<?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 Michael Dempsey on Medium]]></title>
        <description><![CDATA[Stories by Michael Dempsey on Medium]]></description>
        <link>https://medium.com/@mhdempsey?source=rss-6e6d7bd0dfb6------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*_Y2nvJq5ztUiVY9jhw6O8g.jpeg</url>
            <title>Stories by Michael Dempsey on Medium</title>
            <link>https://medium.com/@mhdempsey?source=rss-6e6d7bd0dfb6------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 09:09:02 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@mhdempsey/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[Robotics, Food, and People — Announcing our investment in Ono Food Co.]]></title>
            <link>https://medium.com/@mhdempsey/robotics-food-and-people-announcing-our-investment-in-ono-food-co-faf17e57a9a8?source=rss-6e6d7bd0dfb6------2</link>
            <guid isPermaLink="false">https://medium.com/p/faf17e57a9a8</guid>
            <category><![CDATA[food]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[venture-capital]]></category>
            <category><![CDATA[startup]]></category>
            <dc:creator><![CDATA[Michael Dempsey]]></dc:creator>
            <pubDate>Thu, 15 Aug 2019 18:48:49 GMT</pubDate>
            <atom:updated>2019-08-15T18:48:49.187Z</atom:updated>
            <content:encoded><![CDATA[<h3>Robotics, Food, and People — Announcing our investment in Ono Food Co.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KP_4OYQ15GSPmW6yAIc7Bw.png" /></figure><blockquote>“If anything is good for pounding humility into you permanently, it’s the restaurant business.” — Anthony Bourdain</blockquote><p>I’ve been incredibly fortunate to have grown up in the restaurant business. My dad worked his way up from a busboy in his 20s to an operator focused on scaling full-service and QSR concepts across multiple vertically-integrated restaurant groups. My mom has handled the intricacies of licensing and permitting for large-scale restaurant groups to sole proprietors for the past decade. My sister has been working in restaurants for the past 15 years, going from line cook to chef and operations director for full-service restaurants to large hospitality organizations. They all say the same thing: Food is a nightmare to scale, hard work, and an even harder business. <strong>And at the core of all of these difficulties are people.</strong></p><p>In early 2018 I wrote a post titled <a href="https://medium.com/@mhdempsey/how-to-win-with-robots-in-restaurants-9c984f2ee3d1">How To Win With Robots in Restaurants</a> that detailed our thesis on investing in food robotics. I had been thinking about how to scale restaurant groups and meaningfully change the restaurant business across multiple vectors. As many others, I believed robotics was an incredibly compelling way to achieve this. As the component costs have fallen and automation has become a more attainable goal with things as complex as food, I’ve grown increasingly bullish on what this would mean for the industry. I’ve also continued to be bearish on the idea of robotics companies selling to restaurants as well as bringing pure “tech people” to the idea of drastically “disrupting” a business as complex and difficult as food service.</p><p>Our thesis was broken down as four bullet points in that post surrounding owning and operating the entire stack, using scalable hardware, a team with food experience, and a focus on maintaining some semblance of human UX. This framework was the best way in which I believed it would be possible to capture venture-scale value and growth. My guiding principle though, burned into my brain from a family of food service professionals, was that regardless of robotics, ultimately restaurants come down to people.</p><p>People means building a brand that stands for something to both the employees and consumers of the product. In 2019 this means things like high-quality, sometimes local ingredients, an understanding of the story behind the food, the overall experience feeling special, and even how the food impacts your body and health.</p><p>People also means labor and the complicated economics of properly staffing a food service organization to meet the high bar that consumers want with respect to all of the things I mention above. Both macroeconomic factors surrounding wages and rent in urban environments, as well as concept-specific economics have pushed restaurant groups from all stages (fast food, QSR, full-service) to rethink their businesses and distribution mechanisms while also risking alienating a large portion of consumers. These issues have put an upper-bound on the capital efficiency and scalability of high-quality, affordable, and incredible food experiences.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jOcbMc2iQMmr0eHhMJ66rw.jpeg" /><figcaption>Danny and Stephen early on in their build process</figcaption></figure><p>Shortly after I published my post, Stephen Klein emailed me about a new company he was starting with his friend Daniel Fukuba. After chatting with Stephen and Danny, I was sold on the people and their vision.</p><p>Stephen and Danny had a perfect balance as a founding team of what the future of food service looked like to me. Stephen’s deep experience in operating and scaling tangential companies like Instacart to related companies in CafeX, as well as his deep passion and knowledge of the food service industry blew me away. Danny’s past experience as an engineering manager at L2F, having worked on automation for companies from SpaceX to Creator to Zume Pizza similarly displayed some of the deepest understanding of the necessary (and over-engineered) complexities that exist within food robotics, as well as the technical approaches needed to create truly scalable and fault-tolerant automation systems.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WP4uYUbxlzmp3UcwVQZ5Bg.jpeg" /><figcaption>An early sketch of the Ono Blend food truck</figcaption></figure><p>Throughout our conversations, Stephen and Danny surgically broke down food service economics, paired with a vision of how to create a massively scalable, and economically dominant food service brand, starting with modular, mobile restaurants. After spending more time with them, pressure testing their views on food service, technology, and the marriage between the two, Compound became the first investors in Ono Food Co.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*BT3s7k2oj4aBSnTm6NSVLA.jpeg" /></figure><p>Over the past year the team at Ono has consistently hit countless technical and operational milestones at an order of magnitude greater capital efficiency than any other food robotics company in history. Recently, they’ve reached their most important milestone, which will culminate in the launch of their first mobile restaurant serving Ono Blends in Los Angeles. Unlike other competitors that charge 2–3x the price for similar products, some healthy, some loaded with sugar, Ono is making affordable healthy blends (similar to smoothies) available to Los Angeles. And of course, Ono is Hawaiian for “delicious”, so as a heavy beta tester (often via my own request, not for any actual purpose to the company) you know they are delicious.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*119p67ZGqnc5A8zv77vJZA.png" /></figure><p>We’re incredibly excited about the future of Ono and their goal to build the next massive food brand across multiple modalities, concepts and geographies, enabled by automation. While we can’t talk about all the future plans, Ono Blends is just the start.</p><p>To read more about Ono Food Co. <a href="https://onofood.co/">go here</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=faf17e57a9a8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Technology is making us crave the similar, and the unique]]></title>
            <link>https://medium.com/@mhdempsey/technology-is-making-us-crave-the-similar-and-the-unique-b3330a98c19e?source=rss-6e6d7bd0dfb6------2</link>
            <guid isPermaLink="false">https://medium.com/p/b3330a98c19e</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[instagram]]></category>
            <category><![CDATA[identity]]></category>
            <category><![CDATA[social-media]]></category>
            <category><![CDATA[influencers]]></category>
            <dc:creator><![CDATA[Michael Dempsey]]></dc:creator>
            <pubDate>Tue, 26 Feb 2019 20:06:29 GMT</pubDate>
            <atom:updated>2019-02-26T22:17:40.602Z</atom:updated>
            <content:encoded><![CDATA[<h4>What happens when we can facetune to something “better”, generate to something familiar, and filter to something different.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/923/0*Hweq7pfwKXqsEk4v.jpeg" /><figcaption>Sources: <a href="https://engadget.com/2019/02/17/the-big-picture-selfie-harm-rankin-visual-diet/?utm_campaign=the_download.unpaid.engagement&amp;utm_source=hs_email&amp;utm_medium=email&amp;utm_content=69998497&amp;_hsenc=p2ANqtz--xQxwP7bapYVapsttlgMztFo0PZCAbgJVQvT59x0jtGHziQpmPlA7YCci96AuPjwpFGacp5xx65lbkPedmJifQfthi4w&amp;_hsmi=69998497">Engadget</a>, <a href="https://www.dazeddigital.com/beauty/head/article/43138/1/beauty-gan-kylie-jenner">Dazed</a>, <a href="https://www.instagram.com/p/BtyCLRAi8oK/">Instagram</a>, <a href="http://liusi-group.com/pdf/BeautyGAN-camera-ready_2.pdf">BeautyGAN</a></figcaption></figure><p>Over the past two weeks I’ve read multiple pieces that are on opposing sides of what we’re craving as consumers and in our digital self-expression. They all speak to a duality of how we are both starved for individuality as well as driven towards homogeny. These points are communicated to us through some interesting trends in technology, social media, and pop culture.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*JMuqY14jtnJvhPa-.jpg" /><figcaption>Beauty_Gan of Kylie Jenner</figcaption></figure><p><a href="https://dazeddigital.com/beauty/head/article/43138/1/beauty-gan-kylie-jenner">Beauty_GAN</a> (not to be confused with <a href="http://liusi-group.com/pdf/BeautyGAN-camera-ready_2.pdf">BeautyGAN</a>) is a sparsely documented implementation of a GAN that utilizes Instagram makeup trends and then generates new styles, which Dazed put on Kylie Jenner’s face. The resulting imagery feels very cherry picked (not a rarity based off of my experience with GANs) but the key point of the article is that these dataset sizes and inputs are continually built with human in the loop biases. Put more artistically in the article:</p><blockquote><em>“…Beauty_GAN is like a mirror of popular culture, but the reflection staring back at you might not be what you expected. We teach a machine to see us and what it shows us back is not always what we see ourselves.”</em></blockquote><p>Despite what the GAN mirrors back in terms of how “dystopian” something looks, there are two other related pieces that speak to this homogenization of taking the internet’s makeup kingmaker (Kylie) and re-painting the internet’s makeup onto her. Or as Dazed wrote:</p><blockquote><em>“One could argue that, of all the beauty imagery we see on Instagram today, Kylie Jenner’s face, her aesthetic, holds the most influence. Every time someone copies her contour or lip liner there’s a further proliferation that happens. She influences what we think of as beautiful, what exists on Instagram. The Beauty_GAN project sees this inputted into a machine, and then lets the machine take over; the machine creates what it thinks is beauty imagery, and then paints it back onto Kylie’s face. And so, the feedback loop closes.”</em></blockquote><p>And this feedback loop has proliferated into other forms of celebrities as Telegraph highlighted in <a href="https://www.telegraph.co.uk/films/0/alitalil-miquela-terrifying-rise-dead-eyed-digital-cyberbabes/">this article</a>.