<?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 Steven Adler on Medium]]></title>
        <description><![CDATA[Stories by Steven Adler on Medium]]></description>
        <link>https://medium.com/@sjgadler?source=rss-c04544a536f5------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*PGH_jWnHgKPbabysANYPhQ.jpeg</url>
            <title>Stories by Steven Adler on Medium</title>
            <link>https://medium.com/@sjgadler?source=rss-c04544a536f5------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 14:24:56 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@sjgadler/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[The moral limits of (artificial) intelligence]]></title>
            <link>https://medium.com/@sjgadler/the-moral-limits-of-artificial-intelligence-e3bf971a57e0?source=rss-c04544a536f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/e3bf971a57e0</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[philosophy]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Steven Adler]]></dc:creator>
            <pubDate>Thu, 17 Oct 2019 22:28:33 GMT</pubDate>
            <atom:updated>2019-10-17T23:00:21.945Z</atom:updated>
            <content:encoded><![CDATA[<p><em>There are many areas in which increasingly-capable artificial intelligence (AI) will not be a panacea for the moral challenges ahead. Indeed, an increasingly-capable AI may present further challenges, when considering human values, that are best resolved prior to this increase in capabilities</em>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/0*QPS4tN_rETcX_yrS" /></figure><p>A common quip in economics holds that an economist can teach you how to how to optimize for your values, but that they can’t tell you <em>what</em> to value — and this holds no matter the proficiency of the economist. Similarly, I believe we are entering a stage of development in which we need to differentiate among the AI challenges and opportunities ahead: those that are primarily technical dilemmas, and those whose thorny issues will not be unlocked by further advances (and indeed, might even be exacerbated).</p><h4>Increasingly-capable AIs can enable conduct not previously considered by society — and present new values-based challenges</h4><p>Cathy O’Neil’s <em>Weapons of Math Destruction</em> (perhaps moreso about ‘big data’ than about AI or machine learning, though often these overlap) showcases several instances in which increasingly powerful machine techniques open up new capabilities to society, and in turn require humans to question their values.</p><p>For example, O’Neil highlights a rise in algorithmic schedule-setting at restaurants and other retail locations, particularly around part-time workers and allotting them just enough hours that they are both well-utilized <em>and</em> unentitled to benefits.</p><p>While managers certainly could have worked toward this goal prior to algorithmic scheduling, it was far more work manually, and not as likely to have occurred. Now, achieving this outcome is fairly trivial, and at little cost to the manager; how ought society respond?</p><p>The fundamental tension is that something feels <em>off</em> about this conduct to many folks, but there remains a gap in legislation and other norms that allow businesses to operate in such a manner.</p><p>Prior to the rise of the algorithmic techniques, such conduct wasn’t practical or widespread, so society need not have consciously reflected on its chosen outcomes. Now that it has become tenable, a gap may have emerged between “what society wants” and “what society allows”, driven by a new set of unexpected behaviors. Regrettably, it is not difficult to imagine other scenarios in which AI might open up new capabilities with previously-unspoken norms and preferences, resulting in society condoning behaviors it may not have directly chosen.</p><p>On the other hand, when society does have well-grounded values to optimize around (for instance, if each worker had complex shift preferences, and a manager wanted to create an equitable schedule in light of these), the algorithm will likely be more capable than the human manager, who may get overwhelmed by the variables and tradeoffs.</p><p>In this sense, the algorithms underpinning AI systems are neither morally good nor bad — but they do open up new frontiers that should urge societies to be conscientious in the values they select. If AI continues to progress without our consciously reflecting on the ends to which we want to deploy it, we may end up in a future chosen in large part by accident. <em>[A common view in technology governance is that technology is not neutral, but rather can encode the values and preferences of those who design the technology and for whom the technology is principally designed. I think that is quite often true, and also that neutrality is a bit different than what I mean by AI systems being ‘neither (inherently) good nor bad’.]</em></p><h4>Even when people fault-find in the capabilities of today’s AI systems, improving these capabilities may not address the root of a challenge</h4><p>The above presents an example of AI in which increasing capabilities pose new challenges for society — but what about challenges of today’s AI, in which the systems are not sufficiently robust to accomplish their goals? Even in these instances, the technical challenges that seem to be at the heart of the AI use-case may be masking over other complications — and thus be insufficient to deciding society’s appropriate course.</p><p>To ground this conversation in an AI hot topic, consider the case of AI and facial recognition systems: Researchers such as Joy Buolamwini (of the Algorithmic Justice League and the MIT Media Lab) have <a href="https://medium.com/@Joy.Buolamwini/when-ai-fails-on-oprah-serena-williams-and-michelle-obama-its-time-to-face-truth-bf7c2c8a4119">documented</a> that many commercially-available facial recognition systems perform less well on dark faces, and are also prone to misgendering groups such as black women. Even though companies’ products have generally improved over time, many point to these flaws as a reason to avoid facial recognition technologies, as they are not yet robust and sufficiently high-confidence.</p><p>Suppose that these facial recognition systems became wholly accurate, however, and that the systems now functioned perfectly for their intended purposes; would that mean we have resolved the moral challenges of using AI for facial recognition?</p><p>I think the answer is a fairly strong “no, the moral challenges are not resolved” (there are other important issues to contemplate, outlined below), and yet an analogy to uses of human intelligence may prove instructive: Would anyone argue that, if a cop in a small town recognizes a person carrying out a murder and knows this person’s identity, the cop will be doing something wrong by identifying the murderer? If not, then what is wrong with deploying this form of intelligence at scale — indeed, perhaps even at higher accuracy?</p><p>For one, even technically-perfect AI systems may have negative consequences associated with how these systems are deployed. For instance, deploying AI at scale for policing might have disparate impacts on communities of color. If deployment of these technologies were to resemble current patrol densities, we might see poorer, non-white defendants repeatedly get arrested for crimes, whereas wealthier people face fewer of the consequences. While society may reasonably want to enforce and punish crimes, the unequal application of its laws would seem to be a subpar outcome, and one that society may think twice before increasing through automated police systems.</p><p>On the other hand, some have argued that AI-based facial recognition could reduce the disparities of enforcement across neighborhoods by allowing police to be ever-present through relatively lower-cost technology, rather than requiring officers on the street. If this were the case, and the facial recognition technology continued to work flawlessly, would we then have resolved facial recognition’s moral challenges?</p><p>Even in this instance, resolving the technical challenges of facial recognition and also rooting out bias in its deployment may not be enough; we still need to consider the system’s fundamental goals of the system and whether we want to live in a society in which these come to fruition.</p><p>Surely there is upside to deterring crimes and arresting those who commit them, but what of a society in which all movements are captured by closed-circuit cameras and the citizenry are always under surveillance? How might the system as a whole stifle freedom of expression and affect reasonable expectations of privacy/control of one’s personal data?</p><p>I do not have robust answers to these questions, but I do know that society ought contemplate them deeply before proceeding with new technologies — even ones that have passed stringent technical testing. Arguments against technologies may be couched in terms of technical objections, but often times there remain deeply-held values conflicts as well — and all the technical solutions in the world will not disentangle these.</p><p>To be clear, I do not mean to suggest that Joy Buolamwini objects to facial recognition purely on technical grounds; I know she holds other concerns as well. But it’s worth reflecting that, when we see coverage in the media alluding to technical challenges in AI systems, these reports may be illustrating only one barrier — and not necessarily the real crux of whether society ought adopt a certain technology.</p><h4>Thinking critically: are these technical problems or social problems?</h4><p>In sum, I want to encourage reflection on the part of readers and citizens when considering what AI norms one wants to support. When hearing an objection, ask yourself, “Is this a technical problem, or a social problem?” In particular, consider more concretely the interplay between these categories and whether one form of objection may be masking the other. To truly evaluate our values and decide what AI future we wish to live in, it is necessary to be precise in our formulations and to consciously deploy energy to the problems most worth solving.</p><p>One important assumption underpinning this article is that notions of values and morality may not be ‘solved’ by further advances in intelligence — or at least may not be solved immediately, in which case there is potential of serious moral quandaries in the interim. This is of course a field of active research for many, and in my mind is not a closed question. For those looking to read further on this topic, one book that I enjoyed (though it has been a few years since reading it) is Wendall Wallach’s <a href="https://www.oxfordscholarship.com/view/10.1093/acprof:oso/9780195374049.001.0001/acprof-9780195374049"><em>Moral Machines</em></a>.</p><h4>Steven Adler is a former strategy consultant focused across AI, technology, and ethics.</h4><h4>If you want to follow along with Steven’s projects and writings, make sure to follow <a href="https://medium.com/@sjgadler/">this Medium account</a>. Learn more on <a href="http://www.linkedin.com/in/sjgadler">LinkedIn</a>.</h4><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e3bf971a57e0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Artificial Intelligence and Algorithmic Risk Assessment Tools: The Beginnings of a Common Approach…]]></title>
            <link>https://medium.com/@sjgadler/artificial-intelligence-and-algorithmic-risk-assessment-tools-the-beginnings-of-a-common-approach-2fcaae546a72?source=rss-c04544a536f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/2fcaae546a72</guid>
            <category><![CDATA[ethics]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Steven Adler]]></dc:creator>
            <pubDate>Tue, 09 Jul 2019 06:42:12 GMT</pubDate>
            <atom:updated>2019-07-09T06:42:12.005Z</atom:updated>
