<?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 Jerry Chi on Medium]]></title>
        <description><![CDATA[Stories by Jerry Chi on Medium]]></description>
        <link>https://medium.com/@peacej2?source=rss-8faf1fc6cd9c------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*47IDvO6SJKI-MCqu.jpg</url>
            <title>Stories by Jerry Chi on Medium</title>
            <link>https://medium.com/@peacej2?source=rss-8faf1fc6cd9c------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 00:36:07 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@peacej2/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[Why creativity flourishes in Japan]]></title>
            <link>https://peacej2.medium.com/why-creativity-flourishes-in-japan-398afdb68599?source=rss-8faf1fc6cd9c------2</link>
            <guid isPermaLink="false">https://medium.com/p/398afdb68599</guid>
            <category><![CDATA[nft]]></category>
            <category><![CDATA[art]]></category>
            <category><![CDATA[creativity]]></category>
            <category><![CDATA[metaverse]]></category>
            <category><![CDATA[japan]]></category>
            <dc:creator><![CDATA[Jerry Chi]]></dc:creator>
            <pubDate>Mon, 07 Nov 2022 06:42:38 GMT</pubDate>
            <atom:updated>2022-11-07T06:42:38.659Z</atom:updated>
            <content:encoded><![CDATA[<p>Over the last 2 decades, the creativity I’ve come across in Japan repeatedly induces awe and wonder and tickles my brain. Manga, anime, games, music, movies, shows, ads, architecture, paintings, installations, fashion, funky commercial products… one favorite of mine is art by Takashi Murakami.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*pfpqA2giNMTW0Zm9.jpg" /><figcaption>(Source: <a href="https://www.serpentinegalleries.org/whats-on/takashi-murakami-kaikai-kiki/">https://www.serpentinegalleries.org/whats-on/takashi-murakami-kaikai-kiki/</a>)</figcaption></figure><p>There are several reasons why creativity flourishes:</p><h4>Lots of people with spare money and time</h4><p>Japan has the 3rd highest GDP and, among OECD countries, the 3rd highest population (used to be 2nd in both). With efficient and reliable infrastructure, healthy food, and minimal crime, people can focus on their creative work.</p><h4>Heavy exposure to both Eastern and Western culture and arts</h4><p>Japan has been influenced heavily by China, US, Europe, etc. over centuries. Not to mention Japan itself has had a vibrant art/culture scene since over 1000 years ago. The Western influence accelerated due to the Meiji Restoration of 1868 and the US occupation post-WW2. Even today, there are a huge number of e.g. French culture fanatics/hobbyists in Japan. The culture of appreciation for the arts (esp. Western and Japanese) is deeply ingrained now, as can be seen in the <a href="https://www.statista.com/statistics/1201825/top-countries-by-number-of-museums-worldwide/">numerous excellent museums</a>.</p><p>It’s also notable that the Japanese language is extremely flexible/expressive and includes almost all Chinese concepts/characters while also using a huge number of loan words from English, French, German, etc. In practice, 4 writing systems are usually used simultaneously (kanji, hiragana, katakana, and the English alphabet).</p><h4>Metropolitan areas as creative hubs</h4><p>Greater Tokyo is the <a href="https://en.wikipedia.org/wiki/List_of_largest_cities">world’s largest metropolitan area, with ~37 million people</a>. (Osaka is ranked 10th). By riding in one of the world’s best public transport systems, any of those 37M people can easily meet up and share their ideas, paintings, clothes, robots, etc. whether at a huge event like Tokyo Game Show or a tiny informal meetup over beers. Even if you are into some weird, long-tail, obscure topic, such as crafting 19th century French maid costumes, you can probably find at least a few people among that 37M to explore that topic with. Of course, video conferencing and COVID affect how much this matters, but I’d say physical interaction still matters hugely in terms of communication, exchanging ideas, and creating meaningful outlets for creative expression.</p><h4>Manga publishing industry practices</h4><p>In Japan, it has been possible since about 50 years ago for a single creator with no track record and no money to test his/her creative ideas on a large population.</p><p>The typical example of this is an aspiring manga artist submitting a couple pages to the editors of a weekly/monthly manga magazine such as Shonen Jump, which can choose to include those pages (along with manga by other artsits) in the weekly issue. The marginal cost is almost zero (there’s just opportunity cost of not including another artist’s work). If reader feedback is good, the next chapter can be included the next week as well, and perhaps the author could be on their way to fame and fortune. (This process was depicted/dramatized in <a href="https://en.wikipedia.org/wiki/Bakuman">BAKUMAN</a>, a manga/movie about publishing manga.) Then, if the manga really takes off, companies like Bandai Namco skillfully kick the monetization into high gear with shows/movies, toys, clothes, games, etc. and all that money feeds back into making the creative industry stronger.</p><p><strong>This rapid, low-cost trial and error of creative ideas on a huge readership is the equivalent of having tons of people throw tons of ideas on the wall and quickly seeing which ones stick</strong>. Contrast this to most countries, where it often takes a creator’s track record and maybe some special connections (to e.g. a publishing company or a movie studio) and a significant commitment of budget to get creative material in front of a large audience.</p><p>The culture of manga/anime has become deeply ingrained over decades, so it’s normal for the average adult in Japan to enjoy such content, whereas in some countries, most cartoons are considered to be mostly for kids, limiting the industry potential.</p><p>Granted, the internet has provided new avenues of reaching audiences, so perhaps the Japanese advantage is not as big anymore, but I’d argue that it’s hard for a given piece of online content to gain traction/virality in the vast ocean of content without creative industry pros supporting it, so the Japanese manga/anime/IP industrial complex is still an important differentiator for Japan.</p><p>Speaking of the internet, it has been a major way for the fandom of Japanese manga/anime fandom to blossom globally, and there has been an explosion of Japan-inspired cartoons, games, fan art, <a href="https://nonfungible.com/news/art/nfts-japanese-culture-rising-diverse-community">NFTs</a>, etc. made by non-Japanese people.</p><h4>Openness to metaverses / NFTs</h4><p>In contrast to some countries, where there is considerable pushback from regulators and gaming communities on the use of NFTs, people in Japan are mostly open to the use of NFTs, and the Japanese government is <a href="https://www.metaverse-style.com/english/8033">increasingly pro-NFT / pro-Web3</a> as well. This unlocks new ways for creators to be rewarded and engage with communities. The idea of virtual non-human characters and virtual worlds is also familiar to most Japanese people, partially due to the gaming/anime/otaku culture and mainstream content such as the excellent Mamoru Hosoda metaverse movies <a href="https://www.youtube.com/watch?v=UbmdYasVDDg">Summer Wars</a> and <a href="https://www.youtube.com/watch?v=ChneY1MSVFw">Belle</a>. Also, virtual Youtubers have taken off in Japan like nowhere else (the most famous one, <a href="https://www.youtube.com/channel/UC4YaOt1yT-ZeyB0OmxHgolA">Kizuna AI</a>, has over 3M subscribers). Probably this metaverse/NFT-friendly environment is one reason why metaverse giant <a href="https://techcrunch.com/2022/08/25/animoca-brands-japan-funding/?guccounter=1">Animoca Brands was able to raise $45M at a $500M valuation</a> for its Japan branch.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/970/0*DsNvg7Z4Rhqgk8CN" /><figcaption>(screenshot from the movie Belle)</figcaption></figure><h4>“Shokunin” spirit</h4><p>Shokunin (職人) means “artisan” or “craftsman”. To embody the shokunin spirit as a creator is to take great pride in one’s work and mastery of it, even the tiny, menial details. More explanation <a href="https://medium.com/@richbltn/shokunin-spirit-ac72a49acffd">here</a>. Although this can arguably be in some ways detrimental to creativity (e.g. spend years perfecting how to cut sushi perfectly à la <a href="https://en.wikipedia.org/wiki/Jiro_Dreams_of_Sushi">Jiro Dreams of Sushi</a> rather than experimenting with new sushi dishes), in some other ways the shokunin attitude and mastery can lead to unique creative works that are quite hard for replicate due to the unique degree and combination of skills required. For example, yosegi-zaiku (寄木細工), a type of Japanese marquetry, which is basically combining pieces of wood, sounds simple but damn I bet it’s hard to execute:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/900/0*pqcOcteuzv7O5zzR.jpg" /><figcaption>(Source: <a href="https://www.yosegi.jp/">https://www.yosegi.jp/</a>)</figcaption></figure><h4>Stay tuned</h4><p>Next I will post about the past and future of creative/generative AI in Japan. Feel free to subscribe or follow me at <a href="https://twitter.com/peacej">https://twitter.com/peacej</a> ♫</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=398afdb68599" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Trying out AI music generation with Dance Diffusion]]></title>