</p><blockquote><em>“Lil Miquela, after all, is the ultimate embodiment of homogenised, Instagram-friendly beauty. An ambiguous mix of different ethnicities, with on-trend freckles and a body that can be shaped and moulded depending on the body parts required, she can be everything that consumers desire at any given time. “</em></blockquote><p>But the key here isn’t what makes a character like Miquela or <a href="https://www.instagram.com/imma.gram/">Imma.gram</a> compelling. There’s time decaying first order interest points of “is this a robot or a human?!” and the general intrigue of a synthetic being, but then there’s the natural, more commonplace feeling amongst many influencers of “they are kind of like me, but <em>better</em>.”</p><p>And this close similarity is what I believe is akin to the dopamine rushes that gamification experts have hit on for a long time of being partially satiated, but not entirely fulfilled. This phenomenon has partially been described as <a href="https://www.engadget.com/2019/02/17/the-big-picture-selfie-harm-rankin-visual-diet/?utm_campaign=the_download.unpaid.engagement&amp;utm_source=hs_email&amp;utm_medium=email&amp;utm_content=69998497&amp;_hsenc=p2ANqtz--xQxwP7bapYVapsttlgMztFo0PZCAbgJVQvT59x0jtGHziQpmPlA7YCci96AuPjwpFGacp5xx65lbkPedmJifQfthi4w&amp;_hsmi=69998497">Selfie Harm</a>. It keeps us wanting more, liking more, swiping more, for something that we know we likely can’t obtain. <strong>But what happens when we can?</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*NPFD85f_T7tDMwQV.jpeg" /><figcaption>Source: <a href="https://www.engadget.com/2019/02/17/the-big-picture-selfie-harm-rankin-visual-diet/?utm_campaign=the_download.unpaid.engagement&amp;utm_source=hs_email&amp;utm_medium=email&amp;utm_content=69998497&amp;_hsenc=p2ANqtz--xQxwP7bapYVapsttlgMztFo0PZCAbgJVQvT59x0jtGHziQpmPlA7YCci96AuPjwpFGacp5xx65lbkPedmJifQfthi4w&amp;_hsmi=69998497">Engadget</a></figcaption></figure><p>When we’re given tools that allow us to have those “on-trend freckles” of Miquela, or the contour and sizing of Kylie’s lips, or the dyed hair of <a href="http://instagram.com/imma.gram">Imma.gram</a>, then what do we crave? Perhaps difference.</p><p>This is what I believe we’re seeing in pockets of the internet today. We’re seeing massive share numbers generated by very differentiated and unique AR filters that eschew traditional beauty trends. As one of the creators of these filters says in a <a href="https://www.dazeddigital.com/beauty/soul/article/43164/1/why-beauty3000-other-futuristic-filters-taking-over-social-media">Dazed profile</a>:</p><blockquote><em>“These filters can be used in creative new ways that partly break with the expectation of self-depiction on social media…Breaking fixed thought patterns on how we perceive gender and beauty is important and much needed.”</em></blockquote><p>Maybe this is just how we cycle influence. We tire from the popular aesthetic/approach, early-movers push towards a new approach, some subset of the influencers make the jump, while new ones are borne, and on and on we go. Or maybe early pockets of culture are at a pivot point of individuality because for the first time we don’t get to escape and recharge our batteries from the influence.</p><p>Or as <a href="https://www.newyorker.com/magazine/2019/02/11/the-machine-stops">Oliver Sacks put it best</a>:</p><blockquote><em>“(We) have given up, to a great extent, the amenities and achievements of civilization: solitude and leisure, the sanction to be oneself, truly absorbed, whether in contemplating a work of art, a scientific theory, a sunset, or the face of one’s beloved.”</em></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b3330a98c19e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Avatar-First Products & Platforms]]></title>
            <link>https://medium.com/@mhdempsey/avatar-first-products-platforms-723fd637bd35?source=rss-6e6d7bd0dfb6------2</link>
            <guid isPermaLink="false">https://medium.com/p/723fd637bd35</guid>
            <category><![CDATA[consumer-behavior]]></category>
            <category><![CDATA[social-media]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[avatar]]></category>
            <category><![CDATA[mobile]]></category>
            <dc:creator><![CDATA[Michael Dempsey]]></dc:creator>
            <pubDate>Wed, 12 Sep 2018 20:29:54 GMT</pubDate>
            <atom:updated>2018-09-12T20:37:26.397Z</atom:updated>
            <content:encoded><![CDATA[<h4>Thoughts on our future avatar identities and how to build avatar-first products</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4pHfede3-koh6EZXW-WR2A.jpeg" /><figcaption>Me using various avatar apps — From left to right: Memoji, <a href="http://facemoji.co/">Facemoji</a>, <a href="http://facerig.com/">FaceRig</a>, FaceHub, WeMoji, Puppemoji</figcaption></figure><p>Over the past year or so, next-gen avatars becoming mainstream has gone from science fiction to a probable future. While previous avatars were digital characters controlled via input devices in worlds like Second Life, I define next-gen avatars as digital assets that are manipulated mostly in real-time from physical beings (a la Animojis). These avatars will be manipulated today mostly at the head/face level, but will eventually bridge to the full body.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/0*3flTV92J_Q7dgQRQ.jpg" /><figcaption>Apple Animoji</figcaption></figure><p>The emergence of next-gen avatars has been driven by the proliferation of components like the depth sensing camera in the iPhone X and better edge compute for real-time face tracking. These have given birth to things like Animojis and <a href="https://www.theverge.com/2018/6/4/17426092/animoji-update-memoji-announced-apple-wwdc-2018">Memojis</a>, and have put real-time mobile face tracking + modification on a path to widespread availability.</p><p>But how can this new form of human expression and interaction manifest itself in society? What questions must be answered? What should startups do to be the ones to make that happen?</p><h3><strong>Avatars as the new identity layer vs. reaction layer.</strong></h3><p>I don’t believe that one can build an avatar-first product without having a clear view on which layer they are enabling.</p><p><strong>The identity layer</strong> can be put simply as the visual manifestation of names. Whether the user chooses to personify their real selves (<a href="http://twitter.com/mhdempsey">@mhdempsey</a>) or create a new identity that is far abstracted from their physical identity (<a href="https://twitter.com/startupljackson">@StartupLJackson)</a> is a personal choice.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*AgvJhNybUWRJ9V2V.jpg" /><figcaption>Memojis are Apple’s take at the avatar identity layer</figcaption></figure><p>The identity layer will have an avatar (or set of avatars) that connects a “real world” identity to a digital world identity (or identities). In this case, customization, diversity of the avatars, and ability to express yourself in real-time are incredibly important, with UGC likely to create a form of digital economy network effects.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fgiphy.com%2Fembed%2F69sRSgWfLPbfumqkXy%2Ftwitter%2Fiframe&amp;url=https%3A%2F%2Fmedia.giphy.com%2Fmedia%2F69sRSgWfLPbfumqkXy%2Fgiphy.gif&amp;image=https%3A%2F%2Fmedia.giphy.com%2Fmedia%2F69sRSgWfLPbfumqkXy%2Fgiphy.gif&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=giphy" width="435" height="333" frameborder="0" scrolling="no"><a href="https://medium.com/media/578eb690ca418959ea7fb56ef0590992/href">https://medium.com/media/578eb690ca418959ea7fb56ef0590992/href</a></iframe><p><strong>The Reaction Layer</strong> is the next visual manifestation of Gifs. Gifs have risen to prominence as understood social interactions and reactions, but aren’t directly tied to an individual. The first iterations of Animojis were reaction layer avatars. What Apple is now trying to do with Memoji is a simplistic version of the identity layer where one human creates one lifelike digital avatar of themselves.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*8DSRDjL_k_OyfYLq.jpg" /><figcaption><a href="https://www.facebook.com/spaces">Spaces</a> is Facebook’s latest attempt at owning the avatar identity layer</figcaption></figure><p>While the reaction layer seems more likely to live on in a world where large platforms have a stranglehold on identity, I am more interested in the identity layer and its ability to enable new platforms.</p><p>This isn’t to say that we can’t have both, but my view today is that the more valuable company starts as one that builds a product where a digital representation of ourself (or ourselveS) is customizable, carries brand value, easily manipulated, and creates a sense of an emotional tie between the human and the avatar. This feels most likely to happen if you also control a key platform component at the beginning.</p><p>The platform then allows that users digital identity to create defensibility once technology commoditizes and incumbents move into the space. It also tells the mainstream consumer <em>what</em> to do with their avatar (more on this later).</p><p>Whether you eventually open up your avatar identity layer, similar to how Facebook opened up “sign in with Facebook”, is a decision that can only be done from a position of platform strength.</p><p>The identity layer also has a more clear direction in which an economy emerges around it. If you own an avatar-first platform, you can control or profit from the surrounding industries such as digital fashion and accessories, voices, and perhaps even worlds that emerge.</p><h3><strong>Avatars &amp; Social Limits</strong></h3><p><strong>The identity layer allows us as a society to potentially push social barriers or sidestep social biases that we haven’t previously been able to</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/660/0*7AUuWBuKGtfweoUN.gif" /></figure><p>In all other forms of digital media we have figured out ways to target our audiences based on various demographics, and then A/B test our content with those audiences. Avatars enable this to extend this malleability to humans.</p><p>For example, a talented entertainer may only be able to be accepted by a single core audience, whether due to their sex, race, or other biases that exist. With a collection of avatar-first identities, the entertainer could possibly expand their audience outside of their physical limitations.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*5JlxXSDFNpFclrQy" /><figcaption>Black Mirror covered Waldo, a cartoon Avatar of a comedian who runs for political office in their episode <a href="https://en.wikipedia.org/wiki/The_Waldo_Moment">The Waldo Moment</a></figcaption></figure><p>These types of dynamics could extend to simple social media, journalism, and even in some dystopian futures, politics (as <a href="https://en.wikipedia.org/wiki/The_Waldo_Moment">depicted in Black Mirror</a>).</p><p>But before we get ahead of ourselves,<strong> what does an avatar-first product need to look like in my mind?</strong></p><h3>Avatar-First Product Thoughts</h3><h4><strong>Killer Use-Case</strong></h4><p>Across the 20+ face filter/avatar apps I’ve seen, the usage seems to be fairly similar. The filter/avatar is either used as a cherry on top of a social interaction (think Snap filters), or as a funny experience that people churn from after 10–30 minutes (<a href="https://twitter.com/madewithanimoji">Animoji Karaoke</a>). The reality is, people don’t know what to do with their digital selves.