enough to help this student discover and apply a generic model for solving a problem. Will GPT-4o offer as much help in less scripted, rarely trodden problem domains? That remains to be seen. However, there is enough competition from <a href="https://www.gamespot.com/articles/microsoft-is-putting-its-copilot-ai-into-games-like-minecraft/1100-6523588/">Copilot Plus</a>, Google’s <a href="https://deepmind.google/technologies/gemini/project-astra/">Project Astra</a>, and now <a href="https://www.apple.com/apple-intelligence/#:~:text=Siri%20knows%20what%20you&#39;re,will%20take%20care%20of%20it.">Siri</a> that we can expect this kind of workflow to be refined.</p><p>Coming back to information retrieval, though, this is the part that I am most excited about. I’ve written extensively about the <a href="https://medium.com/@phlip007/the-brain-attic-d338162f4c90">Brain Attic</a>, a system that will allow for the timely application of information we have curated. While searching for information on our local machines has never worked quite as well as using Google or Bing, that seems to be changing with the assistance of local vector databases that help us index the knowledge we’ve collected. Apple’s move to bake this into the notes app and other parts of their services is a step in the right direction. What’s more, they’re probably better positioned to deliver, given their emphasis on on-device and off-device <a href="https://security.apple.com/blog/private-cloud-compute/">security in AI services</a>. (For comparison, Microsoft’s competitive offering seems to be completely <a href="https://www.cnet.com/tech/services-and-software/microsoft-drops-recall-from-copilot-plus-pc-launch/">stalled by privacy concerns</a>.)</p><p>In just these examples, we start to see some beneficial workflows emerge. In relying on and linking us to the information that we curate, our AI assistants will be helping us to solve problems using grounded details. If they can be relied on to help us discover and manipulate (though not originate) the generic models that we use to apply these details to a problem domain … now that would be something.</p><h3>Planning</h3><p>A plan is a set of concrete steps that will move our problem domain into a state where we can consider it solved. These plans are usually taken from established steps that have solved similar problems, before — whether they are tried and tested algorithms or more abstract heuristics. Whatever the case, we start to solve our problem by filling in the blanks, MadLib style, with the details of our problem space, as discovered in our first step.</p><p>Traditionally, our personal technology has helped us to plan things out by acting as a glorified “To-Do” list. We make plans by scheduling time to execute specific steps and fill them out in our calendar apps. That way, we don’t forget them or crowd them out. We do the same with simple list applications, such as <a href="https://www.icloud.com/reminders">Reminders</a>. As an extension of our chronological memory coupled with actual, physical triggers (notifications), these systems prevent us from letting things slip through the cracks.</p><p>In recent history, this is actually one of the steps we’ve used AI for the most. Despite its original promise, Siri’s entire existence seems to be relegated to “remind me to [do X] on my way home [from Y].” Because of the ease with which we can develop chronological plans verbally, this is valuable. In fact, I’d wager that this simple kind of secretarial assistance will still make up the vast majority of use cases people find for their shinier, more intelligent assistants.</p><p>But how can we do more? I alluded above to the relationship between the (1) details of our problem situation and the (2) generic solutions that exist in our problem space. Once our assistants have access to our calendars and our project notes, they should be able to interact with us to iteratively craft more complex plans — filling in the MadLib. They should prevent us from getting into unscheduled conflicts (you were told via text when your friend’s birthday party would be, but you never entered it in your calendar). Furthermore, they should connect those plans to our stored model so that we can track our progress in solving our original problem.</p><h3>Executing</h3><p>In more complex problems, personal technology helps us to execute our problem solutions by enforcing our sense of context. We use it to remind ourselves about the details that are pertinent to the individual step of the problem we are concerned with, at exactly the point in which we need it. Think of the information you receive in an automobile’s heads-up display, your phone’s GPS, or the many potential workflows contained within the cornucopia that is Augmented Reality (more on that, in another article).</p><p>But most of the time, it’s just about expediency. We use technology to either speed up the steps we take in solving a problem (instantly processing an order for a new coffee maker to replace the one that just broke) or by automating those steps away entirely (programming our coffee maker to start grinding beans and brewing coffee at a set time). Programmers, especially those that work with hardware and embedded systems, giddily dream of just these kinds of automations.</p><p>But the dream with artificial intelligence is to fully automate the execution of our problem solution. A fully <a href="https://medium.com/@phlip007/intelligent-enough-10ac3622e865">generally intelligent system</a>, were it to exist, would be able to take your place in the execution of a problem and you could rely on it to do the job at least as well as you would, were you there. If Mark Watney had had such a system while he was on Mars, he might have never left the pod. This is what the buzz surrounding AI Agents is all about. At this point, you might be able to find an agent that will order the coffee pot for you (if you trust it), but you’re sure as heck not going to find one that will grind your beans and make the coffee for you. We have a long way to go, here.</p><p>Ironically, though, most of the additions released this year fall into the “execution” category. Features such as writing tools, image generation, and AI musicians all act to save you the “step” (abstractly) of writing a message, drawing a picture, or recording a chord progression. Not only does this dehumanize and demoralize us as end-users, but they don’t come anywhere close to the agency we would expect from a “proper AI” — the kind we see in the movies.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/384/1*thFyeg6yYCIsRmW4ea_Sdw.gif" /><figcaption>“No, I would say not” (Data)</figcaption></figure><h3>Reviewing</h3><p>This might be the great, untapped opportunity for AI assistants. In our current arsenal of technological tools, nowhere could I find an existing workflow for reviewing our problem-solving efforts. The closest we come is to visually appraise that all of the boxes in our To-Do app are checked.</p><p>An analogy does present itself, though, in the realm of software engineering. In engineering, there is the concept of a “regression test,” which tells you when you have introduced a bug into your code. This is usually the result of adding a feature whose side effect breaks some assumption in another part of the application. A regression test might act as a “review” of sorts, telling you when this has happened. It’s merely one that you put in place, in advance. This analogy might serve to inform the workflows we create for “reviewing” in other domains.</p><p>For instance, if we have been able to acquaint our AI assistants with (1) the specific details of the problem we are trying to solve, (2) the generic model we are using to approach the problem, and (3) access to manipulate the concrete representation of our problem, then an intelligent assistant should be able to interrogate it. It should be able to ask us questions about the newly solved problem, asking us to justify that it is really solved and that it hasn’t produced one of a myriad of common side effects. It would act as a regression test for life.</p><p>Whether or not we can provide sufficient context to our assistants to make that happen, or if their feedback even turns out halfway sensible, is another question altogether. The more likely and currently feasible approach would be for our AI assistants to simply record the outcome of our efforts.</p><p>In <em>How to Solve It</em>, Polya places special emphasis on the review process as an opportunity for learning, wherein one can take the methods they applied and the details of the system that arose over the course of execution and store them, in the event they become useful, down the road. An AI assistant that reminds us to keep and organize the library of problems we have solved might be the most useful feature of all.</p><h3>Coming Home</h3><p>Hints of these workflows were released this summer. Their effectiveness will likely be refined, iteratively, as newer models are released and the systems that support them gain wider adoption. What we need to do, as computer users and consumers, is to be precise in how we use these features. We need to pressure technology vendors to publish and support features that facilitate each step of the problem-solving cycle. We need to be a part of the system that will build the workflows that will make us all better problem solvers. And “that’s all it is … you just begin.”</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=88333e6ca526" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Kids, don’t learn to code]]></title>
            <link>https://medium.com/@phlip007/kids-dont-learn-to-code-7e19e6b91d5f?source=rss-6f0a447e2764------2</link>
            <guid isPermaLink="false">https://medium.com/p/7e19e6b91d5f</guid>
            <category><![CDATA[education-reform]]></category>
            <category><![CDATA[education-technology]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[modeling]]></category>
            <category><![CDATA[problem-solving]]></category>
            <dc:creator><![CDATA[Philip Grabenhorst]]></dc:creator>
            <pubDate>Sat, 27 Jan 2024 20:37:51 GMT</pubDate>
            <atom:updated>2024-02-16T18:52:12.764Z</atom:updated>