            <link>https://peacej2.medium.com/trying-out-ai-music-generation-with-dance-diffusion-b58743679a8d?source=rss-8faf1fc6cd9c------2</link>
            <guid isPermaLink="false">https://medium.com/p/b58743679a8d</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[music]]></category>
            <category><![CDATA[stable-diffusion]]></category>
            <category><![CDATA[generative-music]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Jerry Chi]]></dc:creator>
            <pubDate>Thu, 27 Oct 2022 08:06:56 GMT</pubDate>
            <atom:updated>2022-10-27T08:18:58.866Z</atom:updated>
            <content:encoded><![CDATA[<p>In late September the <strong>Dance Diffusion music generation model</strong> (beta) was released by <a href="https://twitter.com/harmonai_org"><strong>Harmonai</strong></a>. Harmonai, backed by Stability AI (which also backed Stable Diffusion), is a deep learning research lab focused on creating open-source generative audio models, supporting researchers and developers with compute grants and community, and providing musicians and audio engineers with new creative tools.</p><p>As an oversimplication, Dance Diffusion is kind of like the audio version of Stable Diffusion or DALLE-2, except simpler. Essentially, one adds random noise to an audio track and then a model is trained to remove that noise. A <a href="https://wandb.ai/wandb_gen/audio/reports/A-Gentle-Introduction-to-Dance-Diffusion--VmlldzoyNjg1Mzky">gentle intro to Dance Diffusion is on the awesome Weights &amp; Biases website</a>.</p><p>Dance Diffusion doesn’t operate on MIDI files (i.e. the digital version of sheet music) which are used by most generative methods, it operates directly on raw audio files, so that allows for almost any sound to be created.</p><p>I think it’s a really cool that a method used for image generation could be applied in the same way to audio. It’s another testament to how fundamental deep learning knowledge can be applied to so many fields.</p><p>Unfortunately, it’s still early days, and it’s tough to generate good-sounding music with this model. Here’s one clip I made by interpolating between the noised representations of two 8-second clips from Harder, Better, Faster, Stronger by Daft Punk (see the awesome <a href="https://www.youtube.com/watch?v=gAjR4_CbPpQ">original music vid on Youtube</a>), using the Dance Diffusion model trained on glitch music from <a href="http://glitch.cool/">glitch.cool</a>.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fw.soundcloud.com%2Fplayer%2F%3Furl%3Dhttps%253A%252F%252Fapi.soundcloud.com%252Ftracks%252F1370899315%26show_artwork%3Dtrue&amp;display_name=SoundCloud&amp;url=https%3A%2F%2Fsoundcloud.com%2Fjerry-chi-99480064%2Fdance-diffusion-interpolation-of-daft-punk%3Fsi%3Df5e16256d4ac49b3a15c53c5f00177de%26utm_source%3Dclipboard%26utm_medium%3Dtext%26utm_campaign%3Dsocial_sharing&amp;image=https%3A%2F%2Fsoundcloud.com%2Fimages%2Ffb_placeholder.png&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=soundcloud" width="800" height="166" frameborder="0" scrolling="no"><a href="https://medium.com/media/d925fe085efa8648167517175971d563/href">https://medium.com/media/d925fe085efa8648167517175971d563/href</a></iframe><p>Someone else was able to make a much better clip, but they used several tools in addition to Dance Diffusion (“gpt-j (lyrics) -&gt; jukebox (w/modded sampler) -&gt; demucs -&gt; pymixconsole -&gt; dance diffusion -&gt; mastering”):</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fw.soundcloud.com%2Fplayer%2F%3Furl%3Dhttps%253A%252F%252Fapi.soundcloud.com%252Ftracks%252F1368508270%26show_artwork%3Dtrue&amp;display_name=SoundCloud&amp;url=https%3A%2F%2Fsoundcloud.com%2Fbaltigor%2Flion-0186-12%3Fsi%3D9c312290ac6a4a1fb1fc3a0c4539c549%26utm_source%3Dclipboard%26utm_medium%3Dtext%26utm_campaign%3Dsocial_sharing&amp;image=https%3A%2F%2Fi1.sndcdn.com%2Fartworks-c5ryX3IIQyjKtapD-6lbXdw-t500x500.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=soundcloud" width="800" height="166" frameborder="0" scrolling="no"><a href="https://medium.com/media/052bf3f9bc1d71c2b9c234e630523400/href">https://medium.com/media/052bf3f9bc1d71c2b9c234e630523400/href</a></iframe><p>This Dance Diffusion model is much smaller and simpler than Stable Diffusion, and it’s still the early beta version, so I think that explains why it’s not that good yet. Another reason is that there’s not that much non-copyrighted music out there in the public to use as training data.</p><p>Anyway, stay tuned because I will share more about AI music generation!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/1*5JAVkkpwHTLkcp0UeYCYtw.png" /><figcaption>Stable Diffusion prompt: “Pop art of daft punk at a vaporwave neon futuristic cyberpunk Tokyo bustling street at night cyberart by liam wong, rendered in octane, 3d render, trending on cgsociety, blender 3d”</figcaption></figure><p>This post was cross-posted to my personal website at <a href="https://jerrychi.com/">https://jerrychi.com/</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b58743679a8d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Positive trends in Japan]]></title>
            <link>https://peacej2.medium.com/positive-trends-in-japan-7e2c4259d5e8?source=rss-8faf1fc6cd9c------2</link>
            <guid isPermaLink="false">https://medium.com/p/7e2c4259d5e8</guid>
            <category><![CDATA[tokyo]]></category>
            <category><![CDATA[economy]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[japan]]></category>
            <category><![CDATA[fintech]]></category>
            <dc:creator><![CDATA[Jerry Chi]]></dc:creator>
            <pubDate>Sun, 23 Oct 2022 14:09:27 GMT</pubDate>
            <atom:updated>2022-10-23T14:09:27.643Z</atom:updated>
            <content:encoded><![CDATA[<p>There’s a neverending onslaught of bad news of negative trends in Japan. Depreciating yen, aging population, ballooning debt, stagnant GDP, lower ranking in digital competitiveness, dwindling fish population, declining music industry, political scandals… it seems everything is worsening.</p><p>But the news media tends to give a negative bias to everything. Let’s talk about the positives for a moment. There are already many articles talking about the positive <strong>current state </strong>of Japan (great healthcare, delicious food, low crime rate, good public transport infrastructure). But what is actually <strong>improving</strong> over time?</p><h3>Startups and venture capital</h3><p>Startup funding and top talent working at startups has increased dramatically over the past few years (ignoring 2022 for now). Almost a 10x increase in 10 years! (granted, still tiny compared to the US, where startup funding was $330B in 2021).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Zr3oaH9KGXbW5mTC_If8wQ.png" /><figcaption>Fundraising by Japan startups (units: 100M yen)</figcaption></figure><p>It’s also cool how various new government initiatives are trying to help (e.g. startup visas, zero-interest loans for foreign entrepreneurs in Tokyo, Shibuya Startup Support Office with bilingual support). Stuff like the Google Launchpad Accelerator (at which I mentored), Google for Startups Campus Tokyo, Coral Capital and its accelerator Coral Reef also opened over the last couple of years.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KXnFqNxqacvpY-l9li1FOg.png" /><figcaption><a href="https://shibuya-startup-support.jp/">https://shibuya-startup-support.jp/</a></figcaption></figure><h3>More solar and nuclear power generation</h3><p>Nuclear was reduced after the Fukushima incident, but finally it’s on the uptrend again, increasing energy independence and reducing carbon emissions (hopefully the risks will be better managed…)</p><h3>More and more awesome buildings</h3><p>As if Tokyo didn’t already have enough super nice and cool-looking buildings with interesting stores and yummy food, more and better ones keep getting built. Especially in Shibuya (PARCO, Shibuya Scramble Square, MIYASHITA PARK mall, etc.) Don’t know enough to comment on non-Tokyo cities, unfortunately.</p><h3>Even better infrastructure</h3><p>The “linear motor car” (magnetic levitation) trains are being built and the Tokyo-Nagoya line will open in 2027 with a speed of 500 km/h, 2x as fast as current bullet trains.</p><h3>Better games and better anime.</h3><p>The technology keeps getting better and I am repeatedly wowed by the creativity and execution quality. Check out Belle, Attack on Titan, and Cyberpunk: Edgerunners.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*pzVfYKs49W0Zuvab.png" /><figcaption>Night City in Cyberpunk: Edgerunners</figcaption></figure><h3>Crypto / web3</h3><ul><li>Regulation: several Japanese government agencies and lawmakers are working hard to improve regulations to make Japan an attractive place for web3/crypto startups and companies. I spoke with a lawmaker and read their policy whitepaper, and the proposed policy improvements are actually quite well thought-out. (<a href="https://www.metaverse-style.com/english/8033">article</a>, <a href="https://www.taira-m.jp/Japan%27s%20NFT%20Whitepaper_E_050122.pdf">white paper in English</a>)</li><li>The general crypto scene: startups building cool stuff, conferences, parties, events, online communities… the vibe is great, and a much is growing, even during the bear market.