            <content:encoded><![CDATA[<h3>Artificial Intelligence and Algorithmic Risk Assessment Tools: The Beginnings of a Common Approach to High-Stakes Domains?</h3><p>I have the great pleasure of working as Chief of Staff at the <a href="https://www.partnershiponai.org/">Partnership on AI</a>, a nonprofit dedicated to ethical AI use at many of our partner organizations (groups such as Google, the ACLU, etc). <em>[The views represented in this piece are my own and not necessarily those of my employer.]</em></p><p>Recently we <a href="https://www.partnershiponai.org/artificial-intelligence-research-and-ethics-community-calls-for-standards-in-criminal-justice-risk-assessment-tools/">published a report</a> on artificial intelligence technologies that have become increasingly prevalent in the US criminal justice system, calling for caution and outlining a series of ten unfulfilled requirements such a system should meet before being used in high-stakes domains (for instance, that predictive tools should use confidence intervals to reflect their level of uncertainty to users).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/592/1*PBpcj7Q-rFixmC2mprPJFA.png" /><figcaption>This is no ordinary “scales of justice” illustration, but rather one composed of words — and technology-related words at that!</figcaption></figure><p>Intriguingly, I wonder whether this report may advance progress not only in the criminal justice realm from which it hails, but in other high-stakes domains as well. <strong>That is, to what extent may ethical progress generalize across important AI use-cases?</strong></p><p>To make these musings a bit more concrete, let me highlight the three categories of concerns we raised about these tools, which are often used for purposes such as determining risk in pre-trial detention assessments:</p><ol><li>Concerns about the validity, accuracy, and bias in the tools themselves;</li><li>Issues with the interface between the tools and the humans who interact with them; and</li><li>Questions of governance, transparency, and accountability.</li></ol><p>Of note, very little in the high-level descriptions of these categories is specific to criminal justice (though certainly we get more specific to criminal justice in the recommendations themselves, and in the underlying research that informs these recommendations). Does this mean we could port similar forms of recommendations to other highly consequential settings, like AI for healthcare?</p><p>In my personal view, the answer is quite likely yes: A framework of tool validity; tool interface (how humans come into contact with the tool); and tool accountability seems to encompass many concerns common to other domains.</p><p>For instance, in the aforementioned healthcare example, one may wish to ensure that medical data are drawn from a relevant and reflective cross-section of society; that doctors and nurses are properly trained on the systems that will help to augment their medical expertise; and that there are notions of responsibility and accountability in place for the instances in which bad outcomes befall patients. The core principles of our report seem to hold in this instance, at least in broad strokes, though when it comes to complicated technological deployments, we know the devil is often in the details.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/0*ar337kHIBVjzUDFy.jpg" /><figcaption>I like this illustration because it demonstrates that 1) robots are capable of pondering, and 2) stethoscopes are the materials most worth pondering, once we reach the age of sentient robots. (Yes, this is sarcasm.)</figcaption></figure><p>Another way of breaking down challenges for AI usage in these high-stakes domains, albeit somewhat orthogonal to the above, would be to bucket concerns about an AI deployment by the major ethical camps they implicate: Considering the intentions of the relevant actors (deontological); the likely consequences of deploying the technology (consequentialist); and perhaps some virtue ethics strains of what it says about us as a society to be using technology in a certain way (virtue ethics).</p><p>In this sense, one may also come up with a useful, replicable framework for considering AI use-cases, hopefully without first needing to solve the many intractable challenges of ethics! (For an interesting analysis of all the different ways groups have begun to grapple with approaches to AI ethics, I recommend the <a href="https://ai-hr.cyber.harvard.edu/images/primp-viz.pdf">Principled Artificial Intelligence project</a> from the Berkman Klein Center, one of PAI’s Partners.)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/568/1*PtacO59rXZg4AtFWw_1ZPA.png" /><figcaption>From the <a href="https://ai-hr.cyber.harvard.edu/images/primp-viz.pdf">Principled Artificial Intelligence project</a> of the Berkman Klein Center.</figcaption></figure><p>One useful outcome of such frameworks is that, depending how well the one picks apart the different layers of an AI system, these frameworks can allow for more fine-grained conversation and introspection about where exactly one is uncomfortable with an AI deployment.</p><p>For instance, facial recognition is a highly-fraught topic in the AI ecosystem today, replete with objections at many different levels of abstraction — and I have a hypothesis that some parties are engaging with one another as ships passing in the night, rather than being able to drill into their actual disagreements.</p><p>Consider a privacy idealist who objects to public surveillance with AI regardless of the technology’s prospective ability to be racially unbiased — or indeed, even if it were usable to counter racial bias. There are important nuances to views such as these, and it is my hope that a sufficiently good framework for AI use-cases could lend itself to easily mapping the various objections people have about certain applications, and in turn better understanding and engaging with others in debate.</p><p>As you may have inferred from the speculative tone of this post, I certainly haven’t cracked this problem of a common framework just yet, though it is one that occupies a lot of my thinking time. If you have any ideas along this front or want to grab coffee and chat, let me know — I’m always looking to expand my thinking here.</p><h4>Steven Adler is a former strategy consultant focused across AI, technology, and ethics.</h4><h4>If you want to follow along with Steven’s projects and writings, make sure to follow <a href="https://medium.com/@sjgadler/">this Medium account</a>. Learn more on <a href="http://www.linkedin.com/in/sjgadler">LinkedIn</a>.</h4><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2fcaae546a72" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Corpus Callosum, and Thoughts on Human-AI Collaboration & Consciousness]]></title>
            <link>https://medium.com/@sjgadler/the-corpus-callosum-and-thoughts-on-human-ai-collaboration-consciousness-6d8f8e613200?source=rss-c04544a536f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/6d8f8e613200</guid>
            <category><![CDATA[consciousness]]></category>
            <category><![CDATA[neuroscience]]></category>
            <category><![CDATA[psychology]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Steven Adler]]></dc:creator>
            <pubDate>Mon, 03 Dec 2018 06:04:51 GMT</pubDate>
            <atom:updated>2019-02-26T06:26:34.019Z</atom:updated>
            <content:encoded><![CDATA[<p>This week I read <a href="https://brianchristian.org/the-most-human-human/">The Most Human Human</a>, by Brian Christian, the co-author of <a href="http://algorithmstoliveby.com/">Algorithms to Live By</a>. In the course of reading it, my mind ran with some interesting ideas, which I am quite sure are half-baked, but which I’m recording here for posterity and on the off-chance they spark a connection for someone else.</p><p>Using canonical studies about split-brain patients — those who have had their corpus callosum severed, and thus are unable to pass neural signals between the hemispheres of their brain — Brian Christian argues that maintaining one coherent self is often the product of high-bandwidth neural communication: Because the brain’s hemispheres can ferry information back and forth, a human perceives itself as one coherent being. These beings are generally resistant to the tricks and manipulations that fool humans without this means of neural communication.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/323/0*wkeQ60cL0Na0JSCx.jpg" /><figcaption>The brain’s hemispheres are connected by the corpus callosum — a bundle of ~200 Mn electricity-conducting nerve fibers— helping to integrate a series of abilities that generally reside in a particular hemisphere.</figcaption></figure><p>Yet, as Christian points out, this one-self hypotheses — the one self we see is all that exists — is only one of many possible interpretations. Another such interpretation is that, even though the brain’s hemispheres can communicate in a way that produces an appearance of constancy, the hemispheres may remain separate beneath-the-surface nonetheless. The “I” so often expressed by our verbal left hemisphere may not be a representative union of our sets of experiences, but rather be privileging the side that can speak.</p><p>I wonder about the implications of such split-brain research for how we conceive of human-AI collaboration, and in particular for systems in which one agent needs to process information it receives from the other type of agent (for instance, a human rendering a medical diagnosis in which it has been informed of an AI’s diagnostic reading, and some limited information about how it came to this conclusion). To what extent will these agents continue to operate independently from one another, and when they interact, will they be focused upon complementary domains, or rather on providing divergent takes on similar domains?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Y6boV3GhRW-vuHOd.jpg" /><figcaption>The accompanying Economist <a href="https://www.economist.com/leaders/2018/06/07/ai-radiology-and-the-future-of-work">article</a> to this image does a good job of highlighting the various roles of human-AI collaboration in radiology, an oft-mentioned field of AI disruption.</figcaption></figure><p>Separately, if split-brain humans are inclined toward confabulating reasons for their actions — drawing a shovel and saying it is to scoop up after a chicken, when in fact it is because they were seeing a snowy scene — what might this mean for our own abilities to know the drivers of our decisions in an AI-integrated world?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/269/0*WQXRACGldT65HEMB.jpg" /><figcaption>Split-brain patients demonstrate cases in which we aren’t aware of the drivers of our decision-making — and the lengths to which our brain goes to seem rational and internally consistent. <a href="http://thebrain.mcgill.ca/flash/capsules/experience_bleu06.html">Source</a>.</figcaption></figure><p>Currently most AI interfaces are removed enough from our bodies that we can consciously recognize we are being provided with information (though perhaps are not deeply conscious of exactly how we are drawing inferences from this information). In the future, if brain-computer interface companies like Elon Musk’s <a href="https://en.wikipedia.org/wiki/Neuralink">Neuralink</a> succeed, will we have either the physical or temporal latency to clue us into what is happening?</p><p>The corpus callosum and the cases of split-brain patients also raise some interesting questions about consciousness and how we may think about consciousness in the domain of sufficiently-advanced AI systems. This is admittedly <em>slightly</em> out of my depth, but a few relevant factors seem to be: 1) whether consciousness is an emergent property of information-flow and processing (versus, for example, some meta-physical property that doesn’t flow forth from computation), and 2) whether this emergent property of consciousness is ‘substrate independent’ (that is, not reliant upon the particular hardware on which the computation is run).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/740/0*AL1AJw17-FrGV77l" /><figcaption>(Spoiler alert, Black Mirror, Season 3) If human consciousness is indeed substrate independent — alongside many other challenging constraints, like the ability to replicate a <strong><em>particular</em></strong> consciousness on another piece of hardware — perhaps we will end up with a world like San Junipero. For better or for worse.</figcaption></figure><p>If both 1 (emergent property of computation) and 2 (substrate independence) are true, and if we deem split-brain patients to have multiple divergent consciousnesses (from a cursory Google search, the <a href="https://www.psychologytoday.com/us/blog/the-recursive-mind/201711/split-brain-researchers-are-split">psychological field seems to be divided</a>), then we may start experiencing ‘conscious’ (or conscious-seeming) AI systems at a sub-human brain scale of complexity. Granted, there are many animals with sub-human brains that we already generally regard as conscious, and many approaches toward <a href="https://en.wikipedia.org/wiki/Artificial_general_intelligence">AGI (artificial general intelligence)</a> aren’t focused on brain emulation anyway, so perhaps this isn’t all that novel — but it feels worth saying.</p><p>Unfortunately (or fortunately), we might not get too much more data about split-brain patients in the near future — so the above musings will have to suffice for now. I hope the thoughts are of use in sparking ideas and in considering some unusual perspectives on AI &amp; human intelligence:</p><blockquote>The split-brain era is now largely over; because the drug routines for epilepsy are now more effective and more widely used, the split-brain operation is largely restricted to more severe cases, and surgery often involves only a partial section of the commissures.</blockquote><h4>Steven Adler is a former strategy consultant focused across AI, technology, and ethics.</h4><h4>If you want to follow along with Steven’s projects and writings, make sure to follow <a href="https://medium.com/@sjgadler/">this Medium account</a>. Learn more on <a href="http://www.linkedin.com/in/sjgadler">LinkedIn</a>.</h4><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6d8f8e613200" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Care about AI ethics? What you can do, starting today]]></title>