</p><p><strong>Those creating avatar-first products must push users towards specific interactions or use-cases, and then leave the door open for emergent behaviors.</strong></p><p>Live streaming is an area I’m incredibly intrigued by, and other forms of social media storytelling feel like obvious use-cases, but at first it’s quite possible the product must be <em>even more</em> focused than these examples. I’ll leave out anything specific so as to not step on any toes.</p><h4>How should we describe Avatars?</h4><p>When talking about allowing someone to create a digital and visual manifestation of themselves online (and perhaps someday in the real world via augmented reality) there’s a tendency to push users towards the aspirational and use words like “beautiful” or “cool”. This is dangerous.</p><p>Avatars will face mainstream backlash and the proliferation of them will have a dark side related to the same vanity we experience today surrounding social media. Because of this, in the earliest days of avatar-first products, the focus should be around what an avatar <em>enables</em> in a positive light (self-expression, confidentiality, choice) vs. what an avatar <em>fixes</em>, <em>improves</em>, or how how it makes us <em>better</em> than our real world selves.</p><p>In addition, the idea that an avatar can take on a life of its own in the real world feels incredibly significant (one could argue this is what has happened with some <a href="https://medium.com/@mhdempsey/everything-to-know-about-digital-celebrities-and-how-they-could-change-the-world-7b3bf9806073">digital celebrities</a> that are controlled by a single person). If the internet views widespread avatar creation as a tool specifically for anonymity, we could suffer the same fate as other platforms that have brought out the worst in people due to lack of accountability.</p><h4><strong>Design vs. Technology</strong></h4><p>A year ago the possibility of creating a real-time manipulated avatar on a mobile device (vs. through heavy software like <a href="https://facerig.com/">FaceRig</a> or <a href="https://www.slideshare.net/JoeyAboma/bridging-worlds-avatar-idenitity?qid=6b8469dd-5706-48cf-bad1-d2e0dd2ef912&amp;v=&amp;b=&amp;from_search=2">others</a>) seemed like a major triumph.</p><p>There are elements of technological innovation that could result in a <em>better</em> avatar experience, such as tech to allow better interactions or avatar quality, however similar to many consumer apps, technology is not the vector that this battle will be won on.</p><p><strong>While admittedly obvious, I believe that product focus that ushers in a new behavior or interaction, paired with an ecosystem that allows for a continually fresh experience and form of expression is what will win.</strong></p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fstreamable.com%2Fo%2Fkz5b7&amp;url=https%3A%2F%2Fstreamable.com%2Fkz5b7&amp;image=https%3A%2F%2Fcdn-b-east.streamable.com%2Fimage%2Fkz5b7_1.jpg%3Ftoken%3D3Ni1tj1yfh9GCyqTmzLq0A%26expires%3D1536790048&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=streamable" width="266" height="526" frameborder="0" scrolling="no"><a href="https://medium.com/media/b2dfbcbd0661d6291dacbe81d7dbf5c6/href">https://medium.com/media/b2dfbcbd0661d6291dacbe81d7dbf5c6/href</a></iframe><p>I’m still thinking through many moving parts of avatar-first product development and implications, but this post outlines my current state of mind. In meetings over the past few months with founders, product people, and engineers across avatar startups, media companies, and social platforms I found myself repeating these sentiments, so I figured them best to be published and picked apart on the internet.</p><p>As usual, if you’re working on anything in this space, <a href="http://twitter.com/mhdempsey">DM me on twitter</a> or email me at mike (at) compound (dot) vc .</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=723fd637bd35" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Everything to know about digital celebrities and how they could change the world]]></title>
            <link>https://medium.com/@mhdempsey/everything-to-know-about-digital-celebrities-and-how-they-could-change-the-world-7b3bf9806073?source=rss-6e6d7bd0dfb6------2</link>
            <guid isPermaLink="false">https://medium.com/p/7b3bf9806073</guid>
            <category><![CDATA[social-media]]></category>
            <category><![CDATA[medium]]></category>
            <category><![CDATA[digital-celebrities]]></category>
            <category><![CDATA[influencer-marketing]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Michael Dempsey]]></dc:creator>
            <pubDate>Fri, 06 Jul 2018 14:03:15 GMT</pubDate>
            <atom:updated>2018-07-06T14:23:25.130Z</atom:updated>
            <content:encoded><![CDATA[<h4>Hatsune Miku and Gorillaz were just the beginning.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*5jj2f1_Gb-yiRkBb-O5Uaw.jpeg" /><figcaption>From left to right, top to bottom: <a href="http://instagram.com/galaxia.gram">Galaxia</a>, <a href="http://instagram.com/lilmiquela">Miquela</a>, <a href="http://instagram.com/shudu.gram">Shudu</a>, <a href="http://instagram.com/blawko22">Blawko</a>, <a href="http://instagram.com/avalonage">Avalon</a>, <a href="http://instagram.com/brenn.gram">Brenn</a>, <a href="http://instagram.com/lildonnyred">Donny Red</a>, <a href="http://instagram.com/perl.www">Perl</a></figcaption></figure><p>Over the past year Americans have become increasingly aware of the digital celebrities living among us. They’ve infiltrated our social media feeds, twitch streams, advertising campaigns, and even fashion.</p><p>Having been immersed in this space for some time now, I figured it would be helpful to break down this developing market by talking through some of the history of digital celebrities, what’s happening now, and how they could change our world.</p><p><strong>I broke this into four parts:</strong></p><ul><li><strong>Part I: A primer on digital celebrities</strong></li><li><strong>Part II: How do digital celebrities progress?</strong></li><li><strong>Part III: The Cascading Effects of Digital Celebrities</strong></li><li><strong>Part IV: Remaining Questions and Decisions</strong></li></ul><h3>Part I: A primer on digital celebrities</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*RwwvWA52YtuHOgKy.jpg" /><figcaption>S1mOne was a 2002 film about a digital celebrity, Simone.</figcaption></figure><h4><strong>What is a digital celebrity?</strong></h4><p>A digital celebrity is a digitally created avatar/character that is made initially for entertainment purposes. You could call them fake influencers or fake celebrities, but in my mind, because of how inauthentic human celebrities can be, a more precise term is required. I particularly love this quote by Cameron James-Wilson, a famous fashion photographer who we’ll hear more about soon:</p><blockquote>“As soon as you start to retouch a photo, it becomes fantasy, it’s not real,” he says. “We need to get it out of our heads that just because something is labeled as a photo makes it any more real than a digital creation. In some ways, digital creations can be more accurate to real life…”</blockquote><h4><strong>Who are they?</strong></h4><p>A relatively in-depth list of these characters can be found <a href="http://dailymirage.co/list">here</a>.</p><h4><strong>What do they do?</strong></h4><p>The vast majority of these characters today live on social media and post their day-to-day lives and thoughts with the world. Miquela (one of the most famous) has done <a href="https://www.highsnobiety.com/p/lil-miquela-thedropla-barneys-tshirt/">a shirt collab</a>, <a href="https://soundcloud.com/lilmiquela">released music,</a> and <a href="https://www.thecut.com/2018/04/lil-miquela-hack-instagram.html">gotten in fights with other digital celebrities</a>. <a href="https://www.instagram.com/noonoouri/">Noonoouri</a> has <a href="https://www.vogue.com.au/vogue-codes/news/meet-noonoouri-the-virtual-instagram-influencer-loved-by-kim-kardashian-west-and-dior/news-story/dfeb2658e3bbc7b2cfd11d0d26eb5bbc">worked with Kim Kardashian</a>, <a href="https://www.instagram.com/blawko22/">Blawko</a> streamed on Twitch, and a bunch of others post just like your other friends that live on Instagram.</p><h4><strong>Is this new?</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/930/0*DwQf19z1gjDvZsj6.jpg" /><figcaption>Hatsune Miku</figcaption></figure><p>The closest thing to these digital celebrities that many point to is Hatsune Miku, with some Americans also mentioning <a href="http://www.gorillaz.com/">Gorillaz</a>. Hatsune Miku was created in 2007 and is a Japanese holographic pop star that is powered by her large fanbase and easy to use creation tools. Touted a “crowdsourced pop star”, she has generated hundreds of millions to billions of dollars in revenue across all of her entities/merchandise. There are many lessons to be learned from her rise, and a great paper documenting this can be found <a href="http://www.urop.uci.edu/journal/journal13/01_le.pdf">here</a>.</p><h3>Part II: How do digital celebrities progress?</h3><h3>Monetization</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*bYUXDcEe8o41nXkZ0_uDRg.png" /></figure><p>While a large portion of these celebrities are still in the experimental phase, there is a subset that are clearly pushing to monetize their social media presence in some way. Some may take a vertical focus, like food blogger <a href="https://www.instagram.com/p/BjqvG_ij5yl/?taken-by=mcohen77">Melissa Cohen tagging Blue Apron in her posts</a>, while others may just try to build a legion of fans and then figure out how to monetize their following. Either way, every major online vertical that is dominated by some type of influencer will eventually be targeted by a digital celebrity.</p><p>Simple markets like food blogging make sense today, however as technology infiltrates these markets in a meaningful way, the quality of content will increase, making even more markets quite susceptible to the digital vs. human influence wars.</p><h4>Digital Influencers can scale past humans.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*RDfpJoLmTupv9lUXqhz-Xw.png" /><figcaption><a href="http://instagram.com/perl.www">Perl</a></figcaption></figure><p>Many influencers or social media personalities have attempted to break out of their domain and into the mainstream areas of music, TV, or film. Few have succeeded.</p><p>Often the talent agencies tell me that the level of talent/skill set to achieve stardom in these areas is a difficult barrier to entry, along with the strong brand association and fanbase of a given character that may make it hard for them to work within a larger portfolio of brands.</p><p>With digital celebrities, no longer can you age out of your demographic. You can freeze time, fast forward, or rewind, creating both a more relatable character as well as a much more in-depth storytelling mechanism where history, present, and future are all available for creation.</p><p>With potentially different humans controlling different skillsets of your celebrity, you then have the potential to close the gap from influencer/social media personality to mainstream talent.</p><p>Miquela’s pop album was an example of this and others will certainly follow suit over the next few years.</p><h3>Ninja on Twitter</h3><p>Wanna know the struggles of streaming over other jobs? I left for less than 48 hours and lost 40,000 subscribers on twitch. I&#39;ll be back today (Wednesday) grinding again.