</li><li>Lots of companies, big and small, new and old, with plans for blockchain gaming, the metaverse, NFTs….e.g. check out the <a href="https://www.oasys.games/">list of companies</a> involved with the Oasys blockchain started by doublejump.tokyo (disclosure: I’m a small investor)</li></ul><h3>Fintech</h3><p>Not limited to crypto, lots of financial tech improvements have happened. Japan is often called a “cash-based society”, but smartphone-based payments (NFC-based, QR code-based, etc.) such as PayPay and LINE Pay have increased dramatically, as well as credit card use. The percentage of cashless payments increased from 13% in 2010 to 33% in 2021 (granted, still lower than other countries). Adoption of tools such as Money Forward ME (personal finance app like mint.com in the US) or Freee (SaaS for small business accounting etc.) have exploded and saved countless hours for people dealing with paper before. Also, the national government introduced the My Number card for all residents with a private key on a secure chip inside it, and this can be used for secure remote tax filing, for obtaining government documents at any convenience store, and as a health insurance card. This is the same security concept as buying a Ledger or Trezor hardware device with a private key to access cryptocurrency.</p><h3>Machine learning / AI</h3><p>There’s a strong uptrend in the number of startups and patents related to machine learning as well as revenue in the AI software/services category. <a href="https://it.impress.co.jp/articles/-/23212">In 2021, the AI systems market grew 26% YoY to 277B yen</a>.</p><p>Although Japan is still far behind the US etc., there are unique opportunities for AI that Japan can especially excel at, e.g. AI for manufacturing, fishing, and elderly care.</p><p>One area I’m particularly excited about is creative AI. There are already <a href="https://docs.google.com/document/d/1-P58p-5yx8OXE0_0RN_D5Zg3lGqYQ88xtxOXClX5c_I/edit#heading=h.o2zonez60r1r">a bunch of Japanese companies </a>(mostly startups) doing impressive work on this, e.g. Qosmo which works on music AI and Rinna (spun out from Microsoft) which works on chatbots and more recently text-to-image (e.g. they fine-tuned Stable Diffusion to make Japanese Stable Diffusion).</p><h3>Wrap up</h3><p>Not all hope is lost in Japan! Lots of stuff happening to be excited about now and in the future. And there’s a lot I didn’t cover in this article.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7e2c4259d5e8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How Stable Diffusion (Latent Diffusion) works]]></title>
            <link>https://peacej2.medium.com/how-stable-diffusion-latent-diffusion-works-991bec22988f?source=rss-8faf1fc6cd9c------2</link>
            <guid isPermaLink="false">https://medium.com/p/991bec22988f</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[generative-art]]></category>
            <category><![CDATA[ai-art]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Jerry Chi]]></dc:creator>
            <pubDate>Thu, 25 Aug 2022 06:06:04 GMT</pubDate>
            <atom:updated>2022-10-07T04:45:13.448Z</atom:updated>
            <content:encoded><![CDATA[<p>The <a href="https://huggingface.co/CompVis/stable-diffusion">Stable Diffusion model</a>, just released a few days ago, is all the rage right now, with tons of people generating all sorts of amazing high-quality images, sometimes on par with or even better than OpenAI’s DALL-E 2.</p><p>UPDATE 2022–10–07: a new, more beginner-friendly article on Stable Diffusion was posted at <a href="https://jalammar.github.io/illustrated-stable-diffusion/"><br>https://jalammar.github.io/illustrated-stable-diffusion/</a> .</p><p>But how does it actually work from a technical perspective?</p><p>Basically, it uses a variational autoencoder (VAE) combined with a denoising diffusion model. The key idea is that using diffusion models in pixel space (the raw image) is not the most efficient approach, since there are many barely perceptible small details that are not efficiently learned by a diffusion model. Rather, we can use a VAE to map images into a latent space (a form of compression), and then train the diffusion model in the (much smaller) latent space of images.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/578/0*iltP4V3p0cGAkL_O" /><figcaption>Source: <a href="https://huggingface.co/blog/stable_diffusion">https://huggingface.co/blog/stable_diffusion</a></figcaption></figure><p>Stable Diffusion is <a href="https://github.com/CompVis/stable-diffusion">basically a special case / specific configuration of Latent Diffusion</a>. A lot of effort went into making it very high-quality and easy to use for the masses.</p><p>The above explanation barely scratches the surface. <a href="https://docs.google.com/document/d/1x4iHe9mdyqpuINRN2EYMuG6_0JSBoNnjtDdSte18Ugc/edit#"><strong>For more in-depth details on Stable Diffusion / Latent Diffusion, please see this google doc I made.</strong></a></p><p>Also, I compiled various AI art resources (both technical and non-technical) at <a href="https://tinyurl.com/creative-ai-links">https://tinyurl.com/creative-ai-links</a>.</p><p>Of course, we need to wrap up with some actual AI art :) The below image was generated using Stable Diffusion at <a href="https://beta.dreamstudio.ai/">https://beta.dreamstudio.ai/</a> with the prompt “Character portrait of a graceful and pretty Korean princess with gorgeous detailed eyes and flowing hair, fantasy setting, color page, tankobon, 4k, tone mapping, doll, akihiko yoshida, james jean, andrei riabovitchev, marc simonetti, yoshitaka amano”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ePNHU-hG80IQrfE8m7J-1w.jpeg" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=991bec22988f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Easy blog migration from Medium to your own site using Python]]></title>
            <link>https://peacej2.medium.com/easy-blog-migration-from-medium-to-your-own-site-using-python-894d617f6ff3?source=rss-8faf1fc6cd9c------2</link>
            <guid isPermaLink="false">https://medium.com/p/894d617f6ff3</guid>
            <category><![CDATA[static-site-generator]]></category>
            <category><![CDATA[web-development]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[web-framework]]></category>
            <category><![CDATA[blogging]]></category>
            <dc:creator><![CDATA[Jerry Chi]]></dc:creator>
            <pubDate>Wed, 17 Aug 2022 05:30:16 GMT</pubDate>
            <atom:updated>2022-10-18T01:36:23.185Z</atom:updated>
            <content:encoded><![CDATA[<p>Recently I migrated my existing articles from Medium to my own new website (<a href="https://jerrychi.com/">https://jerrychi.com</a>) in a <strong>quick, low-code way</strong>. I wanted to have both blog articles and permanent pages (e.g. “About Me” page) and <strong>high customizability while spending minimal effort and money</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/770/1*IKMuTOm9TULLsIVPxuaDWw.png" /><figcaption>snapshot of jerrychi.com</figcaption></figure><p>You can do this too! Here’s how:</p><p>Register a domain name. I used <a href="https://domains.google/">https://domains.google/</a> to register <a href="https://jerrychi.com/">jerrychi.com</a>, paying about $12 per year. This is the only thing I’m paying for.</p><p>Install required Python libraries including <a href="https://getpelican.com/">Pelican</a> (the Python-based static site generator for creating personal blogs/websites) on your local laptop. E.g. pip install -r requirements.txt (see my <a href="https://github.com/peacej/blog6/blob/main/requirements.txt">requirements.txt</a>). I used Python 3.8.6 in a <a href="https://conda.io/projects/conda/en/latest/user-guide/concepts/environments.html">conda environment</a> but other setups should work too.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/193/1*F-NLYAtzFjXLDYa4PLzXOw.png" /><figcaption>Pelican logo</figcaption></figure><p>Auto-convert any Medium articles you choose to Markdown files (which are compatible with Pelican, which will then auto-convert them to HTML/Javascript) using the medium_to_markdown tool described at <a href="https://willkoehrsen.github.io/writing/markdown/converting-medium-posts-to-markdown-for-your-blog/">https://willkoehrsen.github.io/writing/markdown/converting-medium-posts-to-markdown-for-your-blog /</a> . I did <a href="https://github.com/peacej/blog6/blob/main/medium_to_markdown.py">slightly tweak his code</a> to e.g. filter out a few specific words. This doesn’t require any special access to your Medium account; it’s just scraping the text from the public internet.</p><p>Push your code to a new Github repo.</p><ul><li>You can host the website on <a href="https://docs.github.com/en/pages/getting-started-with-github-pages/about-github-pages">Github Pages</a> for free, up to 1 GB of storage (nothing beyond that even if you want to pay). If I ever need more than 1 GB, I might migrate to Heroku.</li><li>You can <a href="https://docs.github.com/en/pages/configuring-a-custom-domain-for-your-github-pages-site">use your own custom domain</a> (in my case <a href="https://jerrychi.com/">jerrychi.com</a>)</li></ul><p>We need a master or main git branch for all the code and then a gh-pages branch to store the actual website assets that are served when people visit your site. Thankfully one can just use a tool called ghp-import to automate pushing the only the needed to gh-pages branch. You can see <a href="https://github.com/peacej/blog6/blob/main/buildpush.sh">how I use this tool</a>. After pushing to github, usually changes are reflected in a few seconds on my website.