<?xml version="1.0" encoding="utf-8"?><feed xmlns="http://www.w3.org/2005/Atom" xml:lang="en"><generator uri="https://jekyllrb.com/" version="4.3.4">Jekyll</generator><link href="https://kyleygao.com/feed.xml" rel="self" type="application/atom+xml"/><link href="https://kyleygao.com/" rel="alternate" type="text/html" hreflang="en"/><updated>2026-08-16T00:59:25+00:00</updated><id>https://kyleygao.com/feed.xml</id><title type="html">blank</title><subtitle>Kyle Gao&apos;s personal website. </subtitle><entry><title type="html">Trip to Aomori (青森)</title><link href="https://kyleygao.com/blog/2026/aomori-japan/" rel="alternate" type="text/html" title="Trip to Aomori (青森)"/><published>2026-08-15T00:00:00+00:00</published><updated>2026-08-15T00:00:00+00:00</updated><id>https://kyleygao.com/blog/2026/aomori-japan</id><content type="html" xml:base="https://kyleygao.com/blog/2026/aomori-japan/"><![CDATA[<p>We just had an amazing 10-day deep dive into the Aomori prefecture of Japan. It’s my 4th time in Japan and my wife’s 8th (maybe 9th?), and we know we will be back many more times, so we are fortunate to be able to stay in a less touristy region for a longer time. Aomori prefecture’s biggest city is the same-named Aomori city. It’s an agriculturally rich area famous for its apples, rice, seafood, and the Nebuta summer festival.</p> <figure> <picture> <source class="responsive-img-srcset" srcset="/assets/img/2026-08-aomori/winding-red-torii-gates-480.webp 480w,/assets/img/2026-08-aomori/winding-red-torii-gates-800.webp 800w,/assets/img/2026-08-aomori/winding-red-torii-gates-1400.webp 1400w," type="image/webp" sizes="95vw"/> <img src="/assets/img/2026-08-aomori/winding-red-torii-gates.jpg" width="100%" height="auto" style=" object-fit: contain; max-height: 50vh; " data-zoomable="" loading="lazy" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"/> </picture> </figure> <p>This is Takayama Inari Shrine in the Tsugaru (津軽) region, from the same branch of Inari shrines as the famous Fushimi Inari in Kyoto. Most of the torii were donated in the last 5 years, so they are very new. The day we visited, it was drizzling and there was almost no one else besides us — a very quiet and serene feeling. Author Osamu Dazai, who’s famous for his novel <em>No Longer Human</em> and his multiple suicide attempts, was born in this region too — into a very rich family!</p> <figure> <picture> <source class="responsive-img-srcset" srcset="/assets/img/2026-08-aomori/nara-girl-mural-480.webp 480w,/assets/img/2026-08-aomori/nara-girl-mural-800.webp 800w,/assets/img/2026-08-aomori/nara-girl-mural-1400.webp 1400w," type="image/webp" sizes="95vw"/> <img src="/assets/img/2026-08-aomori/nara-girl-mural.jpg" width="100%" height="auto" style=" object-fit: contain; max-height: 50vh; " data-zoomable="" loading="lazy" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"/> </picture> </figure> <p>Yoshitomo Nara was born in Hirosaki, a city in Aomori, and he has helped cultivate an art culture in the region that feels warm and connected to the local population. We visited 3 museums — in Aomori, Hirosaki, and Towada city — that feature pieces from Nara and other local artists. One piece in the Hirosaki Museum of Contemporary Art was a reproduced model of an American-themed local bar that Nara frequented as a student and that later closed. The bar features 70s-80s rock ‘n’ roll music, Maker’s Mark and Suntory whiskey, and memories of young Nara and other Hirosaki boys.</p> <figure> <picture> <source class="responsive-img-srcset" srcset="/assets/img/2026-08-aomori/gull-on-stone-gull-statue-480.webp 480w,/assets/img/2026-08-aomori/gull-on-stone-gull-statue-800.webp 800w,/assets/img/2026-08-aomori/gull-on-stone-gull-statue-1400.webp 1400w," type="image/webp" sizes="95vw"/> <img src="/assets/img/2026-08-aomori/gull-on-stone-gull-statue.jpg" width="100%" height="auto" style=" object-fit: contain; max-height: 50vh; " data-zoomable="" loading="lazy" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"/> </picture> </figure> <p>When we took the Shinkansen from Tokyo to Aomori city, Hachinohe was the last stop before Aomori. Only later did I learn that Hachinohe was the northernmost stop of the Tohoku Shinkansen line until it was extended to Aomori city in 2010 (then to Hakodate, Hokkaido in 2016). Because more people live in Aomori city, Hachinohe used to be a lively city — people would take the fast train from Tokyo, get off there, and visit local businesses before continuing their trip to Aomori city. After the Shinkansen extension the city became much less popular (the hotel we stayed at there was terrible!).