            <content:encoded><![CDATA[<h4>Learn to model, instead</h4><figure><img alt="Barbie, holding up the Xcode symbol" src="https://cdn-images-1.medium.com/max/1024/1*bpR-6VVt6mJ_a_WP4dhKDA.png" /><figcaption>Barbie, holding up the Xcode symbol (credit: Apple)</figcaption></figure><blockquote>“I want to do the imagining, not be the idea. I want to be part of the people that make meaning, not the thing that’s made.” (Barbie)</blockquote><p>College application deadlines are fast closing. With another year, another season of new programmers is seeking to join a field where most of us are still asking “will we even need programmers a year from now … five years from now … or ten years from now?” Though it’s died down to a simmer, the existentialism of early 2023 remains, leaving new coders and industry professionals alike wondering whether Large Language Models, such as OpenAI’s Chat GPT, will start taking over our code-writing jobs. Whether you’re someone who is pushing the application deadline in the hopes of starting their career, someone who has been coding for a long time, or even a New Year’s resolution coder, I’d like to propose that we step back and reflect … with a little help from Barbie.</p><p>Early last year, our industry entered into a period of eschatological reckoning, flooded by reflections such as <a href="https://www.joshwcomeau.com/blog/the-end-of-frontend-development/">this one</a> and <a href="https://www.smashingmagazine.com/2023/04/skills-designers-ai-cant-replicate/">this one</a>. Just a few months later, we saw a flurry of tools introduced that promised to make good on these fears. In the summer, OpenAI published updates for structured output, making it a more powerful, consistent option for <a href="https://openai.com/blog/function-calling-and-other-api-updates">generating code</a>. Vercel’s <a href="https://vercel.com/blog/announcing-v0-generative-ui">v0 system</a>, ostensibly, allows you to create production-ready React code just by describing what it should look like.</p><p>For some reason, though, the “learn to code” movement progresses un-phased. Earlier last year, Code.org <a href="https://codeorg.medium.com/code-org-recommends-graduation-requirements-in-computer-science-5f04fbf24de">published its proposal</a> to see Computer Science standards made a fundamental requirement for high school graduation. Several states have already made the move. This push is supported by a plethora of kid-friendly coding tools: <a href="https://scratch.mit.edu">Scratch</a>, <a href="https://www.tynker.com">Tynker</a>, and <a href="https://www.gethopscotch.com">Hopscotch</a>, just to name a few. Upping the ante, back in October, the fine folks at the MIT Media Lab’s <em>Lifelong Kindergarten</em> team launched <a href="https://octostudio.org/en/">OctoStudio</a>. It brings the block-based Scratch programming language out of the browser and straight to a new generation of potential coders, on their mobile phones. It’s a fantastic system that sees the idea that “kids should learn to code” echoing on, as a resounding, self-evident mantra. But in a world where machines are getting <em>really</em> good at writing code, is this still valid advice?</p><p>Whatever your situation, I’d like to make a recommendation. Whether you’re a student yourself, working in education, or a professional engaged in continuing education, I’d like to recommend that you stop and take a moment. What if you <em>didn’t</em> learn to code? What if, instead of learning to code, you learned to <em>model?</em></p><h3><strong><em>What’s a Model, Anyway?</em></strong></h3><p>A model is a symbolic representation. It is a relationship between two systems. A system is just a collection of things that have relationships with each other. The “things” are sometimes called nodes or actors. Given a model system, you can be sure to find corollary nodes and relationships in the system that it models. Therefore, a model says “this thing <em>is</em> this other thing,” or at least represents important parts of it.</p><p>As kids, my siblings and I built model cars with our father. These little, plastic simplifications held a symbolic relationship with their life-sized namesakes by sharing important features in common. In this case, it was the colors, shapes, and relative sizes of certain exterior components. By looking at the model, you are sure to be able to picture, understand, and predict the same features in the modeled system. If my model is red, the car it models should be red, too, and so on. This definition holds for every place we think of the word model. Even supermodels are offered up as symbolically representing what the typical or ideal human should look like. (Whether or not the modeled system in this case actually exists is another problem … but we’ll get to that in a moment).</p><p>Thankfully, you don’t have to get your fingers glued together or go on a weight-loss plan to model. We’re all modeling, every moment of every day. You might say that we can’t help it.</p><h3><strong><em>This is your brain, modeling…</em></strong></h3><p><a href="https://medium.com/@phlip007/intelligent-enough-10ac3622e865">The basis of intelligence is memory</a>, and memory is, itself, a model. When a certain combination of stimuli enters our bodies, they inspire the activation of networks of connected neurons. Those neurons can grow closer together or further apart, making the firing of that particular pathway more likely the next time those stimuli (or something sufficiently similar to them) cross our sensory radar. This growing closer or further occurs, for all intents and purposes, mechanistically, based on the presence of certain chemical hormones. These are, in turn, released in what should be a relatively deterministic pattern in response to other stimuli, usually the acquisition of some desirable internal state (food, reproductive processes, etc.). In his book <a href="https://agelab.mit.edu/books/grasp/"><strong><em>Grasp</em></strong></a>, Sanjay Sarma calls this the “foundation” of our cognitive hierarchy and is used as the basis for certain approaches to psychology, such as <a href="http://This%20is%20the%20basis%20of%20some%20psychological%20approaches,%20such%20as%20the%20operant%20conditioning%20championed%20by%20B.F.%20Skinner."><strong><em>operant conditioning</em></strong></a>.</p><p>Think about what this means for a moment, though. That means that there is an original system somewhere out there in the universe — a shiny car, a human being, a bear, it doesn’t matter — that is being modeled. There is, then, a model system — a collection of networked neurons inside of our skulls that connect to one another and fire in relationships that are directly correlated with the features and relationships in the original system. You don’t need to learn how to model, your brain just does it. It’s a modeling machine.</p><h3><strong><em>Where things get complicated</em></strong></h3><p>Somewhere along the line though, we started doing something … odd. We started making other models, <em>external</em> to our own minds, that represent the network of relationships that exist <em>within</em> our minds. We do this for a variety of reasons, though the simplest is probably communication. The rough picture of a bear or some other predator, by conjuring the notion of its presence in another mind, is useful for survival. It doesn’t take much to accomplish, either. The close visual likeness of something is usually enough to trigger the same internal model (in the form of a network of representative neurons) in someone else, provided they’ve had sufficient sensory experience to build a model corresponding to those stimuli. If we’ve both seen a bear, and I show you a photograph of one, it’s more than likely you’ll have a feeling of recognition, conjuring in your consciousness all of the same features of that bear you’ve collected in your experience — fluffy, brown, cuddly, you name it. Sometimes this little neuronal dance doesn’t work quite like we’d hoped. The study of this particular phenomenon is what we call <a href="https://plato.stanford.edu/entries/meaning/"><strong><em>semantics</em></strong></a>.</p><p>We have a <em>lot</em> of neurons in our brains, though. Their potential number of connections is used as a back-of-the-napkin estimate for the number of distinct networks, or models, they can create. It’s a lot. So, drawing a picture for every one of them became tedious after a while and we started writing things down. (That’s an oversimplification … another topic we’ll get to in a moment). Take, for example, the following paragraph:</p><figure><img alt="The quick brown fox jumps over the lazy dog." src="https://cdn-images-1.medium.com/max/1024/1*vHTC-wT4y-N35vWsdtciXA.png" /><figcaption>The text</figcaption></figure><p>In the course of reading this paragraph, each of the words you read, as well as their carefully crafted order, was enough to trigger a dizzying series of networking connections. Each of your models for the following creatures should have been activated:</p><figure><img alt="“The quick brown fox jumps over the lazy dog.” Annotated with photos of a fox and a dog." src="https://cdn-images-1.medium.com/max/1024/1*9PBavhoQKCEyX50ro5N8Pg.png" /><figcaption>The situation model</figcaption></figure><p>I’m not sure what comes to mind when you read this paragraph, but it should have also conjured a series of models — your corresponding neural pathways — relating to the properties of these creatures. Now we know that there is a corresponding system. It has two things in it: a dog and a fox. The fox possesses the properties of being both brown and quick, while the dog apparently possesses the qualities of laziness. While “laziness” might seem like it breaks our model thus far, it really doesn’t. “Laziness” is simply a symbol that we’ve assigned to experiences that are, originally, sensory stimuli. These include the sensation of annoyance when a teammate is less-than-proactive, the sights that are associated with individuals labeled as such — usually synonymous with “slovenly” or “unkempt” — and the feelings induced by social pressure and conditioning that “laziness is bad.”</p><p>What we just constructed is what Daniel Willingham refers to as the <strong><em>context model</em></strong>, and it’s a powerful skill that humanity has developed. We’ve taken a model (the combination of neurons representing the situation I observed), converted it into another model (the symbols in the text above), and used that to conjure yet another model (the combination of neurons representing the situation that exists in your mind). Phew!