</p><ul><li>Tip: style changes to CSS files etc. may require an “<a href="https://gsuitetips.com/tips/chrome/chrome:-empty-cache-and-hard-reload/#:~:text=Normal%20Reload%3A%20Uses%20Cached%20Data,be%20re%2Ddownloaded%20as%20required.">empty cache and hard reload</a>” in Chrome to be visually reflected on the website due to caching.</li></ul><p>Pelican also supports add-ons, themes, etc.</p><ul><li>See the <a href="https://github.com/getpelican/pelican-themes">pelican-themes repo</a> for instructions on themes.</li><li>I chose to use the <a href="https://github.com/alexandrevicenzi/Flex/">Flex theme</a> which includes many nifty features such a mobile-first responsive UI. I cloned this Flex repo directly rather than cloning from the pelican-themes repo which pins Flex at an older version.</li><li>I also edited the pelican-templates/Flex/static/stylesheet/style.less file to customize the font style (my first time dealing with the <a href="https://lesscss.org/">Less language</a> but it was easy to edit).</li></ul><p>That’s it! Now you have a nice shiny new personal website~</p><p>So, why did I choose Pelican? I also considered Nikola and Django:</p><ul><li>Nikola: the second-most popular Python static site generator after Pelican. I did initially try building with Nikola but got frustrated with the documentation being a bit unclearly worded and the lack of resources/answers when e.g. googling around for an issue. Pelican is more established and popular with more discussion/answers for potential issues.</li><li>Django: the most popular Python web backend framework. I did take the <a href="https://docs.djangoproject.com/en/4.1/intro/tutorial01/">Django official tutorial</a>, which was enjoyable and extremely well-written. However, Django is much more complicated than Pelican (which is natural since Django is a general-purpose framework for creating any sort of webpage) and I’d likely end up combining Django with a frontend framework like React (meaning one more framework to learn). The complexity and work required was overkill for needs: a simple personal website. I’ll come back to learn Django+React if I need a much fancier site.</li></ul><p>UPDATE 2022–10–18: it seems the above method no longer properly converts the image links to Markdown format. My preferred way around this is just manually editing the Markdown files after conversion to add back the images (takes me about 20 seconds). Another way is to use something like <a href="https://github.com/gautamdhameja/medium-2-md">https://github.com/gautamdhameja/medium-2-md</a> but that adds the hassle of manually exporting all your data from the Medium settings page each time you want to convert a new article.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=894d617f6ff3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Crypto and DeFi useful resources]]></title>
            <link>https://peacej2.medium.com/crypto-and-defi-useful-resources-114b7ed2eae?source=rss-8faf1fc6cd9c------2</link>
            <guid isPermaLink="false">https://medium.com/p/114b7ed2eae</guid>
            <category><![CDATA[yield-farming]]></category>
            <category><![CDATA[blockchain]]></category>
            <category><![CDATA[cryptocurrency]]></category>
            <category><![CDATA[defi]]></category>
            <category><![CDATA[ethereum]]></category>
            <dc:creator><![CDATA[Jerry Chi]]></dc:creator>
            <pubDate>Wed, 11 Aug 2021 05:59:01 GMT</pubDate>
            <atom:updated>2021-08-11T05:59:01.808Z</atom:updated>
            <content:encoded><![CDATA[<p>I’ve made a <a href="https://docs.google.com/document/d/1-dv-1XlpfPaT4ax6o1m65Ip8FmeXgKygueimGSWMu-c/edit#">google doc</a> listing out useful resources / tools / links for those interested in cryptocurrencies, blockchain, and DeFi (decentralized finance).</p><p>This doc is meant to be newbie-friendly but should also contain some useful tidbits for experienced people as well. I’ll keep updating it periodically.</p><p>See you down in the rabbit hole!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/728/1*tzo83NVKzuykBTAsCqU4Bw.png" /><figcaption>Accurate depiction of how blockchain works</figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=114b7ed2eae" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[My resolutions for 2021]]></title>
            <link>https://peacej2.medium.com/my-resolutions-for-2021-a3e5c641e9f5?source=rss-8faf1fc6cd9c------2</link>
            <guid isPermaLink="false">https://medium.com/p/a3e5c641e9f5</guid>
            <category><![CDATA[new-year-resolution]]></category>
            <category><![CDATA[self-improvement]]></category>
            <category><![CDATA[life-purpose]]></category>
            <dc:creator><![CDATA[Jerry Chi]]></dc:creator>
            <pubDate>Fri, 01 Jan 2021 08:11:39 GMT</pubDate>
            <atom:updated>2021-01-01T08:12:41.047Z</atom:updated>
            <content:encoded><![CDATA[<ol><li><strong>Better health</strong>. Including both physical and emotional/mental health. Meditate regularly.</li><li><strong>Study / play with machine learning. </strong>Including fun, non-work-related stuff, such as the <a href="https://www.deeplearning.ai/generative-adversarial-networks-specialization/">GANs specialization from deeplearning.ai</a>.</li><li><strong>Aim lower to aim higher</strong>. Most of my life has been busy with trying to hit short-term goals: get an A on the next test, reach the KPI goal for the next work project, etc. While that’s important, that may not be the best way to reach ambitious long-term life goals that are even more meaningful; in fact, when one is busy, it’s easy to forget to even think about what those goals should be. Forcing myself to be less busy and less ambitious in the short term may help me ideate and achieve something greater or more meaningful in the long term.</li><li><strong>Refine my life purpose. </strong>I know it has something to do with machine learning / AI and creativity, but I’m not sure what it is yet. One thing to do is the <a href="https://www.actualized.org/life-purpose-course">Ultimate Life Purpose course by actualized.org</a>, which I recently started.</li></ol><p>Let’s all have a fulfilling 2021! ♫</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/0*2KW5GQXngxUm54eI.jpg" /><figcaption>me trying to ideate something</figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a3e5c641e9f5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How machine learning may “save” humanity]]></title>
            <link>https://peacej2.medium.com/how-machine-learning-may-save-humanity-860026ae68f7?source=rss-8faf1fc6cd9c------2</link>
            <guid isPermaLink="false">https://medium.com/p/860026ae68f7</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[futurism]]></category>
            <category><![CDATA[extraterrestrial-life]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[compression]]></category>
            <dc:creator><![CDATA[Jerry Chi]]></dc:creator>
            <pubDate>Fri, 12 Jun 2020 11:34:28 GMT</pubDate>
            <atom:updated>2020-06-12T11:35:28.337Z</atom:updated>
            <content:encoded><![CDATA[<p>I think humanity will end eventually (whether it takes hundreds or thousands of years). So what’ll we do when the end is near? We’d want to leave our legacy; in other words, <strong>we’d want to save humanity’s knowledge and history as data and send it into space, hoping that aliens somewhere will receive it</strong>. That sounds a lot better than just disappearing into the void without a trace.</p><p>This begs the questions:</p><ol><li>what information to send</li><li>how to compress/represent the information/data</li><li>how to send it (including the <a href="https://arxiv.org/pdf/1101.4968.pdf">transmission strategy</a>: using lasers vs. omnidirectional radio transmitters, direction to point in, etc.)</li></ol><p>In fact, the approach for #2 (compression) affects #1 and #3. For example, we could decide to select more information for sending if we know it can be efficiently compressed, and if the compression algorithm allows fault-tolerant decompression, that might allow aliens to extract useful information even with incomplete reception of the transmitted data.</p><h3>The compression tradeoff</h3><p>We’d want to convey useful information on the <strong>essence</strong> of humanity with <strong>lossy compression</strong> (i.e. we aren’t able to perfectly recover the original data when decompressing). I speculate that the tradeoff between useful information and lossiness of compression may look something like the below, so we’d want to target the highest point in the curve.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*peVnP6QaPft--FpKzvLveA.png" /><figcaption>(<a href="https://docs.google.com/spreadsheets/d/1bvpLlxeYF_njHC4oPDhXIgC7bGO5ZLejPPKl9OxFGTw/edit#gid=0">spreadsheet for the above</a>)</figcaption></figure><p>As you increase lossiness, the information, when decompressed/recovered, becomes less true to the original data and thus less “useful,” but it allows for more efficient compression algorithms, which allows you to fit more usefulness into each bit. The definition of what is “useful” info about humanity is a philosophical debate which I won’t get into now.</p><h3>Machine learning to the rescue</h3><p>Embeddings (essentially, a list of numbers) generated by deep neural networks (e.g. <a href="https://medium.com/@peacej2/generating-fake-k-pop-faces-part-1-6202dc27f0eb">autoencoders</a> or language models such as <a href="http://jalammar.github.io/illustrated-bert/">BERT</a>) can be seen as a form of lossy compression (where the degree of compression can be arbitrarily chosen). For example, you can represent an image or the profile of a human with an embedding. As <a href="https://arxiv.org/abs/1206.5538">representation learning</a> techniques advance, perhaps we’ll be able to accurately represent a human’s behavior/personality with an embedding (queue <a href="https://www.youtube.com/watch?v=rYelEUVQ50g">Westworld theme song</a>).