</p><p>In addition to having the ability to scale talents, <strong>digital celebrities can scale utilization</strong>. While scarcity can drive the price of services up for a celebrity, there are some where volume is what creates revenue. A recent example of this is when famous Twitch streamer <a href="http://twitter.com/ninja">Ninja</a> took 2 days off from streaming and thus lost 40,000 subscribers, costing him over $100k. When multiple people can utilize a likeness across different talent classes, you can imagine the possibilities that are unlocked by removing the volatility and deficiencies of humans.</p><p>Now these celebrities can be on tour, while filming commercials, modeling their new fashion line, or even starring in <em>their own</em> TV show or movie.</p><h4>Building your own universe</h4><p>In addition to creating a single character that can be monetized as an entity, another strategy will be telling a multi-character story and building your own universe. The lowered barriers of social media allow a cadence of storytelling to happen on a daily or even hourly basis, even if done just through photos (as mostly occurs today). I view this similar to popular text storytelling platforms such as <a href="https://itunes.apple.com/us/app/hooked/id1024818709?mt=8">Hooked</a>, but instead there are characters with a <em>physical</em> manifestation, living on the same platforms as your friends, that can further connect the audience to their story.</p><p>By building a character universe, you allow for some form of virality to emerge via word of mouth growth of a given character, and thus the IP value of the universe.</p><p>Example: If I follow an Instagram account that details the life of a scientist named Rick Sanchez, but my friend follows his grandson Morty, we inherently will want to talk more about the other people showing up in their stories despite our affinities for different characters. We’re invested into the Rick &amp; Morty universe and once Morty’s dad, Jerry creates an Instagram account, we’ll <em>both</em> likely follow it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/897/0*JP0E-nj81QPHW8Hd" /><figcaption>Bermuda (left) &amp; Miquela (right) are both created by <a href="http://brud.fyi/">Brud</a></figcaption></figure><p>A more specific example of this is the ongoing interactions of <a href="http://brud.fyi/">Brud</a> creations, Miquela, Bermuda, and Blawko. This allows you build a flagship character (in this case, Miquela) and then spin up multiple characters with a much faster ramp time than an entirely new account. Similar to how Marvel has created what feels like 100 movies about different Avengers characters in the past 3 years.</p><p><a href="http://crypttv.com/">CryptTV </a>is a tangential example of this as well. CryptTV uses rapid iteration and the lowered barrier of online distribution to test and expand upon universes and characters focused on horror.</p><h4>Digital Celebrities must get more innovative</h4><p>While many are running today’s playbook of simple storytelling one photo, inspirational quote, or anecdote at a time, there has been very little innovation on how to utilize the digital component to create a truly unique product. Creators must innovate.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FDyhp8E4WcLM%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DDyhp8E4WcLM&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FDyhp8E4WcLM%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/46ef6ac20da65eb15c599371c85945d7/href">https://medium.com/media/46ef6ac20da65eb15c599371c85945d7/href</a></iframe><p>Cameron-James Wilson is showing early signs of this with <a href="https://www.instagram.com/thediigitals/">The Diigitals</a>, an all-digital modeling agency. He began by creating a black supermodel, <a href="http://instagram.com/shudu.gram">Shudu</a>, and recently released two new models for his digital agency in <a href="http://instagram.com/brenn.gram">Brenn</a>, a plus-sized model, and <a href="http://instagram.com/galaxia.gram">Galaxia</a>, the world’s first alien model.</p><h3>Michael Dempsey on Twitter</h3><p>And it&#39;s here. Two more models including one non-humanoid.</p><p>Shudu, originally created as a commentary on minority representation within the modeling industry, has created <a href="https://wwd.com/eye/people/shudu-digital-fashion-model-avatar-1202683320/">a fair amount of controversy</a>, but Galaxia could be part of a different thought experiment.</p><p>Could an Alien supermodel help us identify what we care about as humans when it comes to the personas we let influence us? Could it perhaps even change how we define beauty?</p><p>If an alien-figure can eradicate <a href="https://en.wikipedia.org/wiki/Uncanny_valley">the uncanny valley problem</a> and rise to prominence, it sets the stage for many other celebrity types, and potentially removes the need to create celebrities in ultra-realistic, humanoid-focused, high-resolution 3D models.</p><h3>Part III: The Cascading Effects of Digital Celebrities</h3><h4>The New Digital Identity Layer</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*whEEm4uhtDQfjEr4.jpg" /><figcaption>Samsung AR Emoji (left) vs. Apple’s Memoji (right)</figcaption></figure><p>One of the interesting takeaways from this wave of both newly created digital characters as well as lowered barrier to digital recreation of human entities is what do humans need in order to create a connection to other intelligent beings?</p><p>I’ve already written that <a href="http://notes.michaeldempsey.me/post/172866628574/were-ready-for-ai-friends">we’re ready for AI friends</a> and believe that digital celebrities could lead to a world where digital identity layers taking on physical traits is more socially accepted.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*VwBmYe_mm5BB0PPH.jpg" /><figcaption>Rigs used to capture and map facial details have gone from this to a mobile phone. (via Avatar, released in 2009)</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vjAQMm9VoMegOxxHOfcMzw.png" /><figcaption>Via <a href="http://facemoji.co/">Facemoji</a></figcaption></figure><p>Bitmojis were a small step towards this future, but now Memojis, AR Emoji, and everything in between have shown the major tech companies beginning to recognize this trend, while startups like <a href="https://facemoji.co/">Facemoji</a> have their own unique take on this thesis. These technologies could allow us to either create entirely new characters/identities by implementing voice modulation, or build economies on top of this digital identity via avatar layer.</p><h4><strong>Human celebrities becoming digital celebrities</strong></h4><p>Human celebrities could also utilize their likeness to create new alter-egos or digital manifestations of themselves that live online. In a world where many stars are seeing their appeal to younger generations fade vs. those on social media consistently “connecting with their fans”, a digital persona could be the perfect balance between connection, effort, and privacy. Some celebrities like Daft Punk and Marshmello have managed to maintain relative privacy by doing this in an analog way already. There are issues to be hashed out around verification, but that is another topic <a href="https://twitter.com/mhdempsey/status/851476978519744515">I’ve covered ad nauseum</a>.</p><h4>The Displacement of Human Celebrities and influencer economies</h4><p>As this segment of the market matures, I do believe the idea of “digital vs. human influence wars” will become more of a reality. Some have expressed concern that these celebrities could take the jobs of humans, and that is likely true in some cases.</p><p>The micro-economies that have emerged within social media are ripe for disruption and digital counterparts have clear competitive advantages.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3DZqs_5-QtLng-x8wS_OhQ.jpeg" /><figcaption>Lil Donny Red, created by Jake Paul’s Team10</figcaption></figure><p>Interestingly enough, we’ve seen famed YouTube star Jake Paul’s Team10 experiment with this by creating <a href="https://www.instagram.com/lildonnyred/">LilDonnyRed</a>, a funny red-headed digital celebrity.</p><p>Cameron James-Wilson, briefly touched on the idea of human displacement when he was criticized for Shudu taking modeling jobs from the underrepresented humans she was originally meant to advocate for:</p><blockquote>“I don’t think Shudu or other 3-D models take anything away…Eventually, if this catches on, people will be more and more employed. OK, one model might not be employed, but maybe 50 other people are [working] in order to take that kind of space. You have to be very lucky to be a model. You have to be born genetically beautiful in every way and pretty much perfect. What kind of world would be better: one where somebody who works hard, who’s a talented artist can create a model or somebody who’s born lucky?”</blockquote><h4><strong>Proliferation of Celebrities into everyday life</strong></h4><p>Digital celebrities will broadly be utilized for entertainment purposes first and then as major media or marketing vehicles, but then what?</p><p>With development of machine learning tools we could see the AI friend future I discuss start enabling these beings to scale towards commercial tools and more.</p><p>Why should your customer service agent not at least sound like your favorite celebrity? Or your digital therapist? There is still the issue around the economics of creating scarcity for your celebrity, however it wouldn’t surprise me to see (or at least hear) some of your favorite celebrities serving you a targeted ad or helping you with a problem, sometime in the future.</p><h3><strong>Part IV: Remaining Questions and Decisions</strong></h3><p>There are multiple open-ended questions left to be answered or that I’ve alluded to in this post.</p><p><strong>In general we must better understand:</strong></p><ul><li><strong>Consumption: </strong>Are the best mediums to consume this content already built today, or is there a place where these types of celebrities should live on the internet?</li><li><strong>Society: </strong>As we’re shifting from a <a href="https://machinelearnings.co/moving-from-a-default-trust-to-default-skeptic-society-cbf9adeb4971">default trust to default skeptic society</a>, is now the best time to be flooding our lives with digital characters representing themselves as real? Will consumers grow tired of things they know they have a 0% chance of physically interacting with?</li><li><strong>Economics:</strong> Will the economics of a significantly more scalable celebrity play out in practice and how could they change existing industries? As it is today influencers are largely seen on a different tier from an earnings perspective as mainstream celebrities. Could we be creating another tier that will not be able to monetize at the level that human counterparts can?</li></ul><p><strong>Creators must make decisions related to:</strong></p><ul><li><strong>How the form factor allows a new type of innovative entertainment or experience or connection.</strong></li><li><strong>How transparent to be around the “false” nature of these characters and their stories.</strong></li><li><strong>Whether to build something hyper-realistic and take aim at the Uncanny Valley effect, or tell stories with clearly non-human entities.</strong></li><li><strong>The overall creation process and how it can be scaled. I have </strong><a href="https://twitter.com/mhdempsey/status/928342329735475200"><strong>many thoughts on this</strong></a><strong>. Stay tuned.</strong></li></ul><p>While we are still very early in the world of digital celebrities, they will continue to push into many different areas our lives. Digital celebrities may feel like something out of a dystopian sci-fi novel today, but I look forward to seeing how they change how we consume and interact with content and each other moving forward.</p><p>If you’re working on or building anything related to the topics discussed in this post, I’d love to hear from you <a href="http://twitter.com/mhdempsey">on twitter </a>or via email.</p><p>Stay tuned for a future announcement on something we’ve been working on at Compound with a team of experienced entertainers, content producers, and technologists.</p><p>—</p><p>Thank you to <a href="https://medium.com/u/86aabb3e4f5c">Dylan Flinn</a> and <a href="https://medium.com/u/f9c0c051bfef">Vijay Sundaram</a>. Vijay and I discussed this topic at length in <a href="https://soundcloud.com/v0concepts/ai-generative-celebrities">a prior audio recording</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7b3bf9806073" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Scaling Autonomous Vehicles with Machine Learning — Our Investment in Wayve]]></title>