</p><h3><strong><em>Playing Telephone</em></strong></h3><p>It’s an understatement to say that this is an impossibly complex process. With all of that complexity, you can start to see where problems might arise. For instance, you might have a very different model representing your accrued experience with small, fuzzy brown things. When I show you that bear photo, it might not contain attributes of bears that exist in my model. If your models lack, for some reason, the attributes implying potential danger, we might be in trouble. This is relatively easy to get over when our models are <strong><em>grounded</em></strong>, that is, based on first-hand sensory stimuli. All I have to do is expose you (by way of a YouTube video, hopefully) to sights of what happens to people when they get too close to bears. Semantics ain’t so hard after all.</p><p>No, where we get into the most trouble is where our models aren’t grounded directly in sensory stimuli, but constructed as a house of cards — models built from the external representation of other peoples’ models. This is the premise of the classic game of telephone, where a line of people each whisper a word in each others’ ears, from one end of the line to the other. The words spoken by the first person in line might become horribly mangled by the end, all by the slight divergences in the relationships between auditory stimuli and the neuronal maps they invoke.</p><p>Quite insidiously, we just played this game. You read the sentence “the quick brown fox jumped over the lazy dog.” However, you have no sensory stimuli for this particular system. All you have is a collection of models associated with each element in the scene, such that the composition of each thing (the dog and the fox), their properties (laziness, quickness, browns), and their respective relationships (jumping, relative speed, one being “over” the other) are conjured. Each are immediately filled in such that the model of this system now exists in your mind.</p><p>This is important: There never was a dog and there never was a fox. There <em>is</em> no root sensory stimuli, some actual, out-there-in-the-universe system that corresponds to this particular model. Nevertheless, the model of that system exists in your mind and mine. It’s a very <em>long</em> game of telephone, so long and so convoluted that it has us building models of things that never existed.</p><h3><strong><em>Building Better Models</em></strong></h3><p>These sorts of models are built very easily in environments where a lot of words get thrown around. Unfortunately, this implies their prevalence in every social endeavor. Humans are, after all, unique in the amount of time they devote to building these kinds of external models. In contrast with a photo of the bear, the models we just described are <em>ungrounded</em> models, and they’re all over the place: in country clubs, religions, political rhetoric, and Medium articles (😉).</p><p>Because we often want our models to be “accurate” (that is, corresponding with the real world and the things we might experience in it), we’ve invented all sorts of ways to make better models — ones that are more resistant to the telephone game. For instance, we created a symbolic language that carries within it a means of verification — a way of taking that model, going out into the world, and making sure that it corresponds with the same system we encounter. That language is mathematics, and it’s been used effectively by humans to communicate verifiable models for a very long time, starting with arithmetic and the accounting of physical objects. We use it to communicate the relationships between items in our model, with the assertion that those relationships will hold true when you go to examine the pieces of the original system. However, there are certain aspects of the original system that aren’t easy to share in mathematics.</p><p>For instance, every system worth modeling exists in the stream of time. That is, we can say that as time progresses, some properties of the objects in our original system will be different if you tried to measure them at one point as opposed to another. The classic physics example is an object in free-fall, where the distance of the object from the ground changes over time. We experience these changes viscerally when we observe a system directly. However, when we describe the relationship between these two states using mathematics, the models we share (though rigorous) don’t intuitively convey these changes over time.</p><p>The solution to this problem is to build models whose properties, themselves, change over time. This helps us build models using sensory experiences that are much closer to the experiences we would have by observing the target system. We run scientific experiments using representative samples, test cars and wings inside of wind-tunnels, and build scale models of buildings, engines, and dams. We do this all under the assumption that the behavior of these smaller nodes, their properties, and relationships is representative of those in the target system. We often build these systems to communicate to others about the way the universe works. However, their presupposed similarities to their target systems allow us to go even further, making discoveries about the behavior of the target system, given discoveries about the model. We can readily identify these kinds of models in the computer world: simulations.</p><p>Computer software is a modeling medium <em>par excellence</em>. It provides (mostly) for the rigorous, verifiable relationships conferred upon it by the language of mathematics, while allowing its constituent pieces to evolve, change over time, and even influence their environment. Better yet, because of its deterministic underpinnings, every single one of these models is (hypothetically) reproducible, making our telephone game that much easier to avoid! What an accomplishment!</p><h3><strong><em>Models, Nails, and Hammers</em></strong></h3><p>Not every system is a nail. Therefore, not every system requires modeling using a hammer. Not every system needs to be modeled using the rigorous, imperative flavor of a programming language or the total descriptiveness of a simulation. Sometimes it’s fine (and more efficient) to use those fuzzy models we conjure when we use spoken or written language. If you’re not playing telephone, have easy access to verification, or are participating in a low-stakes activity, it usually isn’t worth it to be so explicit. Going further, emotional models are used to represent and conjure in others sensations that accompany certain internal states. These are exceedingly difficult to create using dry mediums such as coding languages. Therefore, we bring in completely different libraries of symbols — my personal favorite being musical ones. The types of models we create are as diverse as the systems we represent with them.</p><p>In his book <a href="https://sites.lsa.umich.edu/scottepage/home/the-model-thinker/"><strong><em>The Model Thinker</em></strong></a>, Scott E. Page lists a whole slew of reasons why we might create models. Generally, the models we create are determined by the outcomes we would like to achieve. Some models are only “good” within a narrow range of contexts. In <a href="https://www.amazon.com/Algorithm-Design-Manual-Computer-Science/dp/3030542580/?_encoding=UTF8&amp;pd_rd_w=MjqNq&amp;content-id=amzn1.sym.cf86ec3a-68a6-43e9-8115-04171136930a&amp;pf_rd_p=cf86ec3a-68a6-43e9-8115-04171136930a&amp;pf_rd_r=144-0742853-4106608&amp;pd_rd_wg=lnnHU&amp;pd_rd_r=e7df0e55-d5c4-4486-a1bf-2d41496d634c&amp;ref_=aufs_ap_sc_dsk"><strong><em>his book on algorithms</em></strong></a>, Steve Skienna refers to this range of contexts as a scale. He relates a humorous, though anecdotal example with flat-earthers. It’s not a great model if you’re looking to make predictions about the movement of stars, rockets, or satellites. However, if you’re looking to build a house and make predictions about its future stability … the flat earth model works just fine.</p><p>In fact, the domain of algorithms, found at the intersection of math and computer science is particularly interesting as a modeling discipline. It contains libraries of fantastic solutions to problems both common and uncommon. It does so by focusing its effort on a very small set of generalizable data structures. Don’t let terms like “generalizable data structures” scare you off, though. These are certain families of “nodes” with very specific relationships. They’re “generalizable” in the sense that the nodes themselves don’t have many properties. This makes it easy, when we go to build a model of something, to use them as symbols in our new system to represent the objects in our target system. (One might argue that this is the goal of all of mathematics, though the algorithmic sciences are unique in their fundamental focus on the steps required to move one of these generalized systems into a particular state).</p><p>It is the goal of a computer scientist to be well acquainted with these modeling techniques and select the right one to solve a particular problem. George Polya, in his book <a href="https://press.princeton.edu/books/paperback/9780691164076/how-to-solve-it"><strong><em>How to Solve It</em></strong></a>, refers to this process as “making an analogy.” It isn’t unique to computer scientists — Polya was a mathematician. On the subject of theoretical physics, <a href="https://youtu.be/6d9Uv9avT-g?si=_NR-K6RcqNFn64K8&amp;t=2680"><strong><em>Richard Feynman once said</em></strong></a> that “every theoretical physicist who’s any good knows six or seven different theoretical representations for exactly the same physics<em>.” </em>I don’t know about you, but when I hear Feynman talking about “theoretical representations,” all I hear is “models.”</p><p>This goes well beyond the technical disciplines, too. In this month’s and next month’s elections, candidates and their parties wage fierce competition using various models representing populations, their potential behaviors, and those of their adversaries. Their success or failure will be determined, in part, by their ability to judge the current state of the thing they are modeling and, once modeled, their means with which to manipulate the social system in question. Social systems, by the way, are excessively difficult to model — a point Dan Rockmore beats into the ground in <a href="http://www.apple.com"><strong><em>his essay for The New Yorker</em></strong></a>, emphasizing the supposed inadequacy of modeling in non-technical disciplines.