            <link>https://medium.com/artificial-intelligence-ai-for-social-impact/care-about-ai-ethics-what-you-can-do-starting-today-882a0e63d828?source=rss-c04544a536f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/882a0e63d828</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ethics]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[product-management]]></category>
            <dc:creator><![CDATA[Steven Adler]]></dc:creator>
            <pubDate>Tue, 25 Sep 2018 18:51:00 GMT</pubDate>
            <atom:updated>2019-01-22T23:59:20.281Z</atom:updated>
            <content:encoded><![CDATA[<p><em>This week I had the great pleasure of meeting </em><a href="https://en.wikipedia.org/wiki/DJ_Patil"><em>DJ Patil</em></a><em>, Chief Data Scientist of the Obama Administration and former Chief Scientist at LinkedIn. Below is my summary and light analysis of his </em><a href="https://santaclarauniversity.hosted.panopto.com/Panopto/Pages/Viewer.aspx?id=482fe5ba-683a-43f4-a297-a93601163fdb"><em>talk</em></a><em> (ppt slides </em><a href="https://prezi.com/yhktt59vph3c/dataethics/?utm_campaign=share&amp;utm_medium=copy"><em>here</em></a><em>), which covered data science broadly (though I found quite relatable to AI ethics).</em></p><p><em>For further detail, I highly recommend DJ and his co-authors’ (Hilary Mason and Mike Loukides) excellent and free e-book, “Ethics and Data Science”, available </em><a href="https://www.amazon.com/dp/B07GTC8ZN7/ref=cm_sw_r_cp_ep_dp_klAOBb4Z72B4G"><em>here</em></a><em>. Many thanks to the </em><a href="https://www.scu.edu/ethics/about-the-center/"><em>Markkula Center for Applied Ethics at Santa Clara University</em></a><em> and other hosts for helping to convene this discussion, as part of their </em><a href="https://law.scu.edu/hightech/htli-artificial-intelligence-for-social-impact-speaker-series/"><em>AI for Social Impact Speaker Series</em></a><em>.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/465/1*cltiI2mN4ppKMSctfpYHXA.png" /><figcaption>The speaker and hosts of the event, “ How to Do Good AI and Data Science (By Living Your Ethics)” with DJ Patil</figcaption></figure><h4>Key questions for AI ethics / data ethics</h4><ul><li>What standards and practices ought to be promulgated across AI and data science, as they grow in importance as fields?</li><li>What mechanisms exist for promulgating these standards and practices, from the lens of a country/regulatory body, institution (such as a large for-profit or university), and individual?</li><li><strong><em>NB:</em></strong><em> I think the questions about governments and institutions are quite interesting, and deservedly getting much attention. Because the individual question is perhaps rarer and also more directly actionable, I add further color on it below.</em></li></ul><h4>Frameworks for understanding and managing AI/data science’s opportunities and challenges</h4><p><strong>DJ’s “5 Cs”</strong> — “a framework for implementing the golden rule of data” (treat others’ data as you’d like your own data to be treated; what dimensions do people care about when sharing their personal data?)</p><ol><li>Consent — Informed agreement on what data is to be collected and to what end; notions of consent and its limits can vary by geography (e.g., EU and GDPR).</li><li>Clarity — Even when there appears to be consent, how confident are we that end-users understand the terms? This can be affected by technical complexity of usages and by legalese documents, such as form Terms and Conditions.</li><li>Consistency (and trust) — Ensuring systems operate as expected and are reliable, so that users have accurate mental models of how their data will be used and the data-holder’s responsibilities.</li><li>Control (and transparency) — Understanding what data is held, on what dimensions, and to what use, on an ongoing basis.</li><li>Consequences — Anticipating how data might be misused or abused, either by malevolent actors or by well-intentioned actors with unforeseen circumstances.</li></ol><p><strong>DJ’s “data science project checklist” </strong>— what are the questions a data science team should ask before undertaking or launching a project, akin to other product review questions a team may undergo (e.g., will this new launch cannibalize our existing products)</p><blockquote>1. Have we listed how this technology can be attacked or abused?</blockquote><blockquote>2. Have we tested our training data to ensure it is fair and representative?</blockquote><blockquote>3. Have we studied and understood possible sources of bias in our data?</blockquote><blockquote>4. Does our team reflect diversity of opinions, backgrounds, and kinds of thought?</blockquote><blockquote>5. What kind of user consent do we need to collect to use the data?</blockquote><blockquote>6. Do we have a mechanism for gathering consent from users?</blockquote><blockquote>7. Have we explained clearly what users are consenting to?</blockquote><blockquote>8. Do we have a mechanism for redress if people are harmed by the results?</blockquote><blockquote>9. Can we shut down this software in production if it is behaving badly?</blockquote><blockquote>10. Have we tested for fairness with respect to different user groups?</blockquote><blockquote>11. Have we tested for disparate error rates among different user groups?</blockquote><blockquote>12. Do we test and monitor for model drift to ensure our software remains fair over time?</blockquote><blockquote>13. Do we have a plan to protect and secure user data?</blockquote><p>I should note as well here that while cybersecurity is directly present in a few of the above questions (for instance, in questions 8, 9, and 13), this was also a central focus of DJ’s talk, e.g., in describing <a href="https://en.wikipedia.org/wiki/Red_team">‘red-team’</a> approaches to both infiltrating data systems and to how systems may be misused or abused.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1023/0*MICAEx3MB5L52Vuv.jpg" /><figcaption>DJ highlighted the “Hack the Pentagon” program as a success of red-teaming and bringing cybersecurity into the realm of data science — even within a highly-confidential organization.</figcaption></figure><h4>DJ’s suggestions for individuals</h4><p>Given DJ’s perspective and his work across the data science ecosystem, what can one start doing today that might have an impact?</p><ul><li>If working on a data science project, take a look at the checklist outlined above. Could your team use this? If not, what questions need to be added or subtracted to make it more tenable?</li><li>Ask yourself: If, in the course of a project, you were to run into a challenge with the 5 Cs or a question from the checklist, do you know the resolution mechanism? If not, can you speak with your manager or team about formalizing a process (including, perhaps, a neutral third-party mediator/ombudsperson)?</li><li>When interviewing candidates to join your team, can you screen for amount of thought given to ethical tradeoffs, and ensure applicants are aligned with the vision you’d like your team to embody? (Where do AI and data ethics fit into your team’s notion of cultural fit?)</li><li>When interviewing as a candidate at companies, can you ask them how they resolve data ethics challenges internally, and how they ensure their products are used to good ends? (DJ raised a good point here that if this question became common practice, you’d certainly see more activity around the field — this ties into the bottom-up, worker-demand side of the economic equation I allude to below.)</li><li>When things go wrong (or narrowly avoid going wrong) in a project, does your team already have a post-mortem process for analysis? If not, can you suggest one oriented around the checklist above? (How can your team ensure appropriate fact-finding in the short-term and proper accountability in the long-term to help inform these practices?)</li></ul><h4>One additional framework for changing AI/data science ethics</h4><p><strong>An economic lens </strong>— though this framework was not called out directly, I view it as an essential theme in DJ’s talk and in my conversations with other participants — how to alter market dynamics (both supply- and demand-side) such that practicing strong data ethics becomes the norm.</p><p>Of course, this perspective is an oversimplification of the realities of institutional life and decision-making. At the same time, I think it is a useful oversimplication for considering how change may come, and where collective action may help.</p><p>As a starting point, I believe that nearly all AI scientists and adjacent stakeholders want very badly to do the ‘right’ thing — both deploying technologies toward great ends and avoiding misuses along the way.</p><p>Despite these inclinations, individuals within institutions can sometimes deviate from ideal behavior, based in part on lacking core tools to decisively answer the ethical questions, and in part from their own economic challenges and difficult incentive-structures.</p><p>To bring systemic change, therefore, we may find greater success when we couch practical solutions within people’s and institutions’ economic realities and ensure that ethics and market pressures function compatibly.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/333/0*yrdHf6jW744fVzRp.jpg" /><figcaption>Market outcomes are not guaranteed to have been ethical — but aligning ethics with market incentives may make for an easier case for action.</figcaption></figure><p>Consider a hypothetical individual at a company, who is evaluated principally by the output of code or product features they ship, without an ethical evaluation factoring in; this individual will have good reasons to overlook gray areas that would delay their practices, particularly if flagging these issues may not even result in clear resolution.</p><p>“Pulling the production-line cord” and halting a project’s progress may not be valued by the company (and in turn by the individual) if these traits do not show up positively in one’s reviews. Individuals need to feel empowered and actively encouraged to take such actions — not just that these actions are permissible. (I felt that my former employer, McKinsey &amp; Company, really excelled on this front, emphasizing the <a href="https://hbr.org/2017/01/true-leaders-believe-dissent-is-an-obligation"><em>obligation to dissent</em></a> at all levels and trying to embody non-hierarchical decision-making.)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/808/0*WnS0_S8mhj2904Pn" /><figcaption>DJ highlighted Toyota’s use of “Andon Cords” in their manufacturing processes to empower any worker to stop production, should they see a sufficiently glaring issue. To encourage this conduct, companies must then reward such workers, rather than punish them as whistleblowers.</figcaption></figure><p>At an institutional level, I suspect that marketplace demands may encourage companies to continue launching AI/data science use cases, even if open questions remain about data usage, so long as the companies continue to be rewarded. (Fear of poor public perception, of course, is a powerful deterrent to hasty launches.)</p><p>Yet in practice, company policy is not as simple as deciding <em>a priori</em> what will maximize profits, and then magically willing it; there is an additional market dynamic at play. Within companies, well-positioned and ideologically-aligned individuals may be able to affect launch trajectory, particularly if they have unique skills or a lot of clout in the organization. These mechanisms of influence include by individually refusing to work on certain projects, or by helping to spark conversation against certain projects. In this sense, powerful workers with a strong sense for ethical guardrails can make a difference, at least within certain companies.