            <link>https://medium.com/@mhdempsey/scaling-autonomous-vehicles-with-machine-learning-our-investment-in-wayve-4d9bcf127932?source=rss-6e6d7bd0dfb6------2</link>
            <guid isPermaLink="false">https://medium.com/p/4d9bcf127932</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[self-driving-cars]]></category>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Michael Dempsey]]></dc:creator>
            <pubDate>Tue, 22 May 2018 13:52:24 GMT</pubDate>
            <atom:updated>2020-07-11T13:38:25.570Z</atom:updated>
            <content:encoded><![CDATA[<h4>From a 2015 email to a seed investment: How Wayve uses novel machine learning to bring autonomous vehicles to the masses.</h4><p><strong><em>TL;DR — Compound led </em></strong><a href="http://wayve.ai/"><strong><em>Wayve’s</em></strong></a><strong><em> seed round. Wayve is an autonomous vehicle company utilizing a unique machine learning approach to drive cars autonomously, with little data, in novel environments.</em></strong></p><p>In 2015 I read <a href="https://arxiv.org/pdf/1511.00561.pdf">a research paper on semantic segmentation for scene understanding called SegNet</a>. I sent one of the authors, <a href="https://alexgkendall.com/">Alex Kendall</a>, an email congratulating him on his work and asking about commercialization plans.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1016/1*p4bVsspeltveIHHeJTzsQg.png" /><figcaption>Not my best cold email.</figcaption></figure><p>Alex said he was excited to get it running in a car soon, but wasn’t sure how.</p><p>Fast forward to 2017 when our Board Partner <a href="https://www.linkedin.com/in/drewjgray/">Drew Gray</a> told me about a team of impressive University of Cambridge PhDs working on a new approach utilizing bleeding-edge machine learning techniques to get cars to drive themselves in novel environments.</p><p>Upon hearing one of their names I quickly searched my inbox and my email about SegNet from years prior popped up. I bumped the thread and re-met Alex and Amar.</p><p>Within an hour of meeting both of them, I realized that they were a rare breed of co-founders; both highly technical, but also incredibly entrepreneurial with a deep desire to commercialize and build a real, big business. After digging deeper into their research, it was clear that what they were building was special. Having seen 20+ AV startups in the prior 2 years, Wayve had by far the most differentiated technical approach I had ever seen. We led the seed round and I joined the board along with <a href="https://www.linkedin.com/in/gabrielmatuschka">Gabe Matuschka</a> at <a href="https://www.fly.vc/">Fly Ventures</a>.</p><h4>Why do we need another autonomous vehicle startup?</h4><p>In my time spent in the AV space I’ve repeatedly heard about the ceilings that various teams have hit along the way. While early teams at 510 Systems, which eventually became Waymo, saw rapid progress to ~80% autonomy, it was only after years of work they realized they needed to essentially start over from a technical perspective to reach 90%+ autonomy.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*aC4QTfwStdsZeUjt." /><figcaption><a href="https://thelastdriverlicenseholder.com/2018/02/01/disengagement-report-2017-the-good-the-bad-the-ugly/">2017 Disengagement report</a> — Disengagements per 1,000 miles.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/752/1*cyMsILbD9ne6-j9t4Qi2uw.png" /><figcaption><a href="https://reader.elsevier.com/reader/sd/C486233F839CF4AFBB86F084B0EE54D09CE69637D6819AD805216BC6F203DD8302805FB79FB39E65E9F68D4F1EE5E021">Autonomous Vehicle Disengagement Trends &amp; Statistics</a></figcaption></figure><p>We’re now starting to see this again with <a href="https://spectrum.ieee.org/cars-that-think/transportation/self-driving/have-selfdriving-cars-stopped-getting-better">slowing progress in disengagement rates</a> (an admittedly flawed metric) and continued failures across <a href="https://www.theverge.com/2018/4/12/17229518/tesla-ntsb-autopilot-crash-investigation-removed">multiple</a> <a href="https://www.theinformation.com/articles/inside-cruises-bumpy-ride-the-limits-of-self-driving-cars">major</a> <a href="http://www.newsweek.com/self-driving-uber-crash-likely-cause-fatal-car-accident-detailed-report-913407">players</a>. The top AV companies today drive in a few cities at most, with $100k+ of sensors, and plenty of existing maps and training data in those cities. And yet, few are on a realistic path towards scalable autonomy.</p><p>Today, the majority of companies utilize deep learning for computer vision (the perception layer) to allow the car to understand what it sees. The more annotated data, the better. They then use a rules-based approach to make path planning and driving actions that the vehicle takes once it understands what it sees. This results in the need to anticipate an almost infinite number of edge-cases across each layer of the self-driving stack. Where humans can generally understand and drive in completely new environments or situations, with this rules-based approach, today’s autonomous vehicles can not generalize nearly as well.</p><p><strong>The traditional rules-based robotics approach is not the answer.</strong></p><p>When looking back at <a href="http://notes.michaeldempsey.me/post/165842323964/the-autonomous-vehicle-engineer-landscape-and-what">the past two years of autonomous vehicle engineering</a> I concluded that if we are going to truly bring AVs to the masses and break through the latest ceilings, we must use more advanced forms of machine learning.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FnsoY0hOfpig%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DnsoY0hOfpig&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FnsoY0hOfpig%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/3a6ced6879c72547f5e434cdecd9f914/href">https://medium.com/media/3a6ced6879c72547f5e434cdecd9f914/href</a></iframe><p><strong>Wayve is building truly scalable autonomy by innovating on the bleeding edge of bayesian deep learning (BDL), reinforcement learning (RL) , and more.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pX95eCZCn0EiT9OKstkQYw.png" /><figcaption>Deep Learning vs. Bayesian Deep Learning — via <a href="https://alexgkendall.com/media/presentations/IJCAI_kendall_2017.pdf">Concrete Problems for Autonomous Safety</a></figcaption></figure><p>By utilizing BDL and RL in an end-to-end model, Wayve is able to estimate uncertainty and propagate actions based on the entire AV framework, including the prediction and decision layers of the vehicle, greatly increasing safety. In AVs, the “state” refers to building out a representation of what is going on (and thus needed to drive), such as where cars and pedestrians are located, what the traffic lights are doing, and other known variables. Most teams hardcode this state of the vehicle today based off of rules (as previously mentioned), and if a given state of the vehicle doesn’t match the hand-coded rules, the vehicle can’t drive. This is why when many vehicles reach situations they haven’t seen before, they stop.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hywpTzAt6ssuZd2zfPQyCA.png" /></figure><p>Wayve utilizes machine learning to learn the state of the vehicle and make a decision from that state to scale to unknown and complex scenarios. This enables a variety of decision making and data efficiencies, including the understanding of multi-agent decision making (i.e. complicated scenarios that arise where multiple vehicles are impacting each other, like a roundabout).</p><p>Put simplistically, Wayve utilizes a unique end-to-end machine learning approach to drive cars autonomously, with little data, in novel environments. This also means that their software enables a car to drive itself safely using only understanding of what it can see, just like humans do. And as of today, their team of incredible PhDs, AI/ML experts, roboticists, and software engineers have been able to do it with less than $5k in sensors per vehicle, on public UK roads.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*zt8EClMuv9iQj1KhJH5OSg.gif" /><figcaption>Wayve’s sensor stack is largely made up of cameras</figcaption></figure><p>While the AV industry is still in its infancy, the early results from the Wayve team have been incredibly promising. I couldn’t be more excited to be along for the proverbial + literal ride, and for the team to pull back the curtain a bit on what they’ve been working on over the past year.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*q8fKcfssn90AmrrRwssh5g.jpeg" /></figure><p><strong>If you’re interested in joining one of the best self-driving teams in the world, </strong><a href="https://jobs.lever.co/wayve"><strong>Wayve is hiring</strong></a><strong>!</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4d9bcf127932" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI.Reverie]]></title>
            <link>https://medium.com/@mhdempsey/ai-reverie-f67251b7cf1b?source=rss-6e6d7bd0dfb6------2</link>
            <guid isPermaLink="false">https://medium.com/p/f67251b7cf1b</guid>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[venture-capital]]></category>
            <dc:creator><![CDATA[Michael Dempsey]]></dc:creator>
            <pubDate>Tue, 08 May 2018 21:49:29 GMT</pubDate>
            <atom:updated>2018-05-08T21:53:10.194Z</atom:updated>