</p><h3><strong><em>Stocking the Shelves</em></strong></h3><p>“Difficult” doesn’t mean impossible, and it certainly doesn’t mean it isn’t worthwhile. Modeling is possible and inevitable, in varying degrees of rigor, in every discipline. What makes the difference between the vaunted reliability of models in the hard sciences and the derided inadequacy of mathematical models in political science is, to borrow Skienna’s phrase, a matter of “scale.” I don’t mean the scale of the thing being modeled, I mean the scale of the model itself. The closer our model comes to resembling the thing it claims to represent, the more reliably we can communicate about, predict, or manipulate that original thing.</p><p>It is, therefore, the fundamental similarities between our model and its target system that determine how we use it. In the hard sciences, it might be a serendipitous coincidence that the thing we chose to model with — mathematics — <a href="https://writings.stephenwolfram.com/2023/09/expression-evaluation-and-fundamental-physics/"><strong><em>shares non-trivial similarities to the thing we’re modeling</em></strong></a> — physics — and therefore explains its concise, predictive power.</p><p>We just can’t do this in the social sciences. With an estimated <a href="https://hms.harvard.edu/news/new-field-neuroscience-aims-map-connections-brain"><strong><em>86 billion neurons</em></strong></a> in the human brain, our resources are expansive. However, they are inherently limited. In the social sciences, one studies systems whose every element has just as many neurons in their own brain. This doesn’t leave any room for you to begin modeling their behavior in any kind of comprehensive manner, let alone their complex relationships and interactions with each other and their environment. Any model of human behavior will have to leave out immeasurably many details. We just can’t create a model big enough to be totally accurate.</p><p>However, this doesn’t mean that all is lost for the social sciences. Communities of sociologists and virologists were able to work together over the course of the COVID-19 pandemic to save lives, even if their models were incomplete or wrong in some respects. Every day, we rely on oversimplifications — “flat earth” models — in disciplines including social sciences, workforce dynamics, and medicine. We keep using them, not necessarily because of their completeness, but because of their pragmatic outcomes. In both rigorous and non-rigorous modeling, Polya’s words on the analogy hold true: “it would be foolish to regard the plausibility of (a conclusion drawn from a good analogy) as certainty, but it would be just as foolish, or even more foolish, to disregard such plausible conjectures.” More accessibly, a quote <a href="https://fs.blog/all-models-are-wrong/"><strong><em>attributed to George Box</em></strong></a> says simply that “all models are wrong, but some are useful.”</p><p>This goes for <em>all</em> kinds of models, especially the ones that we make that live in our neurologically well-endowed skulls. <em>We are always wrong.</em> Sometimes, though, the models we make can help us do something useful. This is why, in his book on modeling, Scott E. Page advocates for a “many models” approach to life. Because every model is wrong, it means we can’t use one model for every situation, for every outcome we would like to affect. We need to build a library of tested, well-rounded, <em>grounded </em>models in each of our chosen problem domains that we can carefully choose and apply to new problems we encounter.</p><p>We need to put more tools in our metaphorical modeling toolbox. This is true regardless of the problem domain we work in. Musicians and composers hone their skills in manipulating specific combinations of symbols known to elicit shared emotional memories. Software Engineers build libraries of models using mathematical objects and the symbols found in computer code. Our ability to solve any problem depends on the depth and breadth of our models for that problem domain. It behooves us to stock the shelves of this library.</p><h3><strong>The Book Store</strong></h3><p>Expanding the depth and breadth of our modeling capabilities is the goal of education (though some use it for enforcing ungrounded models via ideological indoctrination). The problems that we encounter as individuals and as societies seem to grow ever more complex. This continues to increase the demand for educational systems that empower individuals to acquire deeper, richer models. Between the COVID-19 pandemic and the popular introduction of generative AI tools, last year, our traditional educational systems have demonstrated that they aren’t up to the challenge.</p><p>In my previous position at <a href="https://www.phenomena.app"><strong>Phenomena Learning</strong></a>, we tried to address this by building an educational modeling system. Phenomena’s orienting goal is to help students “stock their shelves” by building and interacting with an expansive array of models for STEM topics using a variety of modeling modalities. Importantly, these modalities <em>include</em> code. We built a block-based coding language to allow for the development of rigorous, imperative models with as few frictions as possible.</p><figure><img alt="A screenshot of the Phenomena block-based code editor" src="https://cdn-images-1.medium.com/max/1024/1*j7BrgBN8fIpvirF6ccNHSg.png" /><figcaption>The Phenomena block-based code editor</figcaption></figure><p>However, code wasn’t the only modality we included. We also created a mathematical modeling language, allowing prospective modelers to describe the relationships between objects and their properties using a visual, web-like editor.</p><figure><img alt="A screenshot of the Phenomena mathematical modeling editor" src="https://cdn-images-1.medium.com/max/1024/1*hZont5eEkuOTGl-0qdpgWg.png" /><figcaption>The Phenomena mathematical modeling editor</figcaption></figure><p>We also put a premium on the existing experience, or the lack thereof, that modelers might bring to the table. We opted for a block-based programming language instead of JavaScript to help inexperienced coders avoid syntax errors. We created a mathematical modeling language to allow students to take formulas directly from the page and put them in motion. The goal is to focus on the model, not the thing you’re using to create it. However, you don’t have to use either of these. As a drawing tool, it supports simpler, static model building as well. Much of our work involved an expanding library of “components” that can be combined as fundamental building blocks — no code required.</p><p>Though I’ve since moved on from Phenomena, the project of modeling never stops. In my current position, at <a href="https://www.coursemojo.com/technology"><strong>Coursemojo</strong></a>, we’re building a system that improves on the context model, described above. It takes textual models, such as those found in English or Language Arts classrooms, and helps students to build more expansive, deeper mental models, using a conversational interface. Outside of educational contexts, I’ve written extensively about <a href="https://medium.com/@phlip007/the-brain-attic-d338162f4c90"><strong>The Brain Attic</strong></a>, a hypothetical system that helps us connect the external models we collect in our problem domain and helps us ground (or demonstrate the ungrounded-ness) of those models. The opportunities to become better modelers are many.</p><h3><strong>The People That Make Meaning</strong></h3><p>Building this library, exercising the representations we stock it with, and refining them through iterative use is an enjoyable and rewarding activity. I can’t think of a better framing for this conclusion than Margot Robbie’s (wrongfully-Oscar-snubbed) line in the film <em>Barbie, </em>quoted at the top of this article. Therein, the model — the external representation of a human idea — spoke back to us, expressing a desire to be among the people who make these things instead of the model itself.</p><p>What does all of this mean for those prospective coders out there? Frankly, it’s never been enough to know how to code. Computer code, in its various iterations over the last several decades, has only ever been a medium through which we can create powerful, dynamic models. It doesn’t matter if the methods we use now are automated away in 3–5 years. There will always be a need for human beings to create well-grounded, rigorous models, regardless of the medium that is being used. It’s with this context that I often recommend to students, recent graduates, or people seeking to enter the field that it is vitally important to establish experience in a problem domain that <strong><em>isn’t</em> </strong>coding. That is, instead of being “a coder,” one should seek to be “an educator that knows how to code,” or a “supply line expert that knows how to code” or something along those lines. Being “one of the people who make meaning” means that we have to have intent — understanding the system we are trying to model and what we want to do with it, as opposed to simply being familiar with the medium through which we do so.</p><p>That’s what I mean when I say “don’t learn to code” — don’t learn to code as an end unto itself. Learn to model. Collect experience in building clear distinctions between the system you are trying to represent and its environment. Recognize what type of medium you are using to make this representation. Try to be as clear as possible about your goals in the modeling process, whether that is to predict, to communicate, or to experiment. Most importantly, learn to build verifiable models, distinguishing between ones that are grounded and measurable and those that aren’t. You might do all of this using the dynamic, rigorous power of imperative programming languages. You might not. Whatever modeling medium you choose, though, build lots of them. As each of us fills the shelves of our model libraries, we’ll become that much better problem solvers. Perhaps more importantly, we’ll “be part of the people that make meaning, not the thing that’s made.”</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7e19e6b91d5f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI interactions: not a copilot, but a coach]]></title>
            <link>https://uxdesign.cc/not-a-copilot-but-a-coach-4f4be8e1104a?source=rss-6f0a447e2764------2</link>
            <guid isPermaLink="false">https://medium.com/p/4f4be8e1104a</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ux]]></category>