</p><p>While individuals within companies can make attempts to steer the ship, and indeed often succeed, it remains a reality that many US companies <em>do </em>operate with a focus on shareholder value in many parts of their operations. A framework for AI/data science ethics, at least in the short-term, likely involves operating within that framework, or at least acknowledging its practicalities (say, couching objections in difficulties of recruitment, retention, or implementation feasibility should certain scientists not endorse a mission).</p><p>None of these solutions nor frameworks is a panacea; in some cases, they leave to-be-determined what specific ethics folks ought to embody (what is the vision your team is driving toward in its use of AI/data science, and from where does this hail), and in others they are smaller incremental change, rather than top-down directives that may eventually achieve more scale. (Though taken as a sum, many smaller-scale changes may amount to broader change, or may help to stoke larger-scale actors.)</p><p>Still, I found it both inspiring and empowering to leave DJ’s talk with a concrete set of actions one can take to “live their ethics” in AI and data science more generally. I am very grateful to have had this opportunity, and hope that my synthesis and thoughts are useful to others grappling with these questions.</p><h4>Steven Adler is a former strategy consultant focused across AI, technology, and ethics.</h4><h4>If you want to follow along with Steven’s projects and writings, make sure to follow <a href="https://medium.com/@sjgadler/">this Medium account</a>. Learn more on <a href="http://www.linkedin.com/in/sjgadler">LinkedIn</a>.</h4><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=882a0e63d828" width="1" height="1" alt=""><hr><p><a href="https://medium.com/artificial-intelligence-ai-for-social-impact/care-about-ai-ethics-what-you-can-do-starting-today-882a0e63d828">Care about AI ethics? What you can do, starting today</a> was originally published in <a href="https://medium.com/artificial-intelligence-ai-for-social-impact">Artificial Intelligence (AI) for Law, Social Impact, and Equity</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A perfect week! In crosswords, that is -]]></title>
            <link>https://medium.com/@sjgadler/a-perfect-week-in-crosswords-that-is-6408dbe27c5a?source=rss-c04544a536f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/6408dbe27c5a</guid>
            <category><![CDATA[crossword-puzzles]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[goals]]></category>
            <dc:creator><![CDATA[Steven Adler]]></dc:creator>
            <pubDate>Sun, 09 Sep 2018 20:29:51 GMT</pubDate>
            <atom:updated>2018-09-09T20:30:10.889Z</atom:updated>
            <content:encoded><![CDATA[<p>Most of my musings on Medium relate to AI and what we may expect from the future — but I want to take a break from those topics to celebrate an important accomplishment: I finally achieved a perfect week in the New York Times crossword!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1015/1*0hyTN_jRrnXP-gYjzXtxpA.jpeg" /><figcaption>It won’t be a ‘perfect month’, given my non-completion of the first Saturday and my non-attempt of the first Sunday, but the week is perfect indeed</figcaption></figure><p>This is a goal I’ve been setting toward for a while, likely dating back to last September (a year-to-the-day, actually) when I first completed a New York Times Saturday crossword. I’ve waxed and waned on my commitment to it over time — the thing about a perfect week is that it requires ‘gold stars’, which themselves require completing the puzzle perfectly on the day it arrives, so not much room for schedule flexibility — but it feels great to have knocked this out.</p><p>Interestingly, I drew some inspiration for my approach to the crosswords from my own work in AI (a longer post I’ve been meaning to write, given my interest in the two fields). Suffice it to say, I got stuck in a bit of a local maximum in debugging today’s puzzle, and needed to inject a bit of randomness to shake myself free and keep “climbing that hill”. [<em>NYT crossword spoilers below</em>]</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/289/1*odjlWJrXcuofDZyK7L-NdQ.png" /><figcaption>For some reason, I was convinced I’d seen ‘ABS’ as the answer previously to ‘Small balls’ … Once I shook myself free of that, the rest of the corner came much more quickly</figcaption></figure><p>My overall time for the puzzle was a glacial two hours, of which I’d estimate close to an hour-twenty was spent trying to fill and debug the top-left 3x3 (entirely blank at that point), and scanning the rest of the puzzle for possible mistakes. Once I flipped ‘ABS’ to ‘BBS’, I was probably done in 2–3 minutes.</p><p>One other interesting motif of this puzzle was the inclusion of a ‘rebus’, in which multiple letters occupy one square (sometimes themselves interjecting odd rules to the puzzle, like reading the answer to a clue in a curved manner).</p><p>For instance, this puzzle tackled the idea of ‘LOVEHATERELATIONSHIPS’ (63 Across), and four squares were populated such that they read LOVE horizontally, and HATE vertically (‘fourleafcLOVEr’ crossed with ‘cHATEau’; ‘sLOVEnia’ crossed with ‘wHATElseisnew’ — yes this was a pain).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/661/1*-dhxI1xzSgofidd1nuUotg.png" /><figcaption>When ‘FOURLEAFCLOVER’ did not fit outright for “Symbol of luck” (102 Across), I knew a rebus was likely afoot</figcaption></figure><p>I wonder how an AI could be trained to look for those oddities and to populate a puzzle accordingly. I suppose it helps, as a human, to know that Thursday and Sundays are the puzzles that most frequently include rebus, so we could pass that information along to an AI (or allow it to discover that tendency on its own through analysis of prior puzzles). Similarly, as a human, the tipping-point is often finding an answer you are very sure of, but that also does not fit the grid — so perhaps it wouldn’t be so hard after all.</p><p>Either way, I’m feeling quite proud to have knocked out this goal, and I’ve included the Saturday and Sunday puzzle grids below for anyone curious:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/925/1*fjHClFMEEU9dBYOS7jK2bQ.png" /><figcaption>The Saturday and Sunday grids, with the last-filled-in clues highlighted</figcaption></figure><h4>Steven Adler is a former strategy consultant focused across AI, technology, and ethics.</h4><h4>If you want to follow along with Steven’s projects and writings, make sure to follow <a href="https://medium.com/@sjgadler/">this Medium account</a>. Learn more on <a href="http://www.linkedin.com/in/sjgadler">LinkedIn</a>.</h4><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6408dbe27c5a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[“An Uber for Excel”: Reflections on AI, Augmentation, and Platforms]]></title>
            <link>https://medium.com/@sjgadler/an-uber-for-excel-reflections-on-ai-augmentation-and-platforms-2b333f05a9f1?source=rss-c04544a536f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/2b333f05a9f1</guid>
            <category><![CDATA[freelancing]]></category>
            <category><![CDATA[uber]]></category>
            <category><![CDATA[platform]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Steven Adler]]></dc:creator>
            <pubDate>Sat, 01 Sep 2018 06:45:07 GMT</pubDate>
            <atom:updated>2018-09-01T06:45:07.012Z</atom:updated>
            <content:encoded><![CDATA[<p>Even two years ago this ride may have posed a challenge — ride-share coverage in Westchester, summer 2016, was spotty at best — but on this night there was little difficulty: The driver arrived at my relative’s home outside New York City, pulled up directions via Waze, and we were off to New Jersey.</p><p>As we drove and made light conversation — accompanied by my listening to Presidential-candidate (and fellow Brown alum) Andrew Yang’s <em>The War on Normal People </em>on Audible— I had opportunity to reflect on the “new economy” in which we find ourselves.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*84MXsjGgmUN9zJuc.png" /><figcaption>Well worth a read (or listen) — and not just because Andrew went to Brown.</figcaption></figure><p>Though the driver had never been down to this part of New Jersey, nor had I, we continued our ride with turn-by-turn precision until the hotel came into sight, aided by a host of technologies that go beyond the skills formerly required of a driver. In this manner, the ride exemplified many of the boons — but also the challenges — of a forthcoming labor displacement; it also helped to stimulate my thoughts about transformations we may soon see in other forms of work.</p><p><strong>Technology as a new skill paradigm</strong></p><p>A common broad-strokes refrain about AI’s impact on the job market runs as follows:</p><ul><li><em>Many human jobs will be displaced as AI becomes increasingly capable of automating the tasks a particular job requires.</em></li><li><em>Historically, these automated jobs, which now require fewer people, have been replaced or supplanted by new categories of jobs, including jobs difficult to envision at the outset (e.g., social media managers, search engine optimizers).</em></li><li><em>Though this is the historical pattern, there is not guarantee that this job-replacement must continue forever, and there is disagreement about what percentage of jobs will become automated and at what pace.</em></li></ul><p>While I agree with most tenets of the above, this summary leaves unstated what I believe to be a major shift in the economy: AI and similar technologies are increasingly <em>automating the tasks</em> of a job, yes, but they are also <em>transforming the skills required</em> of many jobs, beyond merely mimicking the human tasks.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/657/1*NI6Cw1qVoDOeKz76v_wmmw.png" /><figcaption>I’ll confess to having Googled “weird jobs that exist today” — and the search results didn’t disappoint.</figcaption></figure><p>Consider the example of a driver: In the past, the core skills of a taxi driver may have been safe, attentive driving; knowledge of a city to efficiently pick up and route passengers; and work-ethic to grind out a full week’s work, even if activity seemed slow or conditions were suboptimal. Certainly these skills remain important today, but there is also more emphasis on literacy with technology, data, and apps/interfaces as an enabling skill:</p><ul><li>Surge-maps provided by ride-share companies and 3rd-parties help drivers to determine the peak hours and locations in which to drive</li><li>Apps such as Waze help drivers navigate efficiently in real-time, even in locales with which they are entirely unfamiliar</li><li>Detailed electronic records allow more-granular means of analyzing one’s financials and spotting trends in income, types of passengers, etc.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/520/0*7Vi88jKnFTtcEdLG.jpg" /><figcaption>Waze offers many features to its drivers — not all of which productivity-enhancing.</figcaption></figure><p>A driver who ‘works for’ Uber and is quite good at actually driving — but who struggles to utilize these tools — will likely not earn their maximum income; the core facets of driving are no longer enough to succeed.</p><p>Certainly part of what we’re seeing here is technology and comfort with it being valued in the driving market (e.g., enough know-how to install Waze, configure it to one’s liking, etc.), but what we’re also seeing is a <em>democratization</em> in expertise, driven by technological augmentation.</p><p>Whereas before a driver could build an advantage by by really learning an area’s routes, patterns, and idiosyncrasies, this data is now readily available without the upfront investment. Drivers and their expertise have become more interchangeable, fed by the platforms that call upon on-demand workers and augment them with the data of thousands of other drivers’ experiences. (This point about augmentation is perhaps best made in Tim O’Reilly’s <em>WTF? What’s the Future and Why It’s Up to Us</em>, with the example of Uber and Lyft drivers being able to see ‘hailing’ passengers, even blocks away and in awful weather conditions. How can a normal worker compete?)</p><p><strong>What does this look like for traditional office work?