            <content:encoded><![CDATA[<h4>Turning daydreams of simulation into scalable synthetic data — Announcing Compound’s investment in AI.Reverie</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ADTnro_zxB5vV7EwKDdnbw.png" /></figure><p>A few years ago while spending time with autonomous vehicle (AV) engineering teams I kept asking what their largest bottlenecks were to improving autonomy. One constant I kept hearing was more annotated training data. The bottleneck of taking sensor feeds and having humans annotate them in order to train deep learning perception models was resulting in <a href="https://www.ft.com/content/36933cfc-620c-11e7-91a7-502f7ee26895">teams of hundreds and even thousands of humans performing this time-consuming task</a>.</p><p>At the same time, I started to learn more about simulation and the value that synthetic data could bring to this pipeline. I wrote about this extensively <a href="https://medium.com/@mhdempsey/how-to-drive-10-billion-miles-in-an-autonomous-vehicle-7c9b81c1ddac">here</a> and set out to make an investment in the space with the thesis that synthetic data would be a key accelerator in the arms race towards autonomy for AVs, and that the addressable market for other robots and autonomous systems would expand over time.</p><p>I’ve said this before, but being thesis driven allows us to move efficiently within categories we’ve gone deep on, but also makes it even more powerful when we can recognize when founders shatter <a href="https://medium.com/@compoundvc/compound-theses-1f97fc56e436">our theses</a>.</p><p>After meeting with 10+ companies from pure simulation to larger scale world building, all related to autonomous vehicles, I finally met Daeil Kim &amp; Paul Walborsky. They introduced me to <a href="http://aireverie.com/">AI.Reverie</a>. My daydreams of all the possibilities of synthetic data had turned into a reality.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F42QZggpb--M%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D42QZggpb--M&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F42QZggpb--M%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/1c89d10295655c3c46e8eab320cdb2e7/href">https://medium.com/media/1c89d10295655c3c46e8eab320cdb2e7/href</a></iframe><p>While AV simulation technology continues to be extremely valuable, Daeil and Paul, along with the rest of the AI.Reverie team have built a simulation platform that creates large scale, high-fidelity worlds algorithmically in order to train machine learning models across markets from more traditional enterprise robotics all the way to <a href="https://www.aireverie.com/case-study/">tracking elephants</a>. Not only was I blown away by the scalability of what they had built, but also the level of diversity at which they could generate virtual worlds based off of real location data.</p><p>Their vision for the future recognized the value of fully annotated, error-free synthetic data and simulation across use-cases I had not seen any other company even conceptualize, let alone <em>commercialize</em>.</p><p>We’re super excited to be backing the entire AI.Reverie team in their pursuit of helping train AI to understand the world.</p><p>You can <a href="https://www.aireverie.com/">check out the website</a> to learn even more about what AI.Reverie is building. To read more about the financing round check out the press release <a href="http://www.prweb.com/releases/2018/05/prweb15471141.htm">here</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f67251b7cf1b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Moving from a default trust to default skeptic society]]></title>
            <link>https://medium.com/@mhdempsey/moving-from-a-default-trust-to-default-skeptic-society-cbf9adeb4971?source=rss-6e6d7bd0dfb6------2</link>
            <guid isPermaLink="false">https://medium.com/p/cbf9adeb4971</guid>
            <category><![CDATA[generative-adversarial]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[frontier-tech]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[future]]></category>
            <dc:creator><![CDATA[Michael Dempsey]]></dc:creator>
            <pubDate>Wed, 14 Feb 2018 16:02:19 GMT</pubDate>
            <atom:updated>2018-02-16T20:58:05.333Z</atom:updated>
            <content:encoded><![CDATA[<h4>What the future of technology does to the future of trust.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ydVA6FCysaqIiApFMXB3eA.jpeg" /></figure><p>I’ve been spending a lot of time thinking about how new forms of AI and machine learning will shape society over the next decade. One area I’ve gone deep on is how <a href="http://notes.michaeldempsey.me/post/159418832409/audio-video-manipulation">the widespread manipulation and generation of digital assets</a> will manifest itself in our daily lives.</p><p>My main takeaway in my research thus far is this:</p><p><strong>We are moving from a default trust society to a default skeptic society. </strong>Or put another way, while today humans largely trust what they hear/see, tomorrow they will not.</p><h4>How did skepticism creep into society today?</h4><p>In my lifetime I’ve seen this shift towards skepticism take place. ~10 years ago, people believed what they could see in relative high fidelity and what they were told in major media outlets. Photoshop made people skeptical of incomprehensible images. Smartphones, social media, and internet publishing has given to the rise of clickbait and “fake news” making society in 2018 increasingly more skeptical about internet content (and <a href="https://www.wired.com/story/inside-facebook-mark-zuckerberg-2-years-of-hell/">the platforms that push it</a>) than in 2017.</p><p>This started with the rise of creative editing tools, has moved to a data-driven understanding of human beings and how they consume information, and ends with how machine learning upgrades the scalability in creating and manipulating <a href="https://twitter.com/mhdempsey/status/851476978519744515">all types of visual and audio assets</a>.</p><p>But today seeing is still believing and we trust certain sources to continue to report the truth and we default believe them.</p><h4><strong>What happens when that trust starts to erode?</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*atz5QbWOSutqOs6AjqdkXA.jpeg" /><figcaption><a href="https://www.youtube.com/watch?v=9Yq67CjDqvw">Synthesizing Obama</a></figcaption></figure><p>What happens once the first person/group displays the power of scalable and believable fake audio or video as a weapon? We recently saw an early sign of this with <a href="https://motherboard.vice.com/en_us/article/bjye8a/reddit-fake-porn-app-daisy-ridley">DeepFake</a>, a technology used to superimpose celebrities’ faces in porn (<a href="https://www.youtube.com/watch?v=BU9YAHigNx8">among other things</a>).</p><p>The result? The internet collectively freaked out and multiple platforms took an uncharacteristically censorship-heavy approach, quickly banning the content. And while deepfakes often manipulate something as sacred as sex acts, the targeting of celebrities in our sex-desensitized world makes me believe that the potential impact on society technologies like this could have will not be realized by similar manipulations, but instead with what could happen next.</p><p>DeepFake and whatever follows it will plant a seed of doubt and skepticism in people’s minds and destroy the idea that you “have to see it to believe it”. But only once false assets are used for something as meaningful as setting off some sort of widespread trigger across financial markets or government agencies, or when they change the outcome of something as large as an election (again?), will that seed blossom into a mindset shift for society.</p><p>Eventually this could trickle down from “important to verify” facts to day-to-day conversations. As a generation that records everything we do in order to create social proof, will our existing digital footprints become worthless as quickly as our feeds became “valuable”? Will the rebirth of trust in these types of assets come from a return to analog mediums?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/815/1*YZ9veeFxBkmSXdjdwCcIDw.jpeg" /><figcaption>by <a href="https://www.instagram.com/p/BdbqJJ8jYrO/?hl=en&amp;taken-by=beeple_crap">Beeple</a></figcaption></figure><p><strong>Perhaps the idea of “real versus not” will mean less</strong> when our worlds are already being overrun by multiple realities and we are hopping between physical and digital worlds.</p><p>These are all interesting exercises to think through, but a lot of how I look at the world as an investor as well as a human opining on the future is through <a href="https://medium.com/@mhdempsey/will-driverless-vehicle-autonomy-commoditize-d68eb4697c3e">inflection points</a>. While I’m not confident in what will be the catalysts for this particular case, I’m fairly confident that we are fast approaching an inflection point in how humans internally compute all new information. From there we have to figure out whether the power of shared beliefs and incentives can keep us a default trust society, or if a new wave of technology will make us default skeptics.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cbf9adeb4971" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to win with robots in restaurants]]></title>
            <link>https://medium.com/@mhdempsey/how-to-win-with-robots-in-restaurants-9c984f2ee3d1?source=rss-6e6d7bd0dfb6------2</link>
            <guid isPermaLink="false">https://medium.com/p/9c984f2ee3d1</guid>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[restaurant]]></category>
            <category><![CDATA[food]]></category>
            <category><![CDATA[automation]]></category>
            <dc:creator><![CDATA[Michael Dempsey]]></dc:creator>
            <pubDate>Wed, 03 Jan 2018 19:09:08 GMT</pubDate>
            <atom:updated>2018-01-03T19:12:14.574Z</atom:updated>
            <content:encoded><![CDATA[<h4>Evolving thoughts on where the venture scale opportunity lies at the intersection of robotics and food service.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yGOipOz7L4L8Z4e50Wr6Ew.png" /></figure><p>The promise of robots creeping their way into the kitchen continues to inch closer to reality, however as a venture investor I’ve been thinking a lot about what the right implementation is for the unique beast that is the food service industry.</p><p>When we think about scaling traditional industries that have a robotics implementation, the question we have to answer, which <a href="https://medium.com/u/f9c0c051bfef">Vijay Sundaram</a> helped me think through, is: <strong>Does the robot help the restaurant group scale so that a traditionally non-venture scale business can achieve lower-cost scale due to an unfair tactical advantage?</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*6mze9rN0LB6FAK3BG80BxA.png" /></figure><p>The key thing with robotics and automation broadly is that they create efficiencies. Restaurants have plenty of inefficiencies vs. traditional tech companies. <a href="http://smallbusiness.chron.com/average-profit-margin-restaurant-13477.html">Margins are low</a>, your product has a high time decay, and there are a wide range of persona profiles on the labor side.</p><h3>So the three problems really are:</h3><ol><li><strong>Labor</strong></li><li><strong>Scaling</strong></li><li><strong>Food cost</strong></li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*YBpmcvzSHgVuOO47JJwDAw.gif" /><figcaption><a href="http://misorobotics.com/">Miso Robotics’ Flippy</a></figcaption></figure><p><strong>For Labor,</strong> robots can create efficiencies to help scale the servers or cashiers, enabling them to improve the customer experience. They can also enable the kitchen to cook quickly, more efficiently, and with better repeatability.</p><p>This repeatability also solves one of the largest friction points in scaling<strong> </strong>food output. However, a similarly important, but less talked about issue <strong>for scaling </strong>most concepts today is actually location scouting. If you look at an emerging restaurant group, location scouting is really hard once you reach a certain mass (~25+ locations). Today, that won’t be automated by robotics.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/786/1*l9Uc6RfJAg9vNuqHdnH4Vw.gif" /><figcaption><a href="https://www.abundantrobotics.com/">Abundant Robotics’ </a>apple picking robot</figcaption></figure><p><strong>For food cost</strong>, it’s less clear. Increasingly restaurant groups want to become more integrated into their supply chain by working with local food sources, but the costs only come down as the actual process costs come down (or at true vertical integration). Agricultural robots can lower food costs due to increased margin on the supplier side, but in terms of looking at a restaurant as an independent entity that’s more of a trickle down effect.</p><h3>So…What is the right path for a food service robotics company?