</p><p>Via training, a neural network can figure out how to efficiently represent the essence of the information. Such an approach is <a href="https://www.forbes.com/sites/nvidia/2019/02/11/accelerating-the-internet-with-ai-based-media-compression/#1a6508107760">already being used to reduce video download sizes</a>. In the future, we’ll have even better machine learning techniques to do this even more efficiently.</p><p>Of course, we’ll need to send not just the embeddings, but also the neural network parameters as well so that the aliens can make sense of the embeddings.</p><p>The information that is represented isn’t limited to just words, images, and numbers. A generative neural network that can accurately generate random new virtual humans (i.e. it “samples” from a realistic distribution of human physique and behavior) or simulate human movement in cities may also be “information” that we want to send.</p><h3>It’s t̶u̶r̶t̶l̶e̶s̶ models all the way down</h3><p>Let’s get meta for a moment. Neural networks themselves are information. In fact, <a href="https://arxiv.org/abs/1906.04358">even neural networks with randomly chosen parameters can perform well</a>, which implies that the architecture of the network is also important information. Furthermore, a single neural network can be a lot of information. For example, <a href="https://medium.com/@Synced/openai-unveils-175-billion-parameter-gpt-3-language-model-3d3f453124cd">OpenAI’s GPT-3 language model has 175 billion parameters</a>. Hundreds of years from now, we might have models with quadrillions of parameters (which could… simulate all of human civilization…?)</p><p>So, we might want to send the aliens:</p><ol><li>the parameters of a neural network compressed using another neural network</li><li>a model which can design/train other models (although getting this to run on an alien’s hardware/software stack might be non-trivial)</li><li>#1 or #2 recursed indefinitely</li></ol><p>(BTW, <a href="https://en.wikipedia.org/wiki/Turtles_all_the_way_down">see this</a> if you didn’t get the joke about turtles 😊)</p><h3>What would aliens do with the data?</h3><p>Hard to say. Maybe they would benefit from some of our technology. Maybe they’d create a computer simulation of earth and humans living on it. Or maybe they’d using advanced bio-engineering to create living things resembling humans… In any case, it would be better than no one ever knowing that humans existed.</p><h3>This might be the most important thing we ever do</h3><p>Do you want to prioritize big, long-term impact over short-term quick wins? If so, let’s extend the thinking by a few years (or millenia). <strong>After humanity perishes, our long-term impact will be essentially zero unless we communicate with extraterrestrial intelligence. </strong>Unless, of course, we create robots that outlive us in a meaningful way…but even then, the physical robots may perish eventually due the sheer distance to inhabitable planets… then, the only way to<strong> save</strong> our legacy is via efficient transmission of information through space. Otherwise, <strong>the universe will be as if humans had never existed.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/782/1*7y_jTOIX7frtWVz5iSoBPQ.png" /><figcaption>Image from <a href="https://art.marcsimonetti.com/science-fiction">https://art.marcsimonetti.com/science-fiction</a></figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=860026ae68f7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The nature of human intelligence from a computer science perspective]]></title>
            <link>https://medium.com/data-science/the-nature-of-human-intelligence-from-a-computer-science-perspective-30f6e9c09146?source=rss-8faf1fc6cd9c------2</link>
            <guid isPermaLink="false">https://medium.com/p/30f6e9c09146</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[intelligence]]></category>
            <category><![CDATA[neuroscience]]></category>
            <category><![CDATA[learning]]></category>
            <dc:creator><![CDATA[Jerry Chi]]></dc:creator>
            <pubDate>Sat, 12 Oct 2019 15:31:02 GMT</pubDate>
            <atom:updated>2019-10-16T05:19:41.401Z</atom:updated>
            <content:encoded><![CDATA[<p>I’m tremendously curious about the long-term future of artificial intelligence. Understanding human intelligence better may be key to predicting the future of artificial intelligence. Thanks to both old books and recent research, my thinking and predictions have evolved quite a bit in the last few years. I now believe that Strong AI will be realized.</p><p>Regarding human intelligence, I now believe that the following statements are probably true. Some of this is just speculation. (For brevity, I’ll omit the words “probably” and “in my opinion” for the rest of the article.)</p><ul><li>All or almost all of human intelligence can be deemed to be some form of computation or data input/output/storage.</li><li>Human intelligence doesn’t just come from the brain. Actually, almost every part of the human body (and even bacteria etc.) plays a role in human intelligence.</li><li>The whole is much greater than the sum of its parts. E.g. each cell has limited intelligence by itself, but together, great intelligence can be achieved.</li><li><strong>Parallel </strong>computation at multiple levels is key and the parallelism is in a way even more impressive than a data center full of GPUs.</li><li>Computation is <strong>distributed </strong>and <strong>decentralized </strong>(like blockchains 😜); there is no centralized component. There is no human analogue of a CPU controlling a whole computer or master node controlling a cluster of computers.</li><li><strong>Recursion / feedback</strong> <strong>loops / self-improvement </strong>are also important at multiple levels</li><li><strong>Attention </strong>and <strong>abstraction </strong>mechanisms greatly decrease the amount of computation required (i.e. they improve the bang for the buck for each unit of computation)</li><li>There are many ways that parts of the human brain/body execute computer science algorithms or mathematical operations</li><li>Human intelligence makes use of both machine learning algorithms and logical reasoning (<a href="https://en.wikipedia.org/wiki/Computer_algebra">symbolic computation</a> etc.)</li><li>Human intelligence would be very difficult to achieve using current computer hardware architecture; there must be a better approach. The human body/brain takes advantage of biology, chemistry, and physics in a way that normal computers can’t (yet).</li></ul><p>Let’s dig into each of these points in more detail.</p><h4>Intelligence is just computation and data input/output/storage</h4><p>In other words, I believe in <a href="https://en.wikipedia.org/wiki/Functionalism_(philosophy_of_mind)"><strong>functionalism</strong></a>. I think rejecting functionalism would mean that you’d need to believe in some concept of a soul or some other non-tangible/non-physical phenomena (which you’d never be able to verify). This relates to the nature of human consciousness. For a nice overview, see <a href="http://omgwai.com/itsdho/Dennett_2001.pdf"><em>Are We Explaining Consciousness yet? (2001 paper by Daniel Dennett)</em></a>.</p><h4>Intelligence doesn’t just come from the brain</h4><p>I used to think that all intelligence comes from the brain, but now I think that is wrong. Let’s consider a thought experiment: if we were able to surgically remove my brain and place it into a robot’s body, would I think/behave/feel the same way? No, I don’t think so (unless we have advanced technology that can synthesize robots that function exactly like human bodies).</p><p>In other words, our bodies are part of our intelligence. For example, recent research shows that <a href="http://The whole is much greater than the sum of its parts.">the gut and its microbiome has significant affects on the brain and human behavior.</a> A finger touching a hot surface will instantly retract before any electric signal reaches the brain. White blood cells wage war against invaders without instructions from the brain. Cells move around, release chemicals, and replicate following something like a computer program (DNA). It doesn’t matter if cells are unconscious; they contribute to the overall intelligence of the human. Also, did you know that <a href="https://www.youtube.com/watch?v=ooA0J6DWWTM&amp;t=13m19s">a single cell can solve mazes and difficult computer science problems</a>?</p><p>Humans don’t just gain intelligence by reading books or listening to lectures. Feeling the world through all senses and sensations (consciously and unconsciously) helps a human understand the world and act intelligently in that world.</p><h4>The whole is much greater than the sum of its parts</h4><p>One human is simply the “animal” level of the <a href="https://waitbutwhy.com/2019/10/idea-labs-echo-chambers.html"><strong>Emergence Tower</strong></a>, where each level is constituted by combining several instances of the level below it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kJXRYzvsw55xVqAgz7T87w.png" /><figcaption>from <a href="https://waitbutwhy.com/2019/10/idea-labs-echo-chambers.html">https://waitbutwhy.com/2019/10/idea-labs-echo-chambers.html</a></figcaption></figure><p>An interesting analogy is how an ant colony is collectively much more intelligent than the sum of each constituent ant’s intelligence; this is perhaps <a href="https://qr.ae/TWYnOU">described most entertainingly in the famous book <em>G</em></a>ö<a href="https://qr.ae/TWYnOU"><em>del, Escher, and Bach</em></a>.