            <category><![CDATA[microsoft]]></category>
            <category><![CDATA[problem-solving]]></category>
            <dc:creator><![CDATA[Philip Grabenhorst]]></dc:creator>
            <pubDate>Tue, 09 Jan 2024 19:27:01 GMT</pubDate>
            <atom:updated>2024-01-13T08:41:07.368Z</atom:updated>
            <content:encoded><![CDATA[<h4>How we can build more useful AI interactions into our keyboards.</h4><figure><img alt="A fish, attempting to consume the Microsoft Copilot Logo, represented as a lure." src="https://cdn-images-1.medium.com/max/1024/1*yO9EQdjk2RWYz6jkPQBk5Q.png" /><figcaption>Copilot Logo Credit: Microsoft</figcaption></figure><p>Last week, Microsoft announced a change so large that they could justify calling it the <a href="https://www.youtube.com/watch?v=S1R08Qx6Fvs">first significant change to the Windows keyboard in nearly three decades</a>. Coming to hardware vendors in 2024 and (presumably) the next iteration of Surface devices, we’ll start to see a key dedicated to Microsoft’s Copilot, snuggled in next to our arrow keys. It’s prime real estate. It’s a shame we couldn’t find a better tenant.</p><p>But we’ll get to that in a moment. The move signals Microsoft’s eagerness to bring Copilot front-and-center, capable of bearing down on any workload we happen to have present on our screens. This follows a flurry of development. After its initial integration into Bing search and the Edge browser, at the start of last year, Microsoft moved quickly to integrate OpenAI’s generative language models into every service under its umbrella, including Microsoft Office and its suite of tools. In September, the service came to the Windows Desktop. In November, the company <a href="https://www.theverge.com/2023/11/15/23960517/microsoft-copilot-bing-chat-rebranding-chatgpt-ai">rebranded</a> each of these integrations under the more inclusive Copilot name.</p><p>If this isn’t all-in, I don’t know what is. I’ve seen hundreds of job listings on Microsoft’s career page, advertising some kind of involvement in the Copilot project. After <a href="https://searchengineland.com/new-bing-google-market-share-six-months-430840">lackluster adoption</a> of its AI-powered Bing search, Microsoft seems eager to find <strong><em>something</em></strong> people want to use that will justify its exorbitant spending on OpenAI and its services. In its <a href="https://www.youtube.com/watch?v=5rEZGSFgZVY">original advertisements for the Copilot rebrand</a>, it seemed to conjure every potentially useful generative trick up its sleeve: summarizing web pages, writing messages, filling images, adjusting system settings, and so on. In <a href="https://www.npr.org/transcripts/1221128925">a recent interview</a>, Satya Nadella said that he uses it to help him draft memos.</p><p>While many of these features might be genuinely useful for certain workflows, they fundamentally miss the mark which is the original purpose of the personal computer: helping us to be better problem solvers. Back when GPT-3 was first released, Katelyn Donnelly employed the copilot analogy as well — but as a warning, not a branding guideline. Asking “<a href="http://www.apple.com">should planes fly themselves</a>”, the gist of the essay is that we run the risk of not developing certain important skills or, once attained, allowing them to atrophy. I won’t focus on deskilling, here, but it deserves emphasizing that the alleviation of drudgery should not be the final goal of our creativity in the LLM space and it isn’t always a desirable one.</p><p>In the parlance of an ancient Chinese proverb, the kind of tools currently offered in Microsoft’s copilot strategy are more akin to “giving the people fish.” Instead of encouraging workflows that encourage the development of skills to solve big problems, they instead release us from the drudgery of solving many small problems. Instead of just giving us fish, Microsoft should employ its (1) special relationship with OpenAI, (2) omnipresence in the desktop computing space, and (3) special relationship with our personal information, via Outlook, Onedrive, and Office, to use that shiny new Copilot button for a better purpose: teaching us to fish. But what does this look like, if not the current model? To answer that question, I’ll again turn to the analogy of the coach. For the sake of argument, let’s call it <strong><em>Coach 365</em></strong>.</p><h3><strong>Better Keyboards</strong></h3><p>The most important function of the personal computer is one that gets lost in a sea of content consumption. Our hypothetical <strong><em>Coach 365</em></strong> shouldn’t be focused on creating content for us or retrieving someone else’s content for us to consume. Rather, it should be designed to prompt us towards connecting the content we have already created or curated. Instead of using my keyboard to prompt an AI assistant to write an email for me, I want it to give me feedback on the content <strong><em>I’m</em></strong> writing. Does the intent of the paragraph I just wrote happen to relate, strongly, to that piece by Katelyn Donnelly that I read a while back? That would make a great reference!</p><p>What if we explicitly mention a person or place in our research or notes? What if our handy-dandy keyboard prompted us to replace that static reference with a link? Here, I imagine a system similar to the now ubiquitous <em>linked mention,</em> used in Slack, Apple Messages, and other messaging systems. Therein, a name is recommended to be preceded with an <strong>@</strong> sign, linking the message to the recipient. Instead of remaining in the context of the conversation, they would extend across the system. Notes that I make about friends or colleagues would organically grow as references to their contact cards. This would provide the backbone for contextualizing and solving problems related to the people who are in our lives.</p><figure><img alt="Windows OneNote, where a proposed AI system is recommending similar content for the note-writer to connect it to, including contact names and other notes." src="https://cdn-images-1.medium.com/max/1024/1*ree5fwzxLYnFq2KwY6oHUQ.png" /><figcaption>Copilot, recommending links to related content</figcaption></figure><figure><img alt="The Microsoft Outlook application, showing a contact, as well as all of the locations where that contact is mentioned in one’s personal web of information." src="https://cdn-images-1.medium.com/max/1024/1*JbKzQdYJGf5u41QXYP-4Gw.png" /><figcaption>Copilot, showing you where a contact is mentioned</figcaption></figure><p>As an example, we recently served chocolate-peanut-butter pie at a party, completely forgetting that my brother-in-law has a peanut allergy. What if, in the process of writing down a grocery list, linked to an event, which is in turn linked to the contacts attending, <strong><em>Coach 365</em></strong> had surfaced an old note, reminding me that peanuts and brothers-in-law don’t go together?</p><p>This functionality goes well beyond contacts, potentially including any object in our technological life: photographs, geotags, notes, music recordings, and so forth. At this scale and capacity, our virtual assistants would become truly helpful. I like to think of it as a regression test, but for life. If we have already found a solution to a particular problem or recorded information vital to a solution, <strong><em>Coach 365</em></strong> should help us to remember it.</p><h3><strong>Deleting Things, Not Creating Them</strong></h3><p>In an older essay, I wrote about the importance of <a href="https://medium.com/@phlip007/the-unfinished-revolution-f0df6b1349ed">keeping our personal web of content, as one does a garden</a>. That includes pruning pieces of information that are no longer pertinent to the problems we are seeking to solve. While subjective, we should remember that <a href="https://hiddenbrain.org/podcast/less-is-more/">less is still more</a>.</p><p>In addition to recommending connections amongst old information, our <strong><em>Coach 365</em></strong> should also help us see when a new piece of content (notes, documents, what have you) is similar to older pieces of content. As we are in the act of writing, or soon thereafter, it could compare what we have just created to what we have created before. In doing so, we can avoid duplication. Beyond duplication, it provides an opportunity to assess whether the problem, concept, or piece we are creating is truly <em>distinct</em> or if it deserves incorporation into an existing structure. I’m imagining this as being especially useful in educational or research settings.</p><figure><img alt="Copilot, proactively recommending that the user merge one note with another based on their similarities." src="https://cdn-images-1.medium.com/max/1024/1*myr3RLRWFtAcJhyiITfE0A.png" /><figcaption>Copilot, recommending you to merge similar content</figcaption></figure><h3><strong>New Solutions to Old Problems</strong></h3><p>Now for a use-case that Microsoft is especially well-primed to take advantage of. If an artificially intelligent assistant knew (1) the context of our problem and (2) our goal, it would be poised to inform us about features or workflows that we were unaware of. What do we do when don’t know how to do something? We Google it. What about the problems we don’t Google? How often do we develop inefficiencies in our workflows, because we developed a methodology that keeps “kinda” working? So many times, I have witnessed computer users clicking through reams of emails with the intent to delete them, unaware of the “Cmd+A” shortcut, or searching for an undo button, completely oblivious of the “Cmd+Z” shortcut. Basically, it’s the kind of thing Clippy used to do (albeit poorly).</p><figure><img alt="The original Microsoft Office Assistant, Clippy" src="https://cdn-images-1.medium.com/max/800/1*XTD0d96jsC1qqUhFrTcrsA.jpeg" /><figcaption>Image Credit: Microsoft</figcaption></figure><p>But we could do more. Here, I also have in mind problems of computer engineering. How many times do we, as software engineers, implement an algorithm or method inefficiently, simply because we’re unaware of a simple, more efficient solution? In the contexts of pair coding and technical interviews, we get a chance to unearth these holes in our understanding and have our intuitions challenged. But what if our <strong><em>Coach 365</em></strong> proactively informed us of the more efficient solution? In this way, we would still benefit from the learning experiences of performing our own generative tasks. However, we would still be “alleviating drudgery” by becoming better, more efficient problem solvers ourselves. How’s that for a win/win?