</strong></p><p>Building upon the lessons of drivers, augmentation, and Uber, we can extend the example to consider how similar types of technology may impact office work. In particular, we ought to recognize that the labor impacts of AI extend beyond automation, in which software can wholly replicate a human’s job or a significant portion of it; we should also be attuned to how AI may reduce the complexity or barriers-to-entry of human jobs in the interim, largely through technological augmentation (prior to wholesale automation).</p><p>For the purpose of this example, I’m going to discuss where my mind goes when I think of in-demand, relatively high-skilled office work: Microsoft Excel. As a former McKinsey consultant, I’m a bit too acquainted with Excel and its utility across many different jobs — business analytics, financial analysis, performance reporting, general data organization, and so on. For professionals across all sorts of jobs, Excel is routinely one of the top tools of the trade — so what would it mean for Excel skills to be “Uber”-ified?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/0*4j1MlN7kigGTa3iG" /><figcaption>Ah, what a pretty spreadsheet — with color-coding of some kind, too. Though I’m not envious of the user trying to operate Excel on a Mac …</figcaption></figure><p>Let’s start by addressing the traditional “Uber for X” conception: One of my realizations in reflecting on the new economy is that, generally, when people discuss “an Uber for X”, this description leaves a bit to be desired. Often this is in reference to a “push button for service” model for an industry, particularly one that cuts out middle-men and in which regular people monetize their own assets through micro-transactions — but I think this overlooks significant aspects of what would make for successful “Uber”-ification. We can step through some aspects of these transformations and extend them to consider the impact on office work.</p><p><em>Reliable on-demand labor</em> is certainly one aspect of successful “Uber”-ification, but for this to succeed at-scale the work products must either be sufficiently commoditized or the labor sufficiently skilled and curated to overcome concerns of platform users. While platforms for office workers, such as Upwork and Catalent, allow for easier sourcing of freelancers than previously existed, today they are still plagued by quality risks and transaction friction. How might office work evolve to meet this market demand, in advance of high levels of automation?</p><ul><li>Many types of office work today are <em>not</em> commoditized, and therefore platforms care deeply about matching the right talent to the right problems. For instance, many sites are implementing AI-powered features, such as highlighting promising candidates for a gig based on a combination of prior reviews, listed skills, and other criteria.</li><li>While this can help certain workers to receive new opportunities, it also erodes stability for workers who had previously built up a niche: Similar to an Uber driven being ‘hailed’ by a passenger out of vision’s range (who perhaps did not know the hot pickup spots), these discovery features may reduce barriers-to-entry of freelance office work, at the cost of existing frontrunners. Workers may no longer need pretty websites or successful SEO to attract attention; they can focus on their core skills, albeit with less of a competitive moat once they do get a foothold going. In this way, a good Excel freelancer may then double down on their Excel skills because their auxiliary skills are now less important.</li><li>To further extend the analogy to core software, imagine a world in which the operating tools, such as Excel, became capable of assessing a particular user’s skills. If Excel could detect and certify skill at particular analyses and work patterns, it could route workers to jobs suited for their skills, such as if Uber automatically matched chattier drivers with passengers who prefer small-talk. In the ideal version for productivity, this enhanced Excel may even be able to guide and instruct workers on analyses they otherwise could not take on. In the same way that Uber’s rating system and GPS navigation help a passenger know what they are getting, and help drivers perform at more consistent levels, these changes to office work would certainly reduce the need (and premium) for a worker’s expertise while stopping short of full automation. Workers would perform at a higher, more-reliable level than when not connected to the technological network.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*2xV4TpjZVp_o_NeX.png" /><figcaption>Upwork does a good job of measuring the skills needed and deployed across its platform — but even more powerful would be skill assessments built into the underlying software.</figcaption></figure><p><em>Democratized skill and efficiency </em>are other major (and related) factors that influence both the amount of people who can perform a given job, and also each individual’s productivity at that job. An increase in either of these factors drives down the demand for any individual worker, and in turn should depress the average wage for a job. (Economics digression: As work becomes sufficiently cheap, one might expect to see an offsetting increase in the total volume of work demanded from firms, as labor for their projects has become cheaper; the net impact on total labor spending is ambiguous depending on what economists call <em>price elasticity of demand </em>for labor at certain wage-levels … But like I said, I digress.)</p><ul><li>In the context of Excel and office work, democratizing skill and efficiency may entail making use of ‘simpler’ software defaults, shortcuts, and interfaces, some of which powered by AI. For instance, perhaps software users could describe analyses in natural language and have the software produce more-complex outputs like Pivot Tables, which today only a subset of users can create. Similarly, Excel could come pre-loaded with adaptive templates that auto-fit column widths and use other presentation-friendly tactics for which McKinsey analysts currently memorize keyboard shortcuts (ALT+h+o+i is a real crowd-pleaser).</li><li>One might also imagine democratization through easier access to add-on software or external research, such as CapIQ for pulling in financials, ideally with simple walk-throughs to guide even an Excel rookie. At McKinsey, we at times augmented our core client service teams with a Research &amp; Insights team, who were experts in market research reports and proprietary data sources. We invested quite heavily in these teams, but if access to these systems becomes less-specialized or built native into software, soon workers without the McKinsey infrastructure will be able to participate in large efficiency gains.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*GH5w3crqWcBiXALk.jpg" /><figcaption>There, you’re basically an Excel wizard now.</figcaption></figure><p>Finally, <em>integration of work across systems</em> is another large driver of efficiency, though often may not be classified as automation depending on the nature of these integrations. For instance, Uber’s incorporation of GPS directions into their app and their automatic reporting of finances from a particular ride certainly aid a driver’s productivity, but it isn’t clear to me that this is what anyone means by automation streamlining work. Regardless, this subarea is fairly interesting for possible productivity gains in office work, in part because of how much time workers spend working across disparate systems (email, Excel, PPT, ERP systems, etc.) and doing work that largely serves to transfer knowledge from one location to another.</p><ul><li>One ‘low-hanging’ but surprisingly difficult problem to crack is to easily port materials between Excel and PowerPoint — the most common presentation mode for the analyses one creates within Excel. (At many consulting firms, the business purchases for third-party software, such as ThinkCell, to help resolve the tediousness of moving data from one system to another.) If this were no longer such a bottleneck, how much more time could each office worker devote to the true knowledge work of their job? Yet we likely do not need full-scale automation (at least I hope not) to achieve this goal.</li><li>Similarly, in a future with enhanced productivity tools, a worker might be able to cross-off an item on Asana or Trello and be instantaneously provided a draft status update, as well as make changes in any linked documents in one fell swoop. Of course, AI/automation is not yet at that level (one can dream), and human judgment continues to be important in deciding how to present information and in what contexts. But in the not-too-far future, we may have technology that reduces the communication load of knowledge work and allows each worker to operate more productively — as well as requires new fluency with technologies and skills unneeded in certain jobs today.</li></ul><p><strong>Wrapping up — what this means for workers today</strong></p><p>I’ve written a lot of words above, but what does it actually mean? I believe there are two important insights any knowledge worker ought to carry in mind:</p><ul><li><em>AI automation of tasks is part of the coming disruption to jobs, but it is only one piece of the equation.</em> A young worker choosing among prospective careers should be attuned not only to how automatable their job seems today — I’ll plug the McKinsey Global Institute’s <a href="https://www.mckinsey.com/featured-insights/digital-disruption/harnessing-automation-for-a-future-that-works">excellent research</a> on this front — but also to how technology might change particular careers and the skills required in advance of full-scale automation.</li><li><em>Given the uncertainty about future careers and skill-sets, the most important traits to cultivate are an ability to learn and a desire to be curious and self-driven in charting one’s path</em>. It is very difficult, if possible at all, to sit here in 2018 and know exactly how any career path will play out — and that’s where adaptability and agility come in. Regrettably, many of these ‘traits’ and opportunities will come cloaked in privilege, as it is much easier to be agile and throw yourself into a new field if you can sustain a new months without income to retrain — a luxury many do not have. But for those who can find time to invest in these traits, I would expect pretty good outcomes (or at least more enjoyment of learning overall).</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/891/1*QGsvKktdzNaLJImTRaPW3Q.png" /><figcaption>An example chart from MGI’s automation research; the <a href="https://www.mckinsey.com/featured-insights/digital-disruption/harnessing-automation-for-a-future-that-works">full report</a> is well worth a read.</figcaption></figure><p>In addition to these insights for workers, I believe there continues to be much to ponder for society about how it ought to engage with platforms and other sources of labor augmentation. Certainly we have seen platforms offer enhanced (or new forms of) productivity, and alongside them often consumer surplus — such as my relief at being able to take the Uber ride mentioned at the top article. But alongside these gains we have also seen the erosion of many workers’ economic sources of distinction, including in areas in which workers may have striven for several years or dedicated significant capital to building out these abilities.</p><p>I am hopeful society will consider how to balance these interests against one another, rather than take a wholesale view of automation/productivity as purely a gift or a curse. Ultimately the choice is with all of us, both as consumers and constituents, to support the services we wish to see in the world. This responsibility means that the onus is on to keep an eye on the nuances of the labor and technology discussion and to check media narratives, whether overly frothy or bleakly pessimistic, when they fail to capture a full picture of reality.</p><h4>Steven Adler is a former strategy consultant focused across AI, technology, and ethics.</h4><h4>If you want to follow along with Steven’s projects and writings, make sure to follow <a href="https://medium.com/@sjgadler/">this Medium account</a>. Learn more on <a href="http://www.linkedin.com/in/sjgadler">LinkedIn</a>.</h4><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2b333f05a9f1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Book recommendations from 2017: Emerging technologies, AI, self-improvement, and more]]></title>
            <link>https://medium.com/@sjgadler/book-recommendations-from-2017-emerging-technologies-ai-self-improvement-and-more-b3163f5a24ef?source=rss-c04544a536f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/b3163f5a24ef</guid>
            <category><![CDATA[books]]></category>
            <category><![CDATA[science-fiction]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[self-improvement]]></category>