</p><p>Related concepts are the <a href="https://en.wikipedia.org/wiki/The_Wisdom_of_Crowds"><strong>wisdom of crowds</strong></a> and computer science concepts such as <strong>evolutionary algorithms</strong>,<strong> </strong><a href="https://www.lamsade.dauphine.fr/~lang/papers/sofsem07.pdf"><strong>computational social choice</strong></a>,<strong> graph computing</strong>, and <strong>neural networks </strong>(where <a href="https://medium.com/intuitionmachine/the-end-of-monolithic-deep-learning-86937c86bc1f">multiple neural networks combined</a> can be more “intelligent” or useful than the simple sum of each constituent network’s intelligence, and one network can be deemed more intelligent than the sum of each constituent layer’s intelligence).</p><h4>Parallel computing</h4><p>If we assume most of the <a href="https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3842595/">cells in the human body can each be deemed to have some sort of computational capacity</a> (a single cell might be dumb and slow by itself, but that’s besides the point), then that implies one human can make 10s of trillions of computations simultaneously (in <strong>parallel</strong>). Compare that with traditional computers:</p><ul><li>A traditional CPU can only process one thing at a time.</li><li>A multithreading CPU might be able to handle 64 simultaneous computations.</li><li>A decent GPU might be able to handle <a href="https://stackoverflow.com/questions/6490572/cuda-how-many-concurrent-threads-in-total">tens of thousands of simultaneous computations</a>.</li><li>A data warehouse full of GPUs might be able to handle 100 million simultaneous computations.</li></ul><p>Parallel computing is key because it allows humans to e.g. react to things very quickly. E.g. if you saw an object flying towards your head, you would immediately try to dodge, not process each retina cell’s data one by one before acting. Of course, humans can’t apply efficient parallel processing to any arbitrary computation; humans are set up (via evolution) to use parallel processing well only for certain types of tasks.</p><h4>Distributed and decentralized computation</h4><p>This is true both for the body overall and for the brain. If we think of each computational unit (e.g. a cell, an organ, a piece of the brain) as an individual agent, then the agents are both <strong>competing </strong>and <strong>collaborating</strong>. At any given time, your gut might be competing with part of your brain to decide on the next feeling/thought/action, and one part of your brain could collaborate with another part to form a fuller thought, and then compete with yet another part of the brain to enter your consciousness and dominate your thought. Such collaborative and competitive mechanisms are what enable the whole to be greater than the sum of its parts (as described above).</p><p>Some would argue that the thalamus / prefrontal cortex acts as a centralized processor, but this is <a href="https://www.researchgate.net/post/Is_the_thalamus_REALY_the_CPU_of_the_Brain">debatable</a>.</p><h4>Recursion / feedback loops / self-improvement</h4><p>There are at least 2 ways to apply this concept.</p><ol><li>The brain seems <a href="http://www.letras.ufrj.br/poslinguistica/recursion/papers/17-friederici.pdf">set up to perform tasks requiring recursive logic </a>(e.g. speaking or understanding phrases like “Peter knew that Maria who loved Hans contacted Johann”)</li><li>The brain is able to <a href="https://www.bbc.com/future/article/20151123-the-brains-miracle-superpowers-of-self-improvement">self-improve</a>. I.e. part of neuroplasticity is realized via the brain changing itself (indirectly?). For example, having some thoughts can <a href="https://www.scirp.org/journal/PaperInformation.aspx?PaperID=73936">result in some enzymes being released, which then cause changes in the number of synapses and synaptic connections</a>.</li></ol><p>I have a feeling that feedback loops are at play in other ways as well.</p><p>By the way, many people believe that self-improving AI is the key to better AI (and to conscious AI and Strong AI). From a practical standpoint, AI that writes or helps write AI software is a great example (as <a href="http://www.youtube.com/watch?v=BH4ItyTVMEs&amp;t=42m46s">François Chollet, the creator of Keras, believes</a>).</p><h4>Attention and Abstraction</h4><p>Humans are good at minimizing wasteful computations. This follows naturally from evolution. A human might have 10 objects in her visual field, and 10 details about each of those objects. But human would naturally 1. limit <strong>attention </strong>to only one of those objects (e.g. attractive potential mate or a dangerous rock flying towards one’s face) and 2. make a simple <strong>abstraction </strong>of the object while ignoring the details (e.g. understand that it’s a dangerous, fast-moving object — no need to think about the exact shape or patterns on the rock). Attention and abstraction mechanisms can sometimes be weaknesses (e.g. sometimes humans jump to conclusions prematurely or gloss over important details) but most of the time they contribute positively to human intelligence.</p><p>Computers can use similar approaches. The famous 2017 paper <a href="https://arxiv.org/abs/1706.03762"><em>Attention is All You Need</em></a><em> </em>led to huge advances in natural language processing. The 2018 paper <a href="https://arxiv.org/abs/1803.10122"><em>World Models</em></a><em> </em>described advances in using neural networks to create <strong>abstract </strong>models of the world that are computationally tractable and combining them with other neural networks to perform tasks well.</p><h4>Computer science algorithms and mathematical operations</h4><p>For example, the <a href="https://towardsdatascience.com/human-like-machine-hearing-with-ai-1-3-a5713af6e2f8">ear amplifies certain sound frequencies and then acts as multiple bandpass filters</a> to separate sounds into components. The visual cortex performs multi-layer convolutions (kind of like pooling pixels together) which inspired convolutional neural networks (CNNs), a core part of modern AI for image processing. (See the <a href="http://www.deeplearningbook.org/contents/convnets.html">Neuroscientific Basis for Convolutional Networks on page 358 of the Deep Learning book</a>.)</p><p>You could view such phenomena either as “the body uses such mathematical approaches” or, reversing the order, you could say “these mathematical models describe the body’s functionality” or “these algorithms are inspired by biology” but I don’t think it matters which view you choose; my main point is that the body can do cool stuff which contributes to human intelligence.</p><p>For more on this, check out the 2012 paper <a href="https://www.currentscience.ac.in/Volumes/103/04/0370.pdf"><em>Computation Algorithms Inspired by Biological Processes</em></a>.</p><h4>Machine learning algorithms and symbolic reasoning</h4><p>Broadly, there are 2 major approaches to AI:</p><ol><li>statistical machine learning (learning from the data, e.g. learning to differentiate cats from dogs by looking at the pixels of many labeled photos)</li><li>symbolic AI / logical reasoning / <a href="https://en.wikipedia.org/wiki/Computer_algebra">computer algebra</a> (e.g. manipulate 2y=x into y=0.5x or deduce from “all men are mortal” and “John is a man” that “John is mortal.”)</li></ol><p>I think that humans use some form of both of the above. Only using one of the above would be insufficient to explain the intelligence humans have demonstrated. Statistical machine learning is nowadays often implemented via neural networks which are inspired by the human brain. As for symbolic AI, I guess humans must also be able to use something similar to that, since humans are apparently able to do algebraic manipulations and make logical deductions pretty quickly.</p><p>The question is, how exactly do we combine #1 and #2 ? There are many possible approaches. For example, see this video: <a href="https://www.youtube.com/watch?v=ETHrFxiFIUM">Bridging Machine Learning and Logical Reasoning by Abductive Learning</a> (2019). In this approach, “ machine learning models learn to perceive primitive logical facts from the raw data, while logical reasoning is able to correct the wrongly perceived facts for improving the machine learning model.” Does the human brain do something similar? Maybe. Another approach is <a href="https://medium.com/@ybergquist/opencog-vs-openai-2-different-paths-to-human-level-ai-d36deafbf966">OpenCog</a>, which includes both symbolic AI and neural networks as part of a larger connected “hypergraph.”</p><h4>Taking advantage of biology, chemistry, and physics in the real world</h4><p>A laughable but illustrative example is the <a href="https://en.wikipedia.org/wiki/Spaghetti_sort">Spaghetti Sort algorithm</a> which can be implemented by a human but not by a computer without a physical body. The algorithm involves holding a bunch of uncooked spaghetti rods of different lengths and slamming down the bunch against a table. A human can sort the rods by length with O(n) time complexity. In contrast, the fastest computer algorithm for sorting a list of numbers is O(n log n), which is worse. The computer is not able to take advantage of e.g. gravity which constantly acts on all the rods at the same time (<strong>in parallel</strong>). Of course you could use a computer to simulate gravity’s effect on all the objects, but that is computationally expensive.