</p> <p>Nevertheless, Hachinohe is still a major fishing city known for its mackerel and squid. We woke up at 5am to visit the largest morning market at the Hachinohe port, and the self-service seafood restaurant in the Hasshoku Center was super fun and delicious (you pay a rental fee to get access to a charcoal grill, and then buy fresh seafood from the market).</p> <p>A little east of the Hachinohe morning market is the Kabushima Shrine. It was a small island that was later connected to the mainland by land reclamation. The island and shrine are famous for being the breeding ground of black-tailed gulls (umineko, 海猫), A LOT OF them — it has been a designated national natural monument for over 100 years!</p> <p>When you go to the shrine there are umbrellas you can borrow to prevent bird shit from falling on you, and if you actually get hit by bird shit, you can get a certificate from the shrine for having good luck! Fortunately and unfortunately, we didn’t get hit by bird shit!</p> <figure> <picture> <source class="responsive-img-srcset" srcset="/assets/img/2026-08-aomori/canoes-on-the-lake-480.webp 480w,/assets/img/2026-08-aomori/canoes-on-the-lake-800.webp 800w,/assets/img/2026-08-aomori/canoes-on-the-lake-1400.webp 1400w," type="image/webp" sizes="95vw"/> <img src="/assets/img/2026-08-aomori/canoes-on-the-lake.jpg" width="100%" height="auto" style=" object-fit: contain; max-height: 50vh; " data-zoomable="" loading="lazy" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"/> </picture> </figure> <p>We spent 4 days in the Towada-Hachimantai National Park, which features the amazing Lake Towada and Oirase Stream, and great hot spring hotels. Staying in the traditional Tsuta hot spring hotel, canoeing on the lake, and hiking along the stream was an amazing detox from the busy city life and, uh, the AI stress.</p> <figure> <picture> <source class="responsive-img-srcset" srcset="/assets/img/2026-08-aomori/tachineputa-float-at-dusk-480.webp 480w,/assets/img/2026-08-aomori/tachineputa-float-at-dusk-800.webp 800w,/assets/img/2026-08-aomori/tachineputa-float-at-dusk-1400.webp 1400w," type="image/webp" sizes="95vw"/> <img src="/assets/img/2026-08-aomori/tachineputa-float-at-dusk.jpg" width="100%" height="auto" style=" object-fit: contain; max-height: 50vh; " data-zoomable="" loading="lazy" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"/> </picture> </figure> <p>Finally, the highlight of the trip was the Nebuta festivals. It’s a local tradition where giant warrior floats (nebuta) are paraded with traditional flute music and dance during a week in summer, to expel the sleep demon so that people can work hard and produce more agricultural output. Every city has its own special nebuta: Hirosaki’s (called neputa there) are fan-shaped, Goshogawara’s are very tall, and Aomori’s, the largest festival in the region, are wide and mighty.</p> <p>All the nebuta are built by local businesses and organizations, with participation from their employees, employees’ relatives, and local people. Many kids and older people are also involved, sometimes multiple generations of the same family. On the days of the festival people dress up in traditional clothes, restaurants open carts along the street to sell delicious street food, and people cheer and have fun as the nebuta pass.</p> <p>We saw all three cities’ parades. For the Aomori one we reserved good seats for the 2-hour-long parade, but Goshogawara’s was the most impressive for me. In Goshogawara the nebuta are tall — actually 23 meters tall and almost 20 tons heavy — and they are called Tachineputa. The craft of making Tachineputa was almost lost after two big fires in the city, but after a blueprint was rediscovered in 1993, three guys who were proud of their hometown tradition committed to recovering the craft and restarted the tradition. I am jealous that they have a local tradition that’s worth being proud of and can be passed down through generations.</p>]]></content><author><name></name></author><category term="article"/><category term="photo"/><category term="photo"/><summary type="html"><![CDATA[We just had an amazing 10-day deep dive into the Aomori prefecture of Japan. It’s my 4th time in Japan and my wife’s 8th (maybe 9th?), and we know we will be back many more times, so we are fortunate to be able to stay in a less touristy region for a longer time. Aomori prefecture’s biggest city is the same-named Aomori city. It’s an agriculturally rich area famous for its apples, rice, seafood, and the Nebuta summer festival.]]