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/610/1*dX9t43rjEzAg-Li9yKRPSA.png" /><figcaption>An early build of the <a href="http://momentummachines.com/">Momentum Machines</a> burger robot</figcaption></figure><p><strong>For Labor, look at commodity hardware </strong>— What we’ve seen over the past few years is custom robots that want to help existing restaurant groups automate a specific process. These increase efficiencies, but not in a scalable way because a lot of the existing infrastructure in this industry is going to be really difficult to change. This is similar to what we’ve seen in warehouse automation where robots that require infrastructure change make for a drastically more difficult sales process.</p><p>Instead, we should look at the task that needs to be automated and utilize a combination of commodity hardware/sensors paired with software that does the heavy lifting to adapt the robot to various footprints. This is a <em>really</em> hard technical problem.</p><p>But whether you agree with this approach or you say “ovens are large pieces of custom hardware!”, you’re still selling to restaurants, which is <em>really</em> difficult. Just ask<a href="http://www.wiley.com/legacy/products/subject/business/forbes/kroc.html">Ray Kroc</a>.</p><h3>Own the entire experience = QSR 3.0</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/540/0*j3wGtbbgZd2U_fuN.jpg" /></figure><p>Speaking of Ray, McDonald’s rose to prominence because of elite execution due to a <a href="http://www.wiley.com/legacy/products/subject/business/forbes/kroc.html">rethinking of how to create and serve consistent food</a>. They rode this operational edge to build a restaurant + real estate empire worth over $130B.</p><p>Of course, over time their edge has commoditized, but the underpinning of the brand was operational excellence. We’ve seen a similar early edge develop with QSR’s like <a href="https://www.tumblr.com/new/sweetgreen.com">Sweetgreen</a> which has used better sourcing, data science, and mobile implementation to gain an edge in UX over a slew of other salad concepts.</p><p>And for the first time in awhile, <strong>operational efficiencies and thus competitive advantages can quantifiably increase in restaurants because of robotics. </strong>This is a huge opportunity that you must own and build your brand off of.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kVfzhxbK9sFRGZgluarA-w.png" /><figcaption><a href="http://zumepizza.com/">Zume Pizza’s</a> pizza assembling robot</figcaption></figure><p>Zume pizza has done a great job at this by building custom robots (with ABB) for pizza assembly while also iterating on the actual delivery/cooking mechanism with pizzas that cook en route.</p><h4><strong>But it’s still unclear how the robot helps scaling.</strong></h4><p>Restaurant people tend to have split beliefs on what allows “venture” scale in restaurants. The operations side believes that location scouting is more important, while the food side not surprisingly believes that consistency out of labor is more important.</p><p>Economically, the robot could enable better cash flow for the group as labor costs fall due to lack of turnover and fewer staff, while also removing difficulty of food consistency, but that doesn’t really scale in terms of the location implementation.</p><h4><em>“Our robot allows us to go where other restaurants can’t”</em></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*Bm-HBydskE-Mp31TZpwZHg.jpeg" /><figcaption><a href="https://www.cafexapp.com/">CafeX</a> in San Francisco, an example of robotics enabling a smaller footprint</figcaption></figure><p>Some people tell the story that robot-driven concepts allow the group to operate in smaller or differentiated real estate footprints in cities. This is where tech people coming into restaurants could get burned.</p><p>While yes, you technically can put a robot-heavy QSR in a hotel lobby, or a smaller footprint, in order to build the next McDonald’s you must never forget that you are selling a lasting food-centric experience. CafeX has taken this approach, however coffee is a fundamentally different product vs. food. Moving towards a hyper-efficient food concept too early could lose mass market appeal, killing the <a href="https://en.wikipedia.org/wiki/Third_place">third place concept</a>, and eroding your long-term defensibility and brand loyalty once automation commoditizes.</p><h3>Summary</h3><p>My thoughts continue to evolve around how robots can impact the food service space, however as it stands today, this note is a quick outline on how we’re thinking about the industry at <a href="http://compound.vc/">Compound</a>. While longer-term automation in some forms could commoditize, today we believe the venture scale opportunity lies in:</p><ul><li>Some form of commodity hardware.</li><li>Owning the entire stack from tech to food service</li><li>Team with operating experience in food service.</li><li>Focus on improving UX while maintaining human feel that lends itself to the mass market in the short term.</li></ul><p><strong>If you have any thoughts on this please don’t hesitate to reach out on </strong><a href="http://twitter.com/mhdempsey"><strong>twitter</strong></a><strong> or email me at mike (at) compound (dot) vc</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9c984f2ee3d1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Tia]]></title>
            <link>https://medium.com/@mhdempsey/tia-5125413e54af?source=rss-6e6d7bd0dfb6------2</link>
            <guid isPermaLink="false">https://medium.com/p/5125413e54af</guid>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[womens-health]]></category>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[vc]]></category>
            <dc:creator><![CDATA[Michael Dempsey]]></dc:creator>
            <pubDate>Tue, 10 Oct 2017 20:19:30 GMT</pubDate>
            <atom:updated>2017-10-10T20:19:30.348Z</atom:updated>
            <content:encoded><![CDATA[<h4>A quick note on our investment in <a href="http://asktia.com/">Tia</a>.</h4><p><strong>In a </strong><a href="http://www.ondeckdaily.com/post/202"><strong>recent discussion</strong></a><strong> related to the </strong><a href="https://medium.com/@ask_tia/backed-by-2-5-million-tia-gets-one-step-closer-to-redesigning-womens-health-15665b66b344"><strong>announcement of Tia’s Seed round</strong></a><strong>, </strong><a href="https://medium.com/u/abcbbcac8a09"><strong>Erik Torenberg</strong></a><strong> asked:</strong></p><blockquote>Investors in the company [@mhdempsey, @hunterwalk, @soopa, @soleio] can you talk about how you got there from an investment POV, and what, if any, initial reservations you had to get over? Consumer healthcare is hard.</blockquote><p><strong>My answer, taken directly from the On Deck Daily conversation is below.</strong></p><p>So in standard VC fashion I have to say, first, it was the team. I’m not going to go into the overly long blog post about how amazing <a href="https://medium.com/u/ce5185024ce9">Carolyn Witte</a> and <a href="https://medium.com/u/99f95688ef57">Felicity Yost</a> are (though I could), but both because of who they are as people, as well as their previous backgrounds, they are pretty uniquely suited to tackle a product and vision like this.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*KqzzlPNtv_9WP6EeAnqhGg.png" /></figure><p>I came across Tia after DMing someone on twitter who was asking about women’s healthcare apps. After downloading the app, I became kind of obsessed with the product experience and used it a bunch. I then read Carolyn’s <a href="https://medium.com/@carolynwitte/my-next-chapter-hacking-womens-health-278f3ea4d3e2">medium post</a> and got connected via my friend <a href="https://medium.com/u/99d9fe3e3a41">John Melas-Kyriazi</a>.</p><p>Tia actually intersected at two theses we’ve thought a lot about at Compound.</p><ol><li>We’ve been thinking about personalized women’s healthcare for a few years now but frankly I have always been looking at it from the lens of family planning as a powerful initial wedge (more <a href="https://medium.com/@mhdempsey/whats-next-for-egg-freezing-69072524788b">here</a>). I guess that might have been an initial reservation?</li><li>I’ve always personally found consumer “chat bots” frustrating, but at Compound, we’ve seen the text interface work very well in healthcare, and we’ve made multiple investments in the space across <a href="https://onward.org/">addiction</a>, <a href="https://www.talkspace.com/">psychotherapy</a>, and <a href="http://cleargenetics.com/">genetic counseling</a>.</li></ol><h4><strong>Thesis 1:</strong></h4><p>Why I generally like to be so thesis-driven is so that I can move efficiently within categories I’ve gone deep on, and also so that I can recognize when a founder/company shatters my thesis. Tia did this. Through the process, Carolyn expanded my beliefs around the idea that you could onboard and grow with your users across various health (and life) journeys and that being the trusted friend/voice/filter for the top of the funnel is a super compelling place to sit in the ecosystem vs. where I was previously so obsessed with focusing in on a single event like family planning and creating a better/more efficient experience.</p><h4><strong>This leads us to thesis 2:</strong></h4><p>Spending time understanding how users were behaving as they talked to their “wingwoman”, Tia, was INCREDIBLE. It’s still early, but the wide variety and depth of engagement from the user base, ranging from traditional healthcare questions to the things women were shy to ask to just day to day conversations I’d have with my friends or my female friends have together, was fascinating to me. I think not everyone always has <em>*one person*</em> that they can turn to across so many different thoughts/problems/experiences, and that idea is really powerful and showed through in the user behavior.</p><p>And last but not least, in addition to the general power of Tia, we also looked around at the increasingly fragmented female-focused healthcare and CPG startup market and think that a major distribution channel like Tia was very compelling from a business model use-case to both sides of the market.</p><p><strong>If you’re interested in learning more, </strong><a href="https://itunes.apple.com/us/app/ask-tia-your-personal-womens-health-assistant/id1184002073?ls=1&amp;mt=8"><strong>download Tia on iOS</strong></a><strong> now.</strong></p><p><strong>Also, </strong><a href="http://www.asktia.com/jobs/"><strong>Tia is hiring across a variety of positions</strong></a><strong>.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5125413e54af" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to drive 10 billion miles in an autonomous vehicle]]></title>
            <link>https://medium.com/@mhdempsey/how-to-drive-10-billion-miles-in-an-autonomous-vehicle-7c9b81c1ddac?source=rss-6e6d7bd0dfb6------2</link>
            <guid isPermaLink="false">https://medium.com/p/7c9b81c1ddac</guid>
            <category><![CDATA[self-driving-cars]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[autonomous-cars]]></category>
            <category><![CDATA[simulation]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Michael Dempsey]]></dc:creator>
            <pubDate>Wed, 31 May 2017 15:29:02 GMT</pubDate>
            <atom:updated>2017-05-31T15:29:02.457Z</atom:updated>