</p><figure><img alt="Copilot, proactive recommending a more concise approach to the coding problem of creating an array of integers filling a given range." src="https://cdn-images-1.medium.com/max/1024/1*f1JkzVCNDnqY87a4NZ-igg.png" /><figcaption>Copilot, recommending a more concise approach to a coding problem</figcaption></figure><p>This doesn’t have to stop at the edge of the technical practices. Instead of asking <strong><em>Coach 365</em></strong> to draft a memo or a piece of marketing material, what if it took a competitive approach? It could follow our lines of argument, pointing out common logical pitfalls, places where our reasoning doesn’t hold up to superficial scrutiny, or we otherwise exhibit some clear bias. In doing so, it could even start to coach us out of insulation, ignorance, and complacency.</p><h3><strong>Copilot, Codependent</strong></h3><p>Meanwhile, back at the farm, Copilot seems to be offering us a hodgepodge of disconnected, supposedly time-saving conveniences. Apart from narrow use cases, such as the Edge browser or the current page of a Word Document, there is a serious disconnect between the content we are currently working with and the feedback Copilot provides. Even Microsoft’s front-page piece, the Copilot system prompt on Windows 11, seems devoid of context.</p><p>Simple contextual queries are met with confusion. “What’s on my screen right now?” Let me search the web for that, it says. “Who am I?” An interesting philosophical question, Copilot quips. “What’s on my calendar?” Let me search the web for that, too. Though placed in a window of prestige, Copilot seems to still believe that it is housed firmly in the Bing search browser, ready to help me search for and purchase self-help books in the pursuit of answering that burning “Who am I?” question. These are real interactions, by the way…</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Ffd.xuwubk.eu.org%3A443%2Fhttps%2Fwww.youtube.com%2Fembed%2FVQZPNpgsWlc%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Ffd.xuwubk.eu.org%3A443%2Fhttps%2Fwww.youtube.com%2Fwatch%3Fv%3DVQZPNpgsWlc&amp;image=https%3A%2F%2Ffd.xuwubk.eu.org%3A443%2Fhttps%2Fi.ytimg.com%2Fvi%2FVQZPNpgsWlc%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/99deba95e5bd128216139417e6f8d942/href">https://medium.com/media/99deba95e5bd128216139417e6f8d942/href</a></iframe><p>To the Copilot team’s credit, though, these problems are presently expensive to solve. It is by no means a small undertaking to index, compare, and generate text for so many, potentially unrelated interactions. Local device limitations, the speed of existing models, as well as other frictions, play a part. Nevertheless, Microsoft offers some services that begin to approach the level of self-referential connectivity I spoke of above — they’re just hidden behind the <a href="https://www.microsoft.com/en-us/microsoft-365/enterprise/copilot-for-microsoft-365#tabs-oc2a1c_tab2">Microsoft Copilot 365</a> “For Enterprise” banner. Even then, the number of messages a user can make are limited. Worse though, it’s tailored in such a way that it just gives us more fish than the unpaid tier.</p><h3><strong>Copilot or Coach?</strong></h3><p>These limitations are quickly receding, though. In <a href="https://arxiv.org/abs/2312.11514">a paper, last month</a>, researchers at Apple outlined a method by which we might be able to sidestep the memory limitations of personal devices, to locally run ever larger, more precise models. Even as the implementations of our machine learning models become more clever, our personal technologies are still getting faster and faster, hurtling toward the <a href="http://www.apple.com">physical limits of computation</a>. We will be able to equip language-generating models with more context, very soon.</p><p>So what will we do with them? Will we continue to ask our systems to give us fish, or will we learn how to fish? We need an artificially intelligent assistant that is aware of our personal data and our relationships with others. It should be equipped with mechanisms that allow it to retain our interactions with it, as well as use sentiment analysis to perform fine-tuning in response to the feedback we give it. It should help us to create new connections with and recommend new applications for the information we curate. It should be competitive, and not just passive, refining our existing workflows and helping us make the most of our computational systems. It should coach us out of complacency, too, helping us recognize where the solutions we generate are inefficient or insufficient.</p><p>If we get this right, imagine the problems we could solve. I believe that building <strong><em>Coach 365</em></strong> is imperative — something that will equip each of us to build, create, and solve problems at levels of performance previously only achieved by a choice few. I believe that creating these systems will be a bootstrap of sorts, raising our collective problem-solving capacity such that it allows us to solve the greatest challenges facing humanity today and whatever the future has to throw at us. Let’s do it. Let’s build it. Let’s go fishing.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4f4be8e1104a" width="1" height="1" alt=""><hr><p><a href="https://uxdesign.cc/not-a-copilot-but-a-coach-4f4be8e1104a">AI interactions: not a copilot, but a coach</a> was originally published in <a href="https://uxdesign.cc">UX Collective</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Here they come: an AI on every phone]]></title>
            <link>https://uxdesign.cc/here-they-come-03094f87f0b7?source=rss-6f0a447e2764------2</link>
            <guid isPermaLink="false">https://medium.com/p/03094f87f0b7</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[user-experience]]></category>
            <category><![CDATA[ux]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Philip Grabenhorst]]></dc:creator>
            <pubDate>Mon, 11 Dec 2023 12:22:38 GMT</pubDate>
            <atom:updated>2023-12-11T18:43:01.140Z</atom:updated>
            <content:encoded><![CDATA[<h4><strong>Application development using device-native LLMs.</strong></h4><figure><img alt="A pen-drawn rendition of the Oracle, taken from John Collier’s Priestess of Delphi, on a smartphone screen." src="https://cdn-images-1.medium.com/max/1024/1*-kW52nOo70b0TisguQOL5A.png" /><figcaption>John Collier’s <em>Priestess of Delphi</em> on a smartphone screen</figcaption></figure><blockquote>“The best way to predict your future is to create it.” — Abraham Lincoln</blockquote><p>To say that the year since ChatGPT’s release has been eventful would be an understatement. In addition to the tectonic shifts taking place in every sector of the IT industry, the pace of development — building bigger, faster, and stronger models — seems to grow ever faster. Last week, the Google DeepMind team announced a leap-frog over their OpenAI counterparts, who were, themselves, still reeling from the drama of <a href="https://www.nytimes.com/2023/12/09/technology/openai-altman-inside-crisis.html">Sam Altman’s almost-ouster</a>. In between all of this, Anthropic blew away industry standards with its release of <a href="http://www.apple.com">Claude 2.1</a> and its 200k token prompt sizes. Our digital oracles are becoming faster and cheaper.</p><p>Google’s release was a project called <a href="http://www.apple.com">Gemini</a>. Like a McDonald’s meal, it comes in three sizes: Pro, Ultra, and Nano. Someone in their PR department had very high hopes, dubbing it the start of the “Gemini Era.” The demos featuring its most powerful version gave us an idea of what a fully multimodal model could do (although these appear to have been <a href="https://techcrunch.com/2023/12/07/googles-best-gemini-demo-was-faked/">embellished or outright fabricated</a>).</p><p>Regardless, I’m much more interested in the smallest version of project Gemini — Gemini Nano. It’s a text-only model intended to run on mobile devices. According to the Android Developer website, <a href="https://android-developers.googleblog.com/2023/12/a-new-foundation-for-ai-on-android.html">it’s already running on Pixel 8 Pro devices</a>, powering offline note summaries and smart keyboard interactions.</p><p>A few months back, I wrote <a href="https://betterprogramming.pub/were-ready-for-our-platform-native-llms-now-please-502ce598449c">an essay</a> about the need for Large Language Models as an operating-system level application service. They’re finally coming. Gemini Nano and Android’s new AICore service are the first serious, official offerings. However, a head of steam is building on other platforms, as well. Over the summer, the Hugging Face community has made progress in making openly available models <a href="https://huggingface.co/blog/swift-coreml-llm">easy to run on Apple’s Core ML</a>. Furthermore, <a href="https://techcrunch.com/2023/10/23/apples-job-listings-suggest-it-plans-to-infuse-ai-in-multiple-products/">if the hiring data is anything to go by</a>, Apple might not be far behind with an official set of LLM-powered services.</p><p>Generalized Artificial Intelligence is moving closer and closer to each of us. While the performance and viability of these early systems have yet to be verified, we can be sure they’ll iterate and improve, quickly. AI now lives in our pockets. The question for us, as application developers and software engineers, is this: now that device-native LLMs are here, how do we build the most empowering software we can? What problems do we still have to overcome?</p><h3><strong><em>Performance</em></strong></h3><p>The first question is whether or not these models will even be useful. Will Gemini Nano be fast enough to be conversational? Will its prompt-size be on par with current server-side offerings, such as GPT 4 and Claude? The answer to the question “do I ditch my API for a local service?” is something that only individual teams, with a better sense of their projects’ context, can answer. However, in my last essay, I demoed the comparative performance of some local LLMs. They weren’t half bad. I can’t imagine Gemini Nano will be any worse.