></summary></entry><entry><title type="html">Using Claude Code with Obsidian - A Perfect Pairing for Knowledge Work</title><link href="https://kyleygao.com/blog/2025/using-claude-code-with-obsidian/" rel="alternate" type="text/html" title="Using Claude Code with Obsidian - A Perfect Pairing for Knowledge Work"/><published>2025-11-22T00:00:00+00:00</published><updated>2025-11-22T00:00:00+00:00</updated><id>https://kyleygao.com/blog/2025/using-claude-code-with-obsidian</id><content type="html" xml:base="https://kyleygao.com/blog/2025/using-claude-code-with-obsidian/"><![CDATA[<blockquote> <p>Disclaimer: These are my personal views and experiences and do not represent the views of Anthropic. This blog post was written together with Claude.</p> </blockquote> <h2 id="the-problem-aspiration-vs-reality-in-note-taking">The Problem: Aspiration vs. Reality in Note-Taking</h2> <p>I’ve been using Obsidian for my personal knowledge base for the past three years. During this time, I’ve built a comprehensive vault containing journals, book reviews, movie reviews, restaurant notes, meeting summaries, and records of interactions with interesting people. Obsidian appeals to me for the same reasons VS Code does for programming—it’s fast, extensible, and stores everything as plain markdown files in the filesystem. It’s a tool that gets out of your way and lets you think.</p> <p>But there’s always been a gap between my aspirations and reality. Here are things I want to do with Obsidian but never have the time or energy for:</p> <ol> <li><strong>Comprehensive backlinking</strong>: I want to link entities (people, places, concepts) everywhere they appear, but I’m too lazy to remember all the entity names and type <code class="language-plaintext highlighter-rouge">[[entity]]</code> consistently.</li> <li><strong>Diligent reviews</strong>: I aspire to write reviews for every book and movie I consume and document every meaningful interaction with people, but often don’t get around to it.</li> <li><strong>Decision frameworks</strong>: After reading <em>Poor Charlie’s Almanack</em>, I wanted to adopt Munger’s mental models checklist for important decisions. The problem is that it’s time-consuming and requires discipline to apply rigorously.</li> <li><strong>Investment retrospectives</strong>: I try to write investment theses for stocks I trade and want to do retrospectives to evaluate my reasoning and learn from outcomes, but rarely follow through.</li> </ol> <p>The pattern is clear: these are all valuable practices that require grunt work, consistency, and discipline—qualities that are easy to aspire to but hard to maintain.</p> <h2 id="why-claude-code--obsidian-works">Why Claude Code + Obsidian Works</h2> <p>At work, I’ve been using Claude Code extensively. In March this year, Dario <a href="https://www.businessinsider.com/anthropic-ceo-ai-90-percent-code-3-to-6-months-2025-3">predicted that AI will write 90% of code in 6 months</a>, and recently he confirmed on 60 Minutes that AI does write 90% of code at Anthropic. This is certainly true for me. Claude Code excels at:</p> <ol> <li>Understanding codebases with complex and large amounts of code</li> <li>Understanding user requests and making changes to achieve them</li> <li>Agentically editing and testing changes to ensure they work</li> </ol> <p>Two weeks ago, something clicked: <strong>why not use Claude Code with my Obsidian vault?</strong></p> <p>At first glance, this seems like an odd pairing. Obsidian is for note-taking, not coding. But the more I thought about it, the more it made sense:</p> <ul> <li><strong>Structure similarity</strong>: An Obsidian vault is essentially a codebase of markdown files. The files are stored in the filesystem and accessible to other applications, including Claude Code. Claude is already excellent at navigating file structures and making surgical edits.</li> <li><strong>AI’s strengths align with my gaps</strong>: The tasks I avoid in Obsidian—organization, consistency, following procedures, grunt work—are exactly what AI excels at. AI doesn’t get tired or bored and doesn’t care about repetitive labor (given sufficient tokens).</li> <li><strong>Natural language interface</strong>: Instead of learning complex Obsidian plugins or writing scripts myself, I can simply tell Claude what I want in natural language.