            <dc:creator><![CDATA[Steven Adler]]></dc:creator>
            <pubDate>Tue, 26 Dec 2017 16:29:33 GMT</pubDate>
            <atom:updated>2018-09-01T06:07:26.328Z</atom:updated>
            <content:encoded><![CDATA[<p>As 2017 winds to a close, I thought it might be useful to document some of my favorite books of the year. Below are a few recommendations from this past year’s reading, broken down roughly across a few categories, with further detail and recommendations beneath the first list.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/750/0*bixseB0RwYFiVJEz.jpg" /><figcaption>I mostly ‘read’ my books via Audible … but I do miss amassing a stack of completed books</figcaption></figure><h4>Books about the growth of particular technologies or figures</h4><p><a href="https://www.amazon.com/Empires-Light-Edison-Westinghouse-Electrify/dp/0375758844/ref=sr_1_1?s=books&amp;ie=UTF8&amp;qid=1514278602&amp;sr=1-1&amp;keywords=empires+of+light">Empires of Light: Edison, Tesla, Westinghouse, and the Race to Electrify the World</a>, by Jill Jonnes</p><h4>Books about the similarities and differences of humans and machines</h4><p><a href="https://www.amazon.com/Life-3-0-Being-Artificial-Intelligence/dp/1101946598/ref=sr_1_1?ie=UTF8&amp;qid=1514304382&amp;sr=8-1&amp;keywords=life+3.0">Life 3.0: Being Human in the Age of Artificial Intelligence</a>, by Max Tegmark</p><h4>Books about self-improvement</h4><p><a href="https://www.amazon.com/Option-Adversity-Building-Resilience-Finding/dp/1524732680/ref=sr_1_1?ie=UTF8&amp;qid=1514280011&amp;sr=8-1&amp;keywords=option+b">Option B: Facing Adversity, Building Resilience, and Finding Joy</a>, by Sheryl Sandberg &amp; Adam Grant</p><h4>Science-fiction books</h4><p><a href="https://www.amazon.com/Ernest-ClinesReady-Player-Hardcover-2011/dp/B0065WC2UI/ref=sr_1_3?ie=UTF8&amp;qid=1514279564&amp;sr=8-3&amp;keywords=ready+player+one">Ready Player One</a>, by Ernest Cline</p><h4>Books about recent events</h4><p><a href="https://www.amazon.com/American-Kingpin-Criminal-Mastermind-Behind/dp/1591848148/ref=sr_1_1?s=books&amp;ie=UTF8&amp;qid=1514304554&amp;sr=1-1&amp;keywords=american+kingpin">American Kingpin: The Epic Hunt for the Criminal Mastermind Behind the Silk Road</a>, by Nick Bilton</p><h4>Books about the growth of particular technologies or figures</h4><p><a href="https://www.amazon.com/Empires-Light-Edison-Westinghouse-Electrify/dp/0375758844/ref=sr_1_1?s=books&amp;ie=UTF8&amp;qid=1514278602&amp;sr=1-1&amp;keywords=empires+of+light">Empires of Light: Edison, Tesla, Westinghouse, and the Race to Electrify the World</a>, by Jill Jonnes</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/0*2zJU4gDEGjU6fYz7.jpg" /></figure><p>Following my time at MIT’s “AI and the Future of Work” conference, I wanted to learn more about the growth of other general purpose technologies — and electricity seemed as good a place to start as any (though this may also have been inspired by the forthcoming movie <a href="https://en.wikipedia.org/wiki/The_Current_War">The Current War</a>, which unfortunately has lackluster reviews).</p><p>A few interesting tidbits:</p><ul><li>Tesla forewent financial gains to keep his AC systems in production; he waived his royalties from Westinghouse when Westinghouse faced financial difficulties and otherwise may not have been able to continue. Not <em>quite</em> the same as the open-sourcing of AI libraries like TensorFlow, but still an interesting story.</li><li>Electricity became embroiled in public relations debates about its use for unseemly means, often peddled by existing business interests (e.g., competitors trying to get Westinghouse’s AC generators used for state executions — “condemning someone to the Westinghouse”)</li><li>JP Morgan likely authorized the merger of Edison General Electric and Thomson-Houston into <a href="https://en.wikipedia.org/wiki/General_Electric">General Electric</a> without Thomas Edison’s sign-off on naming … A detail of which he wasn’t too keen</li></ul><p><em>Other recommendations:</em> <a href="https://www.amazon.com/Pixar-Beyond-Unlikely-Journey-Entertainment/dp/1328745619/ref=sr_1_1?ie=UTF8&amp;qid=1514278532&amp;sr=8-1&amp;keywords=to+pixar+and+beyond">To Pixar and Beyond</a>; <a href="https://www.amazon.com/Everything-Store-Jeff-Bezos-Amazon/dp/0316219282/ref=sr_1_1?s=books&amp;ie=UTF8&amp;qid=1514278553&amp;sr=1-1&amp;keywords=the+everything+store">The Everything Store</a>; <a href="https://www.amazon.com/Elon-Musk-SpaceX-Fantastic-Future/dp/006230125X/ref=sr_1_1?s=books&amp;ie=UTF8&amp;qid=1514278571&amp;sr=1-1&amp;keywords=elon+musk">Elon Musk</a>; <a href="https://www.amazon.com/Undoing-Project-Friendship-Changed-Minds/dp/0393354776/ref=sr_1_1?s=books&amp;ie=UTF8&amp;qid=1514278587&amp;sr=1-1&amp;keywords=the+undoing+project">The Undoing Project</a></p><h4>Books about the similarities and differences of humans and machines</h4><p><a href="https://www.amazon.com/Life-3-0-Being-Artificial-Intelligence/dp/1101946598/ref=sr_1_1?ie=UTF8&amp;qid=1514304382&amp;sr=8-1&amp;keywords=life+3.0">Life 3.0: Being Human in the Age of Artificial Intelligence</a>, by Max Tegmark</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/211/1*fGdmM7LV9N0U7orM4kqyvg.png" /></figure><p>Max Tegmark, co-founder of the Elon Musk-funded Future of Life Institute, considers the consequences of a future in which both human ‘hardware’ and ‘software’ may be designed, rather than purely evolved. Tegmark also questions what a good future may look like and offers ways to increase the probability of a profoundly positive future. <em>(These changes are in contrast to what Tegmark calls today’s Life 2.0, in which ‘software’ may be designed, but ‘hardware’ is evolved: If I want to learn a new language or musical instrument, I can teach myself over time — but I can’t change my DNA, grow an additional limb, or alter my eye color.)</em></p><p>A few of the insights that have stuck with me:</p><ul><li>Human ‘software’ has evolved to solve certain problems without direct mapping to their computational complexity (e.g., humans can recognize a friend’s face instantly — a very complex task — but struggle to calculate 457 * 143).</li><li>Absent existence of a god or some other creator, there isn’t predetermined reason for humans to be at the center of the universe: We could continue our dominion on Earth, or fade into the background vs. a superior species, or cease to exist altogether.</li><li>There is wide support among the AI community for <em>certain</em> AI safety protocols and goals (e.g., not deploying autonomous killer robots), though still disagreement about the relative priority of this research (and reasons for said disagreement). A list of principles and signatories for norms in AI research can be found here:<strong><em> </em></strong><a href="https://futureoflife.org/ai-principles/">https://futureoflife.org/ai-principles/</a></li></ul><p><em>Other recommendations:</em> <a href="https://www.amazon.com/Algorithms-Live-Computer-Science-Decisions/dp/1250118360/ref=sr_1_1?ie=UTF8&amp;qid=1514304513&amp;sr=8-1&amp;keywords=algorithms+to+live+by">Algorithms to Live By</a>; <a href="https://www.amazon.com/Knowledge-Illusion-Never-Think-Alone/dp/039918435X/ref=sr_1_1?s=books&amp;ie=UTF8&amp;qid=1514304528&amp;sr=1-1&amp;keywords=knowledge+illusion">The Knowledge Illusion</a></p><h4>Books about self-improvement</h4><p><a href="https://www.amazon.com/Option-Adversity-Building-Resilience-Finding/dp/1524732680/ref=sr_1_1?ie=UTF8&amp;qid=1514280011&amp;sr=8-1&amp;keywords=option+b">Option B: Facing Adversity, Building Resilience, and Finding Joy</a>, by Sheryl Sandberg &amp; Adam Grant</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/0*WZctk0MSOOAegNfN.jpg" /></figure><p>Sheryl Sandberg encountered a horrible adversity on a global stage (the sudden passing of her husband, <a href="https://en.wikipedia.org/wiki/Dave_Goldberg">Dave Goldberg</a>, CEO of SurveyMonkey); this is her story of recovering from the grief and what others can do when they face unexpected setbacks or disappointments.</p><p><em>Other recommendations: </em><a href="https://www.amazon.com/Grit-Passion-Perseverance-Angela-Duckworth/dp/1501111108/ref=sr_1_1?s=books&amp;ie=UTF8&amp;qid=1514280257&amp;sr=1-1&amp;keywords=grit">Grit</a>; <a href="https://www.amazon.com/Principles-Life-Work-Ray-Dalio/dp/1501124021/ref=sr_1_1?s=books&amp;ie=UTF8&amp;qid=1514280271&amp;sr=1-1&amp;keywords=principles">Principles</a>; <a href="https://www.amazon.com/Happiness-Pursuit-Finding-Quest-Purpose/dp/038534886X/ref=sr_1_1?s=books&amp;ie=UTF8&amp;qid=1514280287&amp;sr=1-1&amp;keywords=happiness+of+pursuit">The Happiness of Pursuit</a></p><h4>Science-fiction books</h4><p><a href="https://www.amazon.com/Ernest-ClinesReady-Player-Hardcover-2011/dp/B0065WC2UI/ref=sr_1_3?ie=UTF8&amp;qid=1514279564&amp;sr=8-3&amp;keywords=ready+player+one">Ready Player One</a>, by Ernest Cline</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/131/0*c19ioo_3y7IquVOc.jpg" /></figure><p>Ready Player One tells the story of a teen living in and hunting through a wholly-immersive Virtual Reality world (e.g., education, media, combat, currencies, etc). He’s on a quest to find <a href="https://en.wikipedia.org/wiki/Easter_egg_(media)">Easter Eggs </a>left behind by the system’s architect and win a life-changingly large prize — and in telling the story, the book helps bring to life what may eventually be the future of VR.</p><p>A few themes I appreciated:</p><ul><li>What are the roles of nations, guilds, and other affinity groups in an increasingly-digital world?</li><li>What are possible impacts of virtual reality across income levels, education, race, etc?</li><li>What are applications of VR for leisure vs. productive use (e.g., new forms of games and media, vs. more-effective techniques for learning)</li></ul><p><em>Other recommendations: </em><a href="https://www.amazon.com/I-Robot-Isaac-Asimov/dp/055338256X/ref=sr_1_1?s=books&amp;ie=UTF8&amp;qid=1514279913&amp;sr=1-1&amp;keywords=i+robot">I, Robot</a>; <a href="https://www.amazon.com/Minority-Report-Collected-Stories-Gollancz/dp/0575075201/ref=sr_1_1?s=books&amp;ie=UTF8&amp;qid=1514279932&amp;sr=1-1&amp;keywords=volume+iv+minority+report">Minority Report</a></p><h4>Books about recent events</h4><p><a href="https://www.amazon.com/American-Kingpin-Criminal-Mastermind-Behind/dp/1591848148/ref=sr_1_1?s=books&amp;ie=UTF8&amp;qid=1514304554&amp;sr=1-1&amp;keywords=american+kingpin">American Kingpin: The Epic Hunt for the Criminal Mastermind Behind the Silk Road</a>, by Nick Bilton</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/155/1*WltuCDuPuAiUACsjt9ePMw.png" /></figure><p>In a year of increased prevalence of Bitcoin and cryptocurrency generally, I found myself fascinated by the story of Ross Ulbricht, aka the Dread Pirate Roberts and founder of the Silk Road (an illicit marketplace that, in the early days of Bitcoin, was a possible cryptocurrency use-case).</p><p>On one hand, Ross is incredibly passionate about his cause, and it’s fairly inspiring how committed he was and how he found the determination to grow his enterprise. On the other (more significant) hand, Ross is also a drug kingpin who routinely committed felonies and attempted to have people killed to fulfill his anarcho-capitalist vision… Not cool.</p><p>Bilton provides a thorough, albeit possibly dramatized, account of Ross’s fall into criminality and the FBI’s attempts to bring him down, with many ‘first-hand’ accounts taken directly from Ross’s diaries.</p><p><em>Other recommendations: </em><a href="https://www.amazon.com/Black-Edge-Inside-Information-Wanted/dp/0812995805/ref=sr_1_1?ie=UTF8&amp;qid=1514304771&amp;sr=8-1&amp;keywords=black+edge">Black Edge</a>; <a href="https://www.amazon.com/What-Happened-Hillary-Rodham-Clinton/dp/1501175564/ref=sr_1_1?s=books&amp;ie=UTF8&amp;qid=1514304783&amp;sr=1-1&amp;keywords=what+happened">What Happened</a></p><h4>What’s next?</h4><p>Here are a few of the books next up on my queue — I’m excited to see what comes of them:</p><p><a href="https://www.amazon.com/Most-Powerful-Idea-World-Invention/dp/0226726347/ref=sr_1_1?s=books&amp;ie=UTF8&amp;qid=1514304809&amp;sr=1-1&amp;keywords=most+powerful+idea+in+the+world">The Most Powerful Idea in the World: A Story of Steam, Industry, and Invention</a>, by William Rosen</p><p><a href="https://www.amazon.com/Titan-Life-John-Rockefeller-Sr/dp/1400077303/ref=sr_1_1?ie=UTF8&amp;qid=1514304913&amp;sr=8-1&amp;keywords=titan+ron+chernow">Titan: The Life of John D. Rockefeller, Sr.</a>, by Ron Chernow</p><p><a href="https://www.amazon.com/Mans-Search-Meaning-Viktor-Frankl/dp/080701429X/ref=sr_1_1?s=books&amp;ie=UTF8&amp;qid=1514305375&amp;sr=1-1&amp;keywords=man%27s+search+for+meaning">Man’s Search for Meaning</a>, by Viktor Frankl</p><h4>Steven Adler is a former strategy consultant focused across AI, technology, and ethics.</h4><h4>If you want to follow along with Steven’s projects and writings, make sure to follow <a href="https://medium.com/@sjgadler/">this Medium account</a>. Learn more on <a href="http://www.linkedin.com/in/sjgadler">LinkedIn</a>.</h4><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b3163f5a24ef" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[To perform like Magnus… relax like Magnus?]]></title>