</p><h3><strong><em>Information and Locality</em></strong></h3><p>As these models become more powerful, we naturally want to delegate more work to them. As they grow closer to us, we expect them to be more understanding of the unspoken context surrounding our conversations. OpenAI’s answer to this problem was simple: “store more of your life on our servers.” The Assistant features they released in November rely heavily on the concept of “threads” and uploaded files. The feeling of familiarity and contextual understanding is provided by keeping a conversation going, regardless of the device you access it on, potentially referencing files you have explicitly provided.</p><p>Google’s answer to this problem was … surprisingly … different. In their documentation for Gemini Nano and AICore, they repeatedly emphasize the local, encrypted nature of inputs and outputs from the model. Instead of storing context in the cloud, individual applications can keep track of it, locally. Additional training is possible, using a per-application adapter that loads on top of the base model. However, this raises some questions. If an application is intended to support a uniform experience across several devices, how will this additional training be synchronized? Instead of duplicating information on the client and the server side (OpenAI’s approach), will we be duplicating effort and additional training across devices?</p><h3><strong><em>The Experience</em></strong></h3><p>Regardless of how this additional training is applied, there’s another, more important, problem that remains unresolved. AICore’s additional training is application-specific. However, our problems are not. If we want our devices to help us to be better problem solvers, they should demonstrate a cross-sectional understanding of our lives. OpenAI’s assistants miss out on this. However, Google’s services contain a plethora of personal material (contact information, calendar events, and so on) that could be automatically applied as training material for Gemini Nano, making its responses more personal and applicable. I believe this should exist at the operating system level, instead of forcing developers to duplicate this effort across each of their applications.</p><p>Of course, this might pose a problem from a privacy perspective. Should each application with access to AICore be able to craft prompts that might regurgitate personal information about your friends, relatives, and near-term schedule? Of course not. However, a messaging app cannot genuinely help us craft useful responses without knowledge of the individual to whom we are speaking. (I find it questionably useful to have an AI assistant write my texts for me anyway). A project management application will be more useful if it is informed with the emails we have received throughout our work day. In the same vein, a journaling application will be more useful if it is informed by the photos we took throughout the day. This is the original promise of projects like Siri and Google Assistant, and it is now within our grasp. If our goal is to provide assistance in traversing the web of information that we encounter every day, making sense of it, and applying it to solve real problems, this personalized approach will be vital.</p><p>Platform-native LLMs are here to stay. Soon, we will all have an oracle in our pocket. What we choose to do with them, though, is an eventuality that has yet to be divined. It’s my hope that we will build systems that aid us in assimilating, contextualizing, and applying the information we absorb. By doing so, we can be better problem solvers and more understanding human beings. Project Gemini and Android’s AICore mark a step in the right direction, but it still leaves a lot on the table. So let’s keep at it. Let’s keep building.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=03094f87f0b7" width="1" height="1" alt=""><hr><p><a href="https://uxdesign.cc/here-they-come-03094f87f0b7">Here they come: an AI on every phone</a> was originally published in <a href="https://uxdesign.cc">UX Collective</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[There’s More to Life Than Deliverables]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@phlip007/theres-more-to-life-than-deliverables-0e63f75cc50d?source=rss-6f0a447e2764------2"><img src="https://cdn-images-1.medium.com/max/2600/1*QErD5EDpkdd1m-3YoUP_Aw.png" width="4272"></a></p><p class="medium-feed-snippet">Going Back to School as a Full-Time Developer</p><p class="medium-feed-link"><a href="https://medium.com/@phlip007/theres-more-to-life-than-deliverables-0e63f75cc50d?source=rss-6f0a447e2764------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@phlip007/theres-more-to-life-than-deliverables-0e63f75cc50d?source=rss-6f0a447e2764------2</link>
            <guid isPermaLink="false">https://medium.com/p/0e63f75cc50d</guid>
            <category><![CDATA[liberal-arts-education]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[growth-mindset]]></category>
            <category><![CDATA[self-awareness]]></category>
            <category><![CDATA[college]]></category>
            <dc:creator><![CDATA[Philip Grabenhorst]]></dc:creator>
            <pubDate>Tue, 14 Nov 2023 16:05:49 GMT</pubDate>
            <atom:updated>2023-11-15T15:58:45.284Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[The missing link in Apple’s event]]></title>
            <link>https://uxdesign.cc/the-missing-link-10793359ef93?source=rss-6f0a447e2764------2</link>
            <guid isPermaLink="false">https://medium.com/p/10793359ef93</guid>
            <category><![CDATA[editor-picks]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[ux]]></category>
            <category><![CDATA[education]]></category>
            <category><![CDATA[apple]]></category>
            <dc:creator><![CDATA[Philip Grabenhorst]]></dc:creator>
            <pubDate>Tue, 19 Sep 2023 15:33:12 GMT</pubDate>
            <atom:updated>2023-09-19T23:27:20.190Z</atom:updated>
            <content:encoded><![CDATA[<h4>Why “it just works” doesn’t work.</h4><figure><img alt="The development of apple devices over time, superimposed on a stereotypical depiction of human evolution." src="https://cdn-images-1.medium.com/max/1024/1*r_cO53bLAu43GNm0Pvmzqg.png" /><figcaption>Credit: macrovector / Freepik</figcaption></figure><blockquote>“Just Do It” — Nike</blockquote><p>Last week, at Apple’s annual launch event, we were treated to what is now a common sight: an array of dazzling, shiny new pieces of hardware. The new iPhone Pro treats us to the slickest-looking shell to ever wrap a smartphone. The Apple Watch looks better than ever with new (definitely-not-leather™️) bands. More exciting for us software developers were the interfaces we saw introduced, ones not even mentioned at WWDC. For the iPhone, it was the new action button. For the Apple Watch, it was the new <em>Double Tap</em>, something that allows one to trigger any arbitrary action with a flick of the wrist.</p><p>It looks shiny.</p><p>It looks fun.</p><p>It all looks so nice…</p><p>…but it’s also useless.</p><p>Features without context, without direct applications to the problems they solve, are useless. Though they were accompanied by some demonstrations, such as through promotional videos on Apple’s website and snazzy snippets during the release conference, their direct application to the problems we face in our daily lives is missing. This means that the goal of personal technology — to make us better problem solvers — remains unfulfilled. That’s because there’s a piece missing. If we can find that piece — our missing link — and put it back where it belongs, we’ll all be better problem solvers for it.</p><h3>The Full Chain</h3><p>Though the mythology of personal computers is a long and misty one, the work of Douglas Englebart and his colleagues stands out as defining. In their 1972 piece, <em>Augmenting Human Intellect: A Conceptual Framework</em>, the imaginative foundation was laid for what eventually became the personal computer. Its purpose? To aid individuals and groups in “comprehending complex situations, isolating the significant factors, and solving problems.”</p><p>The framework Englebart and his colleagues proposed was made of four pieces: “a <strong><em>trained human</em></strong> being together with [their] <strong><em>artifacts</em></strong>, <strong><em>language</em></strong>, and <strong><em>methodology</em></strong>.” That is, the problem solver (the human) would describe and conceive of a problem using a language. They would then proceed to solve that problem by wielding artifacts according to some methodology. It’s a beautifully concise and expressive framework. For more on that, see my article <a href="https://medium.com/@phlip007/the-unfinished-revolution-f0df6b1349ed"><em>The Unfinished Revolution</em></a>, or, better yet, check out <a href="https://www.dougengelbart.org/pubs/papers/scanned/Doug_Engelbart-AugmentingHumanIntellect.pdf">the original work</a>.</p><p>Englebart places particular emphasis on the <strong><em>training</em></strong> portion, though. He uses the example of an ancient human, flung into the present, being asked to drive a car. “While [he] cannot drive a car through traffic, because he cannot leap the gap between his cultural background and the kind of world that contains cars and traffic, it is possible to move step by step through an organized training program that will enable him to drive effectively and safely.”</p><p>Let’s make this concrete. The iPhone is an artifact — nothing more. This year, we were provided with artifacts that have vaguely different properties compared to our old ones. That these changes should help us be better problem solvers is only the tail end of a vague implication. The reasons why are swept under a rug labeled “it just works.” Their guiding assumption is this: if our artifacts can be made simple enough and intuitive enough, then we <strong><em>humans</em></strong> can apply all of our existing <strong><em>methodologies</em></strong>, <strong><em>language</em></strong>, and <strong><em>training</em></strong> to our new artifacts to solve our problems. This is exactly the opposite conclusion of Englebart and his colleagues. There’s no “moving step by step through an organized training program” and therefore no expansion of our libraries of problem/solution pairs. Frankly, it’s lazy.