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/493/0*VPTNlchpXsp39-XW.jpg" /><figcaption>A human performing Spaghetti Sort</figcaption></figure><p>An analogue to Spaghetti Sort is <a href="https://interestingengineering.com/what-is-dna-computing-how-does-it-work-and-why-its-such-a-big-deal">DNA computing</a>. “The power of DNA computing comes from its ability to perform the same operation <strong>simultaneously </strong>on the contents of a test tube…operations can be performed <strong>in parallel </strong>with no added cost” (from <a href="http://www.cim.mcgill.ca/~scott/RIT/dna_computing.html">this paper</a>).</p><p>In other words, stuff (deemed to be computations) happening in the real world (in human bodies) is naturally massively parallel (all happening at the same time) and this is a <strong>huge advantage of humans over traditional computers</strong> (which naturally process things one by one and incur cost to enable parallelism — parallelism which pales in comparison to the real world).</p><p>The above may be the biggest challenge to achieving Strong AI via electronic hardware (if we refrain from speculating about <a href="https://www.quora.com/Can-quantum-computers-help-with-developing-strong-AI">using quantum computing for AI</a>).</p><h3>Closing</h3><p>The more I learn about human intelligence, the more I appreciate all the things it can do. But with the flood of awesome research related to neuroscience and ML/AI (and we’re just getting started), and since we’re getting an increasingly better understanding of the mechanisms of human intelligence, <strong>I do believe that machine intelligence will be able to replicate or supersede many aspects of human intelligence eventually.</strong> Not sure when exactly.</p><h3>Disclaimers</h3><ul><li>I’m not an expert on any of these topics. I’m just a random curious dude.</li><li>I purposefully avoided discussing the definition of intelligence; that could be a long philosophical debate. You can choose what it means to you, and hopefully the above will still make sense.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=30f6e9c09146" width="1" height="1" alt=""><hr><p><a href="https://medium.com/data-science/the-nature-of-human-intelligence-from-a-computer-science-perspective-30f6e9c09146">The nature of human intelligence from a computer science perspective</a> was originally published in <a href="https://medium.com/data-science">TDS Archive</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Guide to working with me (a data science manager)]]></title>
            <link>https://peacej2.medium.com/guide-to-working-with-me-a-data-science-manager-a9ed7d559118?source=rss-8faf1fc6cd9c------2</link>
            <guid isPermaLink="false">https://medium.com/p/a9ed7d559118</guid>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[self-improvement]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[management]]></category>
            <category><![CDATA[teamwork]]></category>
            <dc:creator><![CDATA[Jerry Chi]]></dc:creator>
            <pubDate>Thu, 26 Sep 2019 08:29:09 GMT</pubDate>
            <atom:updated>2019-11-11T02:33:19.310Z</atom:updated>
            <content:encoded><![CDATA[<p>This document is mainly for sharing with my colleagues and anyone I might work with in the future. But hey, I might as well post it on Medium; maybe someone will find it useful.</p><p><strong>Below are principles, values, and best practices I tell myself to follow</strong> (both for working at organizations and working in data science). You don’t need to believe all the same principles, but<strong> if you read them you’ll understand how I think, and that may help us cooperate effectively.</strong> I’ll try to understand how you think too, and we can work out our differences :)</p><h3>Principles / best practices for work</h3><ul><li><strong>Almost nothing is absolute</strong> (including all the below principles). When applying any belief or principle, one should consider exceptions and/or consider moderation/balance.</li><li><strong>Always be open to change. </strong>The world and situations change quickly, and so people should change too. That could include changing my below principles.</li><li><strong>Focus and prioritize</strong>. If one is working on 5 major projects at the same time, that’s probably too many to be very effective. Clear <strong>deprioritization</strong> is a necessary part of good prioritization. In some cases, consider implementing ruthless prioritization (say no — politely, of course — to important projects/people because you have even more important work).</li><li><strong>Be a driving force. </strong>Drive projects and tasks forward and achieve progress in the face of uncertainty/ambiguity. Related concepts are leadership, proactiveness, and taking ownership. But be careful of fatigue or spreading yourself too thin while driving many projects (hence the need for focus).</li><li><strong>Be free. </strong>We are all adults. We should be able to live and work the way we want to. This includes our work hours, the technical tools or Python/R libraries we want to use, etc. However, someone else should step in when someone is abusing their freedom and not producing good work as a result. Also, there are some cases where limiting freedom (by establishing a standardized, scalable approach) can lead to greater organizational effectiveness, so there’s a tradeoff to consider. But in general, I like to err towards the side of more freedom and creating soft guidelines instead of hard rules.</li><li><strong>Be open and transparent</strong> (except if it could cause major issues… e.g. legal issues)</li><li><strong>Be open-minded. </strong>Most<strong> </strong>people tend to become less accepting of conflicting approaches/ideas as they gain experience/expertise (and many are not self-aware that they’ve become less open-minded). Don’t let this happen to you.</li><li><strong>Keep people motivated, </strong>including yourself and the people working with you. If people are not motivated, they probably won’t do good work or they’ll eventually leave the team or company. Take time to think about what motivates you and others (including career goals), and push for change if there’s a motivation problem. There are many reasons for low motivation; you should try to do something about it ASAP if possible.</li><li><strong>Don’t micromanage. Err on the side of over-empowering others. </strong>It’s tricky to decide how much responsibility/authority/independence to <strong>delegate</strong> to someone. <strong>The best leaders delegate well</strong>. I think in most cases it’s better to give too much than too little; empowering people leads to higher motivation, faster work progress, faster personal growth, better long-term results, better team dynamics, and warm fuzzy feelings. Most mistakes can be fixed/dealt with. Of course there are exceptions (don’t let the new intern perform heart surgery all by herself) and it depends somewhat on each individual’s preferences. My thinking is influenced by the <a href="https://www.nbforum.com/nbreport/ilkka-paananen-the-least-powerful-ceo/">culture of my previous employer, Supercell</a>.</li><li><a href="http://schoolofherring.com/2015/07/19/ilkka-paananen-of-supercell-celebrate-learnings-from-failures/"><strong>Celebrate failure</strong></a>. It helps encourage continued risk-taking. But don’t <a href="https://www.entrepreneur.com/article/247933">fetishize failure</a>.</li><li><strong>Don’t only focus on the short term</strong>. Sure, there might be quarterly goals, and those are important, but most great accomplishments in human history took more than a quarter to achieve. Consider blocking time each week for a longer-term goal.</li></ul><h3>Principles for good communication</h3><p>Most communication / interpersonal problems could be avoided if everyone did just the first two below things well.</p><ul><li><strong>Stand in the shoes of your audience. </strong>Before hitting “send” on the e-mail, stop for a minute and think: is everything easily understandable (with sufficient context, explanation of terms, summarization of key points, etc.) for the reader given their limited knowledge? Is it clear to the reader what the next step s/he should take as a result of the e-mail? How would the reader feel after reading? This doesn’t just apply to e-mails or Slack messages. E.g. naming your dashboard is also a form of communication.</li><li><strong>Manage expectations. </strong>All the time, people end up being disappointed/frustrated or lose trust due to misaligned expectations. Most of this is preventable. Ask yourself regularly: are expectations aligned? E.g. is there a difference between your understanding of the expected timing/quality/importance/amount of your work output and a stakeholder’s understanding? If the result will be highly uncertain (as is the case for e.g. using a new machine learning model for the first time), take extra care to ensure people understand that uncertainty and don’t have unfairly high expectations. This is not just about making people happy; it often hurts a team’s effectiveness if someone is deciding a course of action based on unrealistic expectations of a colleague’s work.</li><li><strong>Err on the side of over-communication. </strong>If you’re not sure if something is worth saying/communicating, it’s better to say it just in case. The potential damage from under-communication (loss of trust, major bug, etc.) is typically far greater.