            <link>https://medium.com/@sjgadler/to-perform-like-magnus-relax-like-magnus-1015d9e326ae?source=rss-c04544a536f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/1015d9e326ae</guid>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[peak-performance]]></category>
            <category><![CDATA[chess]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[alphazero]]></category>
            <dc:creator><![CDATA[Steven Adler]]></dc:creator>
            <pubDate>Thu, 14 Dec 2017 23:27:01 GMT</pubDate>
            <atom:updated>2018-09-01T06:08:33.769Z</atom:updated>
            <content:encoded><![CDATA[<h4>Magnus Carlsen is the world’s highest-rated chess player, and he doesn’t spend all day playing chess. What role does that play in his success?</h4><p>Magnus Carlsen, 27, is the highest-rated (human) chess player of all-time and has been world chess champion since 2013. His first championship victory is captured in the very-enjoyable <a href="https://www.netflix.com/search?q=magnus&amp;jbv=80118004&amp;jbp=0&amp;jbr=0"><em>Magnus</em></a>, now available on Netflix everywhere.</p><p>Magnus, as one might expect, puts in hours of deliberate practice in trying to become the best at chess. He is quizzed by his head coach about historical positions; he analyzes his own games; he reads chess news to understand developments in other players.</p><p>But something else stands out about Magnus in the documentary: He spends a good amount of time doing things that aren’t quite chess-related.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/585/1*pxiqM5CARANhuIa8cbXQRQ.png" /><figcaption>Soon-to-be-World Champion Magnus Carlsen at his training camp in South Norway. Magnus asks “Why do I do that?” about a missed volleyball hit — but he could also be asking a more meta question.</figcaption></figure><p>Repeatedly in the film, Magnus is shown playing volleyball, ping pong, and swimming, even in the run-up to critical matches. At times when many others would be inclined to hunker down and cram, Magnus seems to find continued ‘distractions’ in other domains.</p><p>Now, I will not pretend to have a comprehensive accounting of Magnus’s practice hours; perhaps this is only a trivial subsection of Magnus’s week. Additionally, Magnus is certainly not a slacker; he is shown studying chess when surrounded by family, for instance, when it would be easy to study just a bit less.</p><p>But Magnus’s overall preparation style — and the varied activities — does not escape notice of his head coach, particularly compared with the documentary’s portrayal of then-World Champion Vishy Anand: <em>“You know, [Magnus] may not always be the most serious guy in training. But in his head, he has stuff going on. It’s a different kind of approach than maybe other kinds of schools.”</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/584/1*cpyUp1cQAPDPEIrK3ITbOA.png" /><figcaption>Magnus plays ping pong with his head coach in lead-up to the chess world championship.</figcaption></figure><p>Magnus without question puts in legwork for chess, but he also finds room for other activities. Without getting overly-rigorous (and at risk of leaning too pop-psychological), I think there are a few questions to consider for one’s own performance after observing how Magnus spends portions of his preparation time:</p><p><strong><em>Would your performance benefit from time for more-complex realizations to form? </em></strong>As Magnus’s coach notes, his time playing ping pong is not dead time from training, but rather time to gradually work over things without full-force thinking. Plenty of activities involve insight problems that aren’t necessarily best solved by thinking harder or longer; many innovations arise from taking a concept in one domain and applying it to another. For Magnus, sports provide this ‘idle work’ time, just as many people find that taking walks encourages their best thinking. As an additional upside, sports help to keep the body active and healthy.</p><p><strong><em>Would your performance benefit from time to mentally recharge?</em></strong> Though this is not explicitly discussed as a factor in the documentary, I suspect a large reason for Magnus’s breaks is that studying chess is exhausting. On a biological level, Stanford neuroscientist Robert Sapolsky contends that grand masters can burn <a href="https://news.stanford.edu/news/2009/june17/classday-061709.html">thousands of calories per day</a> in the course of a chess tournament (though it need not be several thousands to be significant). Even without the calorie consideration, <a href="https://www.inc.com/melanie-curtin/in-an-8-hour-day-the-average-worker-is-productive-for-this-many-hours.html">many workers are only capable of peak productivity for stints of ~3 hours</a> — though this isn’t ironclad, and someone like Magnus very well may retain focus for longer. At some point, however, everyone will face diminishing returns — and when faced with diminishing returns, why not play ping pong? (Or sleep.)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/439/1*BopjQBcZBNV0t0o4ucAEbQ.png" /><figcaption>This red-lined engine is at a balmy 37C, just like the human body. Coincidence? I think not.</figcaption></figure><p><strong><em>Would your performance be more sustainable with allowances for ‘sub-optimal’ activities?</em></strong> This question is also not discussed in the documentary, but I do wonder if Magnus’s working in more time for sports and friendships will allow him to sustain peak performance for longer. On one hand, time spent in these ways might trade off with time spent training on chess and could in theory lead to worse short-run outcomes (though as discussed above, perhaps not). On the other hand, Magnus is a human being, and if these interests keep him happy, healthy, and motivated to keep pursuing his goals, his hobbies may well end up being instrumental to his success even if locally suboptimal. Put more simply, this is a question of avoiding burnout: If taking on certain stresses and time pressures will cause you to redline, perhaps they aren’t the right path for your long-term career goals.</p><p>I continue to be inspired by Magnus’s journey to the top of the chess world, as well as by the feats of winetasters in <a href="https://www.netflix.com/watch/70273636?trackId=13752289&amp;tctx=0%2C1%2C"><em>Somm</em></a><em> </em>and friends of mine from the national debate community. (Not to mention my friend Max, who recently traveled to Germany to <a href="https://www.wsj.com/articles/chess-novice-challenged-magnus-carlsen-1510866214">play Magnus in a game of chess</a>… Max lost but clearly hasn’t been playing enough ping pong.)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/475/1*UauFfWF0ZETEuWRwBgjKZQ.png" /><figcaption>Wine: It’s serious business. (Screenshot from <em>Somm)</em></figcaption></figure><p>There’s something fascinating about peak performance and the focus it inspires in people — and particularly when people seem to have found a balance between that focus and continuing to achieve their goals.</p><p>Of course, in the future, computers might not face these tradeoffs that can constrain human performance today. (The <a href="https://en.wikipedia.org/wiki/Stockfish_(chess)">Stockfish</a> and <a href="https://www.chess.com/news/view/google-s-alphazero-destroys-stockfish-in-100-game-match">AlphaZero</a> chess engines, for instance, don’t step away from chess to have dinner with their loved ones, at least as far as we know.)</p><p>It’s worth considering, then, how demands on human performance may evolve over time, particularly as computers expand deeper into the realm of human activities and things previously considered art.</p><p>At chess, we can say with confidence that Magnus aka “The Mozart of Chess” and the highest-rated human ever &lt; <a href="https://en.wikipedia.org/wiki/Stockfish_(chess)">Stockfish</a> (the go-to chess engine for human preparation) &lt; <a href="https://www.chess.com/news/view/google-s-alphazero-destroys-stockfish-in-100-game-match">AlphaZero</a> (a self-playing reinforcement learning agent that ran for less than a day [on incredibly high-end hardware]).</p><p>Unfortunately for Magnus’s computer-beating prospects, no amount of sleep or time training in other domains is likely to reverse that — but he might pose a model for us to increase our own productivity in the interim.</p><h4>Steven Adler is a former strategy consultant focused across AI, technology, and ethics.</h4><h4>If you want to follow along with Steven’s projects and writings, make sure to follow <a href="https://medium.com/@sjgadler/">this Medium account</a>. Learn more on <a href="http://www.linkedin.com/in/sjgadler">LinkedIn</a>.</h4><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1015d9e326ae" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[My AI is live — and it’s about as good as I am]]></title>
            <link>https://medium.com/@sjgadler/my-ai-is-live-and-its-about-as-good-as-i-am-81b8bfdd4ee3?source=rss-c04544a536f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/81b8bfdd4ee3</guid>
            <category><![CDATA[strategy-games]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Steven Adler]]></dc:creator>
            <pubDate>Wed, 06 Dec 2017 06:58:40 GMT</pubDate>
            <atom:updated>2018-09-01T06:08:56.738Z</atom:updated>
            <content:encoded><![CDATA[<h4>This post is part of <a href="https://medium.com/@sjgadler/building-a-strategy-game-and-the-ai-to-master-it-58ffbf1ac2f0">ChurrPurr.ai</a>, a challenge to design an online strategy game and the AI to master it. Play the latest version of the game <a href="http://churrpurrdemo.bitballoon.com/">here</a>.</h4><p>I’m proud to say that my AI is now live, and it plays a pretty capable game of Churr-Purr — if I do say so myself.</p><p>Its gameplay is far from perfect, and a person who understands how it plays could likely exploit its tendencies over time. That said, it’s also a far cry from randomness in what began as a relatively complex game.</p><p>I should caveat, however: This AI is not the end-state system I’ve begun <a href="https://medium.com/@sjgadler/going-backwards-to-go-forwards-6fac7635a19e">building with deep reinforcement learning</a>. Instead, I’ve built a baseline AI that later iterations can play against as a yardstick of their progress.</p><p>Allow me to explain …</p><p>For my yardstick AI, I’ve taken the approach of defining relevant rules-of-thumbs for our AI to follow.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/300/1*3bhkaS0EjpF_uQDWdZE9oQ.png" /><figcaption>These people seem pretty excited for the launch of Churr-Purr’s AI.</figcaption></figure><p>For instance, when the AI detects immediate opportunity to create 3-in-a-row, it should do so. Similarly, it should prevent its opponent’s 3-in-a-rows, because those have power to undo whatever move our AI may play instead.</p><p>Accordingly, the AI plays about as well as I do, as its gameplay is based around the rules of thumb I too follow when deciding what move to make.</p><p>The AI still ‘knows’ about the state of the game, what actions are permissible, and so on, but its behaviors are driven by a set of <em>features</em> that I’ve defined, rather than features it’s discovered on its own.</p><p>What does that mean? Let’s consider the two systems:</p><h4>AI feature definition</h4><p>I can teach my AI about how to test the game for certain features and tailor its behavior accordingly — for instance, how to detect duos threatening to become trios, and then block them. This approach comes with some pros and cons:</p><ul><li>Pros: The AI is fairly interpretable; by walking through the code, you can trace what decisions happen and why. The AI can also start off with a relatively robust amount of ‘knowledge’, rather than needing to accrue it through experience.</li><li>Cons: The AI is limited to the quality of the programmer — it won’t learn over time or discover novel methods on its own. If the programmer’s rules of thumb are misguided, the AI will be stuck at a bad baseline without means of improving. Additionally, directing important features for an AI quickly becomes unwieldy — and it isn’t very fun to write a truly comprehensive feature set.