            <content:encoded><![CDATA[<h4>Google, Tesla, Zoox, and many more plan to drive billions of miles autonomously with simulation</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aAJq0dByoCpx38eOjNxuRQ.png" /></figure><p>There are two major bottlenecks for building an autonomous vehicle (AV) today: <strong>Generating supervised data</strong> and<strong> training/testing your model</strong>.</p><p>AV engineers train their models by feeding them with enough data to sufficiently react to multiple scenarios. These range from sunlight hitting sensors at different angles to obstacles flying in front of the car to strange external driver behavior, and much more.</p><p>The problem is it isn’t easy or safe to replicate many scenarios in real-world environments.</p><p>A <a href="http://www.rand.org/content/dam/rand/pubs/research_reports/RR1400/RR1478/RAND_RR1478.pdf">report by RAND</a> found that <strong>AVs would have to be driven hundreds of millions of miles and sometimes billions of miles to demonstrate acceptable reliability.</strong></p><blockquote>“Under even aggressive testing assumptions, existing fleets would take tens and sometimes hundreds of years to drive these miles — an impossible proposition if the aim is to demonstrate their performance prior to releasing them on the roads for consumer use.”</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*oMSr0pu58ZyGJpQw." /><figcaption>Source: <a href="https://www.rand.org/content/dam/rand/pubs/research_reports/RR1400/RR1478/RAND_RR1478.pdf">Driving to Safety</a> by RAND</figcaption></figure><h3>Then how do we get to reliable autonomy?</h3><ul><li><strong>We build within </strong><a href="https://medium.com/frontier-tech/h-edge-cases-in-autonomous-vehicles-37f75fe63b56"><strong>hedge cases</strong></a><strong>, rolling out AVs in constrained environments. </strong>Companies like Uber, Lyft, or Zoox roll out on a city by city basis, operate within constrained borders, lowering the technical barrier of autonomy. This could work for the Ubers of the world, but traditional OEMs will focus on bridging the gap to a shared mobility future by continually releasing advance autonomy features in their vehicles (i.e. Tesla Autopilot), which will be expected by consumers to be less constrained than a single-ride vehicle.</li><li><strong>We bypass the traditional technological approaches of today </strong>which require large amounts of data, and instead build models able to reason and learn with little data. Gary Marcus, <a href="https://techcrunch.com/2016/12/05/uber-acquires-geometric-intelligence-to-create-an-ai-lab/">who’s company was acquired by Uber last year,</a> has spent <a href="https://www.technologyreview.com/s/544606/can-this-man-make-ai-more-human/">years researching this</a>, however this type of learning hasn’t materialized in AVs yet.</li></ul><h4>And then there’s simulation</h4><p>Simulations (from software to hardware to environment), when constructed properly, allow companies to train and test their models on:</p><ul><li><strong>A variety of world scenarios</strong> including traffic, driver behavior, weather, road environment, and more.</li><li><strong>Multiple sensor suites/arrays.</strong> How many LiDARs do I need? Cameras? Radars? Where on the car should they be? Which model hardware do I use?</li><li><strong>With multiple, scalable random permutations. </strong>Without having to put a fleet of cars on the road with safety drivers in tow.</li></ul><h4>But we’re not totally there yet</h4><p>Today products such as <a href="https://vires.com/">Vires</a>, <a href="https://www.tassinternational.com/prescan">TaSS PreScan</a>, <a href="http://www.carsim.com/">CarSim</a>, <a href="http://www.oktal.fr/en/automotive/range-of-simulators/software">Oktal ScanNer</a>, <a href="http://gazebosim.org/">ROS Gazebo</a>, and others allow engineers to simulate sensors, conditions, mechanical builds, and more. While each have their benefits, they miss in areas crucial to simulation for AVs, including an oversimplification of existing sensor outputs, as well as a lack of complex understanding of how environments impact autonomous models.</p><p><a href="https://blogs.microsoft.com/next/2017/02/15/microsoft-shares-open-source-system-training-drones-gadgets-move-safely/#sm.0002udjh9f3pfha10ii19kai2vyrb">Microsoft</a> and <a href="https://github.com/udacity/self-driving-car-sim">Udacity</a> have also created their own simulators purposefully built for AVs/Robotics. However as the tech stack of AVs continues to progress, we have to think about how to implement, interpet, and test these fusions of hardware + software in a high-fidelity way.</p><h4>High-Fidelity Simulated Environments</h4><p>While simulating most sensor perception is hard, the increasingly prevalent sensor on all vehicles continues to be one that is simple to simulate, but difficult to train with.</p><p>Optical cameras are being relied upon to deliver vision as the undelivered promise of low-cost LiDARs and a <a href="https://www.driverless.id/news/2017-lidar-bottleneck-is-causing-modern-day-gold-rush-0177282/">shortage of higher-end units</a> make scalability difficult for OEMs and Tier 1s.</p><p>Training data from simulated cameras is only as good as the input, thus to properly test perception, engineers need photorealistic simulated environments. Building a complex photorealistic environment is incredibly costly and difficult, thus nobody has been able to build this for the purpose of training self driving cars.</p><p>That’s where people like Craig Quiter come in.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bR3vKWsf9GcYxzswwgLG6A.png" /><figcaption>An early email to Craig Quiter, one of the first to build an AV in Grand Theft Auto V</figcaption></figure><p>I met Craig almost a year ago as he was posting about something he built called DeepDrive. I later learned that he was one of the early engineers utilizing a game touting $137M of development costs to model real-world scenarios, with high-fidelity graphics…to autonomously drive cars.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FX4u2DCOLoIg%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DX4u2DCOLoIg&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FX4u2DCOLoIg%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/0030d936435ff7c571d9a02678f80de9/href">https://medium.com/media/0030d936435ff7c571d9a02678f80de9/href</a></iframe><p>A few months later, Craig joined a small startup called Uber, focusing on simulation.</p><p>A team at Princeton detailed the advantages of using GTA V in <a href="http://orfe.princeton.edu/~alaink/SmartDrivingCars/DeepLearning/GTAV_TRB_Final.pdf">their paper</a> quantifying the world as 100 square miles, 4M people, 262 vehicles, 1,167 different living beings, 14 weather conditions, and over 70,000 dynamic road segments in both urban, rural, dessert, and woodland environments.</p><h3>Are simulated miles <em>really</em> useful?</h3><p>There are two conflicting trains of thought about the utility of simulation for AVs.</p><h4>Yes.</h4><p><strong>Simulation can be used for testing rare cases and baseline data</strong>. Rare cases (non-scientific term) are scenarios that are difficult to recreate or random enough that we can generate them with enough permutations, but aren’t so random that with billions of miles simulated, they won’t show up. If we are on a journey to 99% reliability, simulation could theoretically get us a large portion of the way there, with some future iteration of AI/ML allowing our models to react to extreme edge situations without specific prior data.</p><p>Leaving rare cases out of the mix, <strong>simulation is great for building a base dataset and continually testing it in a scaleable way. </strong>Want millions of miles of highway auto-generated across multiple unique permutations? <em>Let’s run those on the cloud for you.</em> This is important as a wave of new OEMs, Tier 1s, and startups seek to build out autonomous systems.</p><p>Related reading: <a href="https://arxiv.org/pdf/1608.01745.pdf">Play and Learn: Using Video Games to Train Computer Vision Models</a></p><h4>No.</h4><p><strong>The counterpoint is that simulated environments aren’t good enough to <em>properly</em> train a model.</strong> Garbage in, garbage out. Often this is a statement of the environment-to-vehicle interaction not ever being replicated as it would in the real-world, or the graphical fidelity being oversimplified.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qI2Y4qiotwEwxatIfgTdJQ.png" /><figcaption>A transfer or virtual image input to realistic from the paper <a href="https://arxiv.org/pdf/1704.03952.pdf">Virtual to Real Reinforcement Learning for Autonomous Driving</a></figcaption></figure><p>To help counter some of the issues around quality of data, <a href="https://arxiv.org/pdf/1704.03952.pdf">researchers are testing the ability to transfer virtual image inputs into realistic ones</a> for improved model training.</p><p>While many government agencies aren’t yet willing to count simulated miles as part of the required # of miles driven autonomously, this could change as regulations around testing become more defined. Google <a href="https://www.theguardian.com/technology/2014/aug/21/google-test-self-driving-cars-virtual-world-matrix">has previously lobbied</a> for this.</p><h3>Simulation Is Necessary</h3><p>I’m of the belief that simulation is valuable if the fidelity is high enough. Admittedly, simulation probably won’t solve the holy grail of the final 1% of autonomy, but if AVs are going to be realistically implemented, the time to 90% competency could be shortened, and maybe even the next 5–8% could allow models to more appropriately recognize or react to a wider range of scenarios.</p><p>And many companies would agree with me.</p><p>Some of the top companies in the world have managed to build significant IP in their simulation tech, with <a href="https://medium.com/waymo/reliving-the-past-how-these-data-centers-drive-us-three-million-miles-each-day-49a8695e8c75">Google driving 3 million+ simulated miles per day</a>, while others including <a href="https://www.tesla.com/careers/job/simulation-engineer-44535">Tesla</a>, <a href="https://jobs.lever.co/zoox/d3479e79-90b0-4416-a1ce-ce18f73bd882">Zoox</a>, <a href="http://comma.ai/positions.html">Comma.ai</a>, <a href="https://www.drive.ai/careers?gh_jid=522207&amp;gh_src=4m1aos1">Drive.ai</a>, and <a href="https://aurora.bamboohr.com/jobs/view.php?id=16">Aurora Innovation</a>, all are actively hiring simulation engineers.</p><h3>Outside of Autonomous Vehicles</h3><p>The use-cases for properly simulated environments expand far outside of AVs. While we can understand how AVs perceive the world around them, we also can theoretically better understand traffic, routing, driver behavior, and even pedestrian behavior.</p><p>Then we can take another step back. With enough specific modules and dynamic beings in a simulated environment, we can better understand how all types of robots (cars, delivery robots, social robots, etc.) will interact autonomously with our real (and digital) world.</p><p>Companies like <a href="http://improbable.io/">Improbable</a> have taken aim at this large-scale world simulation problem, and <a href="https://techcrunch.com/2017/05/11/improbable-grabs-502m-led-by-softbank-at-a-1b-valuation-for-its-virtual-world-spatial-os/">investors have recognized</a> the wide-reaching value in being the architect to our future simulated worlds recently with a $502M financing to the UK-based company.</p><p>We’re just starting to scrape the surface of simulation for AVs. Multiple companies are building internally, a few have startups have begun building standalone software, and with the amount of research being done, I expect a variety of new entrants into the market. Those that figure out simulation could be the early leaders or enablers in better understanding the first, second, and third order effects for enabling autonomy.</p><p><strong>If you’re working on something in the simulation space or have any thoughts please don’t hesitate to reach out via </strong><a href="http://twitter.com/mhdempsey"><strong>twitter</strong></a><strong> or </strong><a href="/service/mailto:mike@compound.vc"><strong>email</strong></a><strong>.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7c9b81c1ddac" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>