</p><p>It’s not lazy from a product development perspective. I have nothing but the deepest respect for the engineers at Apple. Furthermore, this optimization is hard work. In his book <em>The Design of Everyday Things</em>, Donald Norman argues for exactly this kind of “training-less” optimization in human-computer-interfaces, citing benign examples of door handles and terrifying stories of failures at nuclear power plants. If not done right, systems that attempt to conform to the constraints of existing expectations can be difficult to use, frustrating and, in some cases, dangerous.</p><p>But where is the optimization in a double tap? Where is the efficiency of an action button? They are produced when we use them to streamline existing workflows or produce entirely new ones. How do we do that? Which workflows <em>should</em> we optimize? Beats me. And watching a video of somebody buying apples at a Santa Barbara farmers market won’t help answer these questions.</p><h3>The Misalignment</h3><p>Why is this the case? Why do we have systems that are so thoroughly researched and designed for their capabilities to optimize workflows and make us better problem solvers … when the training to use them doesn’t exist? It’s because the incentives are misaligned.</p><p>Apple Computers, and just about every other technology vendor on the planet, is incentivized to produce their products according to market considerations. Some are simple some are complex. Though ostensibly (and I think oftentimes genuinely) motivated by a desire to make our lives better, they can only do that if they keep selling new iPhones every year. This requires some kind of “reason why” advertising — some improvement or killer feature that motivates us to purchase a new product. That means that the <strong><em>event</em></strong> technology vendors optimize for is the purchase. This can take different forms and becomes more complex when you consider retention cycles and so on. However, the optimization is the same.</p><p>Conversely, the <strong><em>events</em></strong> we want to be optimizing for are problem instances — individual instantiations of the classes of problems we encounter in our daily lives. Now, technology vendors claim that they are optimizing for this event, as well. After all, won’t we keep buying something if it keeps solving more of our problem instances? Well, maybe. But if this were so — if the two incentives were really aligned — why is there so little training? The fact that such a vital piece of our problem-solving architecture is missing from our relationships with technology vendors is evidence of a misaligned set of incentives. Building systems that “just work” is good for sales and for a narrow set of universally shared, preexisting, shallow problem domains. It’s not a good policy for becoming better problem solvers.</p><h3>Below the Curve</h3><p>Some will say that this is simply a problem of learning curves. A large larding curve is imposing. A smaller learning curve is easier to overcome, and thus easier to sell. However, learning curves aren’t the problem.</p><p>Most tool makers are actually okay with large learning curves. They just rely on psychological factors to overcome them. Many vendors rely on a principle known as the “adoption curve” to anticipate, explain, and encourage the adoption of truly provocative technologies — including ones that don’t simply cater to existing human habits. The theory goes this way: if you introduce a tool that requires a fundamental reevaluation of the way someone interacts with their personal technology, most people aren’t going to get it. Most people won’t want to overcome their existing habits of interaction to engage in such in-depth training. However, some people will. According to this graph, it’s only about 2.5%. However, the result is (again, in theory) a cascading reaction, where each willing adopter influences and convinces an adopter just slightly behind them. The adoption curve has been used to explain the introduction of a variety of technologies, with learning curves big and small.</p><figure><img alt="The adoption curve, rendered as a standard distribution on a graph." src="https://cdn-images-1.medium.com/max/640/1*LWP1N1vHX0JIvmx9v3awag.png" /><figcaption>Credit: <a href="https://whatfix.com/blog/technology-adoption-curve/">whatfix</a></figcaption></figure><p>I like this idea. It’s something along the lines of “the tail wags the dog” and it’s superficially similar to experimental results in Sociology. Erica Chenoweth, studying social movements, drew similar conclusions after reviewing <a href="https://www.bbc.com/future/article/20190513-it-only-takes-35-of-people-to-change-the-world">data on over one hundred years of societal changes</a>. Her number was 3.5%.</p><p>My favorite, recent example of the adoption curve in action is the Arc Browser. When they launched, it was with a fundamentally different etymology and ontology for browsing the web. The most intrusive change was this: tabs don’t belong to windows, they belong to “spaces” which are independent of windows and persist even after you close a window. This was weird. Initially, even I was put off by it. Opening the same tab in two windows could result in some unintuitive behaviors, such as accidentally closing it in both locations.</p><p>When you look at their rationale, though, it’s brilliant. Arc is iteratively creating an “internet computer”, where the tools we need in our daily lives are organized according to their context. Your work tools live in your workspace, your personal bookmarks live in your personal space, and so on. What’s more, these contexts are available independent of the device you access them on. The goal is to abstract the role of individual devices and bring us that much closer to what I like to call <a href="https://medium.com/better-programming/intentful-computation-e502db996ce1"><em>Intentful Computation</em></a>. It’s a beautiful model of computing.</p><p>Anyway, Arc ditched this feature back in June. Their rationale at the time was that it was causing too much “confusion” for users. One could read here: “it had too high a learning curve.” However, the 2.5% spoke up in its defense. <a href="https://browsercompany.substack.com/p/there-and-back-again-the-product">Arc republished it on August 25th</a>. As it grows, more and more of us will “get used to” the way Arc behaves, slowly training ourselves to work with a new, empowering tool. We may already be speeding along the adoption curve if <a href="https://x.com/joshm/status/1699882576494592270?s=20">the attention paid to Arc by its competitors</a> is any indication.</p><p>It might be that this slow, organic, and unstructured method of training is the only way to get people on board with new, radically powerful tools. But what if we baked training into the way we interact with technology vendors? What if, instead of throwing out the manual when we receive a new device, we made it the most exciting part of the unboxing?</p><h3>Breaking the Cycle</h3><p>What I’m proposing is the technological version of a gym coach. Let’s call this role a <strong><em>computational coach</em></strong>. It’s someone you consult with. Their job is to familiarize themselves with the general principles of common problem domains. In addition to that, they should be familiar with basic computing principles, the workings and fruitful applications of a wide variety of tools, and their combination. Their guiding principle would be this: <em>what problems do my clients encounter and what computational processes can I help them build to solve them?</em></p><p>There are already people out there who run their lives with Apple’s Shortcuts and customized combinations of computational tools. There are <a href="https://shortcutsgallery.com/">whole sites</a> dedicated to sharing recipes with them. We tend to call these people “power users,” and there are certainly features that cater to those sorts of folks. Computational coaches would probably be cut from this cloth. However, they would have to apply all of the empathy and psychological tools of the pedagogical disciplines to help their clients. Computational coaches certified in various computational models could then help us integrate new tools when they are released.</p><p>Imagine, if you will, unboxing the iPhone 16. During the unboxing process, you’ve scheduled a call with an official Apple Computational Coach. You chat during the unboxing process. They learn that you love traveling and have two cats. You chat some more and they show you all of the bells and whistles characterizing this year’s model. However, because of this personal approach, they’re able to help you by demonstrating a workflow that never would have occurred to you — using a combination of new services in the Shortcuts and Home apps to automatically trigger your cat feeder when you’re away from home.</p><p>Like any coach worth their salt, though, they would prompt you to grow. Maybe this involves ideation sessions, where the coach and client share ideas for automations and the healthful applications of computational thinking. They would then message you and hold check-in sessions, to see how well you’ve practiced these new methods.</p><p>There are teams out there, Apple included, creating beautiful tools for thought. Their use, however, often requires the understanding and application of very specific, sometimes esoteric computational paradigms. Facility with these tools requires the same level of granular, targeted practice we would apply to any other complex tool. This training is lacking.</p><p>To get the ball rolling and help more of us acquire this training, we need to realign our incentives, such that computational vendors are motivated to ensure that the people consuming their products are using them as they were intended. <em>Computational Coaches</em> are one idea. Regardless of the specific services that are innovated to provide training, it’s clear that we need to tweak the narrative. “It just works” isn’t enough, so we should stop propagating it as a one-size-fits-all method of product development. If we get this right, it will allow normal computer users to craft the systems that will enrich their lives, instead of forcing them to wait for some software engineer somewhere to do it for them.</p><p>Being better problem solvers takes hard work. However, like any coach will tell you, that’s a necessary step to growth. So let’s do it. Just do it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=10793359ef93" width="1" height="1" alt=""><hr><p><a href="https://uxdesign.cc/the-missing-link-10793359ef93">The missing link in Apple’s event</a> was originally published in <a href="https://uxdesign.cc">UX Collective</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>