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/755/1*tAcM_IBoIqAa-FqPIua85g.png" /><figcaption>I wouldn’t have much faith in an AI that came from my attempts to define features of particular dog breeds … Might be best to leave it to the experts (or the AIs themselves).</figcaption></figure><h4>AI feature discovery</h4><p>An AI without my direct input can also ‘learn’ about duos threatening to become trios, but in a different manner.</p><p>For instance, a reinforcement learning agent may take 1,000s of iterations to discover what this pattern looks like, how to counter it, that not countering it leads to loss of points, and so on. (This is all information I could endow my feature-defined AI with from the outset.)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/379/1*ykjxAF54x0raRX_jUlH3pA.png" /><figcaption>Some examples of duos that the AI may eventually discover — or that a programmer can help an AI to detect upfront.</figcaption></figure><p>For AI using neural networks (NN), this learning happens through a series of value updates to a group of mathematical weights. At the end of this process, there most likely won’t be a certain weight that corresponds to ‘yes the NN detects a duo’ vs. ‘no the NN does not’ — but in aggregate, the decisions will be informed by this detection, albeit in hard-to-pin-down ways (this is what researchers are alluding to when they suggest neural networks are black boxes). <em>(For a good introduction to neural networks, albeit with some technical complexity, check out </em><a href="https://ujjwalkarn.me/2016/08/09/quick-intro-neural-networks/"><em>this article</em></a><em>.)</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/791/1*590A1_2nItX49wqZKOlaFw.png" /><figcaption>This is a fairly standard image of a small neural network. To simplify things, imagine each circle in the input layer is a number; each of these numbers could be passed through several mathematical transformations such that they in aggregate produce an output. For instance, each number on the left might be multiplied by 3 (to get to HL1) and then squared (to get to HL2). An output rule could, for instance, sum all the numbers at HL2 and produce a certain decision if the sum is odd vs. if it is even. Generally, however, transformations and output rules will be a bit more opaque and complex.</figcaption></figure><p>The feature-discovery approach similarly has pros and cons:</p><ul><li>Pros: The AI can discover novel associations unknown to humans and reach higher ceilings of learning; it can also buck convention, such as <a href="https://www.wired.com/2016/03/two-moves-alphago-lee-sedol-redefined-future/">AlphaGo’s notorious “move 37”</a> that puzzled commentators and expert Go players. Additionally, the AI can learn complex features that are difficult to articulate directly.</li><li>Cons: The AI requires time to learn; in a poorly-designed learning environment, it may never fully learn. Additionally, when learning is complete, developers will not necessarily be able to explain reasons for the AI’s actions.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pXx4pbOrtM0EWLKrj-dbqA.png" /><figcaption>Google tells me that this is an image of AlphaGo’s infamous match following move 37. I’d love to explain to you the brilliance behind AlphaGo’s deviation, but if I could do that, I might have a world championship to go win.</figcaption></figure><p>So, which method is better? And does my take on the first method, in which rules of thumb are defined, even count as AI? Good questions.</p><p>To the first question — it largely depends what your objectives are.</p><p>For instance, if my AI needs to operate under tight time constraints, the feature-defined method will get to competency far more quickly than the feature-discovery method.</p><p>Additionally, if you care about knocking out a large percentage of a problem but not necessarily becoming state-of-the-art, feature definition can go a long ways: Spam filters that treat certain notorious words with suspicion aren’t perfect, and there’s certainly risk of false positives — but they also do a pretty good job of rooting out spam. For my baseline AI, which isn’t meant to become Churr-Purr world champion, I think its rules of thumb are enough.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4yPMF2xbOAK0OQZVJdMqyg.png" /><figcaption>Another challenge of defined feature AIs: They can be prone to game-playing by those who want to avoid certain classifications …</figcaption></figure><p>To the second question, of whether this is AI — I think the answer is a bit more complicated, but ultimately, I lean yes.</p><p>I see a few possible objections to whether this method is AI:</p><p><strong><em>Its moves are somewhat pre-determined</em></strong><em> </em>— True, the AI certainly isn’t as freeform as one starting with no conception of the game, but it’s also a large improvement over “<a href="https://en.wikipedia.org/wiki/Symbolic_artificial_intelligence">good old-fashioned AI</a>”. Instead of spelling out state-action pairs in painstaking detail, we can greatly reduce the directions needed by training the AI to look for themes.</p><p><strong><em>It doesn’t really ‘know’ about the game / hasn’t ‘mastered’ it</em></strong><em> </em>— This is an interesting objection but may also be anthropomorphizing the AI a bit. I think a big component of AI is whether the software achieves its intended goals in a way that requires knowledge.</p><p>For instance, software that automates 98% of routine accounting tasks might not be said to have ‘mastered’ the art of accounting, but it’s also pretty clearly a software capable of engaging in accounting. Meanwhile, if I were to bet on who could beat 100 people in simultaneous play, I think my AI is far more capable than I am; it’s less prone to misapplying rules, forgetting the state of the game, and many other mistakes I would surely make.</p><p><strong><em>It doesn’t figure anything out on its own / it doesn’t learn over time</em></strong><em> — </em>I think this is the strongest of the objections, but ultimately, I’m not convinced these are essential components of an AI. Instead, I believe these arguments somewhat fault the AI for not being conscious.</p><p>Consider, for instance, that humans also get knowledge from external sources (books, other people). Certainly most people update their beliefs in light of new feedback, but would we stop considering someone as an intelligent agent if they instead stuck with the same habits and routines? I doubt it. This AI would not be generalizable to other problems or particularly ‘strong’, but it does have some domain expertise. It cannot ‘know’ in the sense of a human, but when judged by its actions, the AI does seem to behave ‘intelligently’.</p><p>Underneath many of these objections is a sort of <a href="https://en.wikipedia.org/wiki/Chinese_room">Chinese room problem</a> in which one gets into very philosophical discussions of what it means to know something, and whether action without understanding truly counts as knowledge.</p><p>While these considerations are fun — and also potentially impactful, as they could be one day part of defining AI’s status in society — they don’t need to be resolved to enjoy the updated functionality of Churr-Purr’s AI.</p><p>I encourage you to give some thought to how you define intelligence and knowledge in your own life, and I also encourage you to <a href="http://churrpurrdemo.bitballoon.com/">check out the AI firsthand</a>. Let me know what you think of its gameplay; I hear it’s got a few tricks up its sleeve …</p><p><em>Read the </em><a href="https://medium.com/@sjgadler/sometimes-the-devil-is-in-the-details-5c0e4e76684c"><em>previous post</em></a><em>.</em></p><h4>Steven Adler is a former strategy consultant focused across AI, technology, and ethics.</h4><h4>If you want to follow along with Steven’s projects and writings, make sure to follow <a href="https://medium.com/@sjgadler/">this Medium account</a>. Learn more on <a href="http://www.linkedin.com/in/sjgadler">LinkedIn</a>.</h4><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=81b8bfdd4ee3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Sometimes the devil is in the details …]]></title>
            <link>https://medium.com/@sjgadler/sometimes-the-devil-is-in-the-details-5c0e4e76684c?source=rss-c04544a536f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/5c0e4e76684c</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[strategy-games]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Steven Adler]]></dc:creator>
            <pubDate>Wed, 08 Nov 2017 04:23:11 GMT</pubDate>
            <atom:updated>2018-09-01T06:09:16.045Z</atom:updated>
            <content:encoded><![CDATA[<h4>This post is part of <a href="https://medium.com/@sjgadler/building-a-strategy-game-and-the-ai-to-master-it-58ffbf1ac2f0">ChurrPurr.ai</a>, a challenge to design an online strategy game and the AI to master it. Play the latest version of the game <a href="http://churrpurrdemo.bitballoon.com/">here</a>.</h4><p>I suppose this post should really be called “the value of line-by-line debugging”, but sometimes it’s more fun to give a sensational title.</p><p>Amid tonight’s confluence of Election Day and a strong desire to watch Shark Tank (don’t hate — <a href="https://medium.com/u/29fd628f3aaa">Mark Cuban</a>’s a fun personality), I nearly didn’t put in work on the AI front. I’m glad that I ultimately did, because while I didn’t make monumental progress, I <em>did</em> debug several pieces that were preventing my reinforcement techniques from running properly.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/619/1*8rWyedKoJQVvQDpdDw09Yg.png" /><figcaption>Mark Cuban is just your normal billionaire wiping down an NBA Championship trophy with a t-shirt — go Mavs</figcaption></figure><p>It all started when I tried rewriting one of the functions from <a href="https://medium.com/@sjgadler/going-backwards-to-go-forwards-6fac7635a19e">REINFORCEjs</a> to take different parameters than expected. Nominally my program didn’t seem to be putting up a fight, but the rewrite also wasn’t quite working … What could be wrong?</p><p>I tried adding a few alerts to debug different values, but none of them would come through when I ran the program. I also tried intentionally calling the function in different ways, only one of which could be right and the other should presumably throw an error, but no dice there either.</p><p>Clearly something was wrong with the interaction across different modules of my code, and I wasn’t quite sure what — so to confirm that theory, I pasted the code into one file, rearranged the pieces in the right order, and tried running it … Voila, errors — but the expected errors.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/312/1*by_MrodwQ694wsP-2ftq4g.png" /><figcaption>Never have I been so happy to see error notifications — though, note to self, I should tag these self-created errors for quicker removal next time … Source: <a href="https://xkcd.com/1739/">https://xkcd.com/1739/</a></figcaption></figure><p>I’ll spare the line-by-line on other changes, but as I traced my code line-by-line through why it wasn’t having the expected outcomes, I noticed: 1) the test-case I was using was (due to a subtle typo) already game over [oops]; 2) the AI’s score-checking function wasn’t running properly to deduce this was game over [oops]; and 3) the AI’s tendency to evaluate invalid moves was having a larger drag on performance than I’d thought.</p><p>Going systematically through these portions was instructive, and I now have a better understanding of the underlying library with which I’m working. The reinforcement learning isn’t quite there yet, but the computer is back to considering only valid moves, and evaluating them properly. That’s certainly progress.</p><p>Now to delve more deeply into the learning parameters and how the AI builds up its memory over time …</p><p><em>Read the </em><a href="https://medium.com/@sjgadler/going-backwards-to-go-forwards-6fac7635a19e"><em>previous post</em></a><em>. Read the </em><a href="https://medium.com/@sjgadler/my-ai-is-live-and-its-about-as-good-as-i-am-81b8bfdd4ee3"><em>next post</em></a><em>.</em></p><h4>Steven Adler is a former strategy consultant focused across AI, technology, and ethics.</h4><h4>If you want to follow along with Steven’s projects and writings, make sure to follow <a href="https://medium.com/@sjgadler/">this Medium account</a>. Learn more on <a href="http://www.linkedin.com/in/sjgadler">LinkedIn</a>.</h4><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5c0e4e76684c" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>