</li><li><strong>Proactively give both positive and negative (critical) feedback. </strong>No one is perfect, no one is completely self-aware, and no one can optimize their work perfectly just by themselves. The most effective teams have members that help each other become better team members over time. <br><strong>Giving critical feedback:</strong> keep the message clear (don’t overly sugar-coat it) while not sounding like an asshole. Provide justification and explanation about how to improve (e.g. don’t just say “your presentation sucked”).<br><strong>Receiving critical feedback: </strong>some people suck at giving feedback. Try not to take it personally and try to turn it into something helpful. Try to give feedback on how they can give you feedback more effectively. Be very welcoming of any critical feedback and proactively invite people to give it to you.</li></ul><h3>Principles / best practices for data science and technical work</h3><ul><li><strong>Good technical work usually requires good communication. </strong>Some people see coding etc. as a hard skill and communication as a soft skill, and some people pride themselves mainly on their hard skills. But if your code is hard to read or if your technical work is hard to follow/understand, I don’t consider that good work, especially in the context of an organization. Making your work reasonably easy to read/understand requires good communication skills.</li><li><strong>Think carefully about the tradeoff between work speed and technical debt. </strong>Sometimes it makes sense to just make a quick and dirty hacky prototype, sometimes it doesn’t. Adjust based on the situation (e.g. ask yourself questions like: will this eventually be used in production? How many people will be reading/using this code, for what purpose?)</li><li><strong>Identify and make the obviously good tradeoffs first. </strong>For example, take the time to design well the folder/file structure and naming conventions for a big project. This might only take 1% of the total project time and reduce the technical debt burden by 20%. It’s surprising how many people fail to make such tradeoffs. Reflect regularly about if you’re making the right tradeoffs. Get feedback from others about it.</li><li><strong>“</strong><a href="https://en.wikiquote.org/wiki/Donald_Knuth"><strong>Premature optimization is the root of all evil</strong></a><strong>.” </strong>True… and insufficient or overly late optimization is bad as well. One needs to think about when and how much optimization is appropriate. E.g. one rule of thumb is the <a href="https://en.wikipedia.org/wiki/Rule_of_three_(computer_programming)">Rule of Three</a>.</li><li><strong>Work should be reproducible and understandable. </strong>If you did a piece of analysis or created an ML model, someone else with sufficient technical chops should be able to come and reproduce what you did with little to no guidance from you. Otherwise, 2 problems will get worse: technical debt and over-dependency on specific individuals.</li><li><strong>Almost all code should be pushed to git</strong> (or another code version control system). There are tons of benefits: reproducibility, accountability, reversibility, searchability, education of team members, etc. Exceptions might be very short, temporary scripts. If it’s too troublesome for you to push your code to git, you need to set things up so that it’s easier to do so (e.g. bash script to reduce the amount of typing when pushing to git, set up auto-sync from your Jupyter server to git, etc.)</li><li><strong>Have a culture of knowledge sharing, </strong>especially on technical topics. Internal/external blog posts, tech talks, tutorials, recommendations for tools to use, etc.</li><li><strong>Appreciate the balance between data-driven and non-data-driven methods.</strong> Sometimes, companies just have to take a risk and act on gut feel . Otherwise, certain risky/innovative things will never be attempted. Often, human decision-making is a combination of data and gut feel or qualitative info, and some of the best software is a combination of ML (based on training data) and rule-based heuristics (based on human expertise).</li><li><strong>Stand on the shoulders of giants</strong>. No need to develop something from scratch when you can reuse something off of github.</li><li><strong>Set up your technical environment to for efficient work</strong>. Sometimes I can find the exact line of code I’m looking for in 5 seconds. Set up <a href="https://winaero.com/blog/add-custom-chrome-searches-for-address-bar-to-save-your-time/">Chrome search aliases</a> for github, google docs, or whatever else you use commonly. Allow efficient use of your command line using bash aliases, <a href="https://ohmyz.sh/">oh-my-zsh</a>, etc. Enable <a href="https://jupyter-contrib-nbextensions.readthedocs.io/en/latest/nbextensions/codefolding/readme.html">code-folding in Jupyter</a>.</li><li><strong>Block off a bit of time for innovative, cutting-edge (and hopefully fun) data science experimentation.</strong> For example, check out a new paper or new Python library and see how it can be applied to problems you’re trying to solve. A common problem in organizations is that, due to time constraints and short-term goals, people are stuck with using the same old approaches/tools repeatedly (which they can implement quickly with relatively high certainty about the results). Data science / ML is a rapidly progressing field, and you will eventually be left behind or get bored if you don’t take time to try new things.</li></ul><h3>Jerry’s personality and quirks</h3><ul><li>Once I took a personality test and the automatically generated test result was that I’m a <strong>“</strong>bull in a china shop”. I hope it’s an overstatement..</li><li><strong>I speak my mind</strong>, sometimes bluntly. If I think something should be changed for the greater good, I may bang the table (figuratively) and argue aggressively about why it should be changed, regardless of the seniority of the people involved. Feel free to tell me to shut up (but please justify why you think I should). I think over time I’ve become more diplomatic, but this is still part of my fundamental personality.</li><li><strong>I have grit</strong>. I can deal with huge setbacks/failures, huge problems, huge challenges, etc. But I won’t sacrifice my health dealing with such problems (I did before and it’s not worth it).</li><li>I have little tolerance for bullshit. Sometimes I’ll fight it patiently if I have to. (Of course, one should take care to not prematurely conclude that something is bullshit.)</li><li>I don’t get embarrassed easily. Some might describe this as <a href="https://www.youtube.com/watch?v=8PxEFti645A">being shameless</a>.</li><li>I’m addicted to self-improvement and learning. Let me know if you have tips :) I still have a long way to go.</li><li>I try to think more about how to do things for the mid/long-term benefit of the organization/ecosystem, not the short-term benefit of any given individual or team. If I’m not doing this, please let me know.</li><li>I like <strong>sharing news articles, reports</strong>, papers, github repos, etc. that I think might be helpful or interesting. Let me know if it’s too much or too distracting. Don’t let it hurt your ability to focus.</li><li>I’m a big fan of <strong>Finnish values</strong> (having worked in Finland for a year). Many Finnish people are humble, blunt, straightforward, concise, and have grit.</li><li>I’m like the <strong>grammar police</strong>. I will point out people’s incorrect grammar/spelling/wording, (in any language that I speak). I do it to help people improve themselves and to facilitate better communication. But if this becomes too annoying, please let me know.</li></ul><h3>Logistical stuff</h3><ul><li>Go ahead and put things in my calendar or reschedule stuff as needed. No need to ask me in advance. My calendar is up to date.</li><li>I may block off a chunk of “focused work time” on my calendar if I feel like I’m being distracted too much.</li><li>Feel free to grab extra time with me if the regular 1-on-1 or team meeting is insufficient.</li><li>Weekly team meetings — I think people should speak up only if one has something to say that others should hear; no need to feel pressured to always give an update on your work. Also, I usually don’t like overly structuring meetings with inflexible agendas, but we can add more structure if needed.</li><li>Feel free to slack/e-mail me at any time on any day. However, I may stop checking slack/e-mail when trying to disconnect from work (on weekends etc.) in which case the best way to reach me for an emergency is via phone or Facebook Messenger.</li></ul><h3>Relevant posts by me</h3><ul><li><a href="https://medium.com/@peacej2/how-to-be-an-awesome-data-science-manager-1c63b059a03c">How to be an awesome data science manager</a></li><li><a href="https://medium.com/@peacej2/how-to-be-an-awesome-data-scientist-930f04760f7e">How to be an awesome data scientist</a></li></ul><p>This guide was partially inspired by the <a href="http://growth.eladgil.com/book/the-role-of-the-ceo/insights-working-with-claire/">Guide to Working with Claire</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/460/0*H13cLJi5Z5CThoew.jpg" /><figcaption>Typical picture of me working with a colleague</figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a9ed7d559118" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>