</li> </ul> <p>I was initially concerned that Claude Code’s system prompt might be too code-oriented, making it awkward for note-taking tasks. To my pleasant surprise, it works remarkably well. Claude understands the context of personal knowledge management and adapts its behavior accordingly.</p> <h2 id="how-i-use-claude-code-with-obsidian">How I Use Claude Code with Obsidian</h2> <p>Here are some concrete examples of how Claude Code has transformed my Obsidian workflow:</p> <h3 id="automated-backlinking">Automated Backlinking</h3> <p>I can ask Claude to “read my journal entry from today and add backlinks to all people, places, and books mentioned.” Claude searches my vault for existing entity notes, creates new ones if needed, and adds proper wiki-links throughout the document. What would take me 10-15 minutes of tedious work happens in seconds.</p> <h3 id="formatting-rough-notes">Formatting Rough Notes</h3> <p>When I finish a book, I do a quick brain dump—bullet points of thoughts, quotes, reactions—without caring about structure or formatting at all. Then I ask Claude to “convert these rough notes into a properly formatted review with sections.” Claude takes my stream-of-consciousness bullets and organizes them into a coherent note with appropriate headings, properly formatted quotes, and logical flow.</p> <h3 id="applying-mental-models">Applying Mental Models</h3> <p>Before making a significant decision, I can ask Claude to “help me apply Munger’s mental models checklist to evaluate [decision].” Claude walks me through each mental model systematically, asking clarifying questions and documenting the reasoning.</p> <h3 id="investment-post-mortems">Investment Post-Mortems</h3> <p>When a stock position closes, I ask Claude to “create a retrospective comparing my initial thesis to what actually happened, and extract lessons learned.” Claude retrieves my original investment note, compares it with current information (using web search if needed), and drafts a thoughtful analysis.</p> <h2 id="the-transformation">The Transformation</h2> <p>The impact on my note-taking practice has been dramatic. Tasks that felt like chores now feel effortless. More importantly, I find myself using Obsidian more frequently because Claude Code removes the friction. Capturing a quick thought doesn’t mean committing to 10 minutes of organization work later—I can just ask Claude to “file this appropriately and add necessary links.”</p> <p>The experience went from something I knew I <em>should</em> do to something I genuinely look forward to. It’s similar to how Claude Code has made programming more enjoyable by handling the tedious parts, letting you focus on the creative and strategic thinking.</p> <p>Give it a try. You might find, as I did, that this combination transforms how you think about personal knowledge management.</p>]]></content><author><name></name></author><category term="article"/><category term="productivity"/><category term="ai"/><category term="tools"/><summary type="html"><![CDATA[How Claude Code transforms the Obsidian note-taking experience through intelligent automation and natural language workflows]]></summary></entry><entry><title type="html">Photos from trip to Asia</title><link href="https://kyleygao.com/blog/2025/photos-from-asia-trip/" rel="alternate" type="text/html" title="Photos from trip to Asia"/><published>2025-05-11T00:00:00+00:00</published><updated>2025-05-11T00:00:00+00:00</updated><id>https://kyleygao.com/blog/2025/photos-from-asia-trip</id><content type="html" xml:base="https://kyleygao.com/blog/2025/photos-from-asia-trip/"><![CDATA[<h2 id="taiwan">Taiwan</h2> <p>Cute foo dog in from of the <a href="https://maps.app.goo.gl/UBKSvnVR4WD7Bxsn7">大稻埕慈聖宮</a>, there is a <a href="https://maps.app.goo.gl/vU6mSbeLbhbJYV5UA">pork trotter noodle shop</a> that’s incredibly dilicious and nofrill.</p> <figure> <picture> <source class="responsive-img-srcset" srcset="/assets/img/2025-05-11-asia/taiwan/foo-dog-480.webp 480w,/assets/img/2025-05-11-asia/taiwan/foo-dog-800.webp 800w,/assets/img/2025-05-11-asia/taiwan/foo-dog-1400.webp 1400w," type="image/webp" sizes="95vw"/> <img src="/assets/img/2025-05-11-asia/taiwan/foo-dog.jpg" width="100%" height="auto" style=" object-fit: contain; max-height: 50vh; " data-zoomable="" loading="lazy" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"/> </picture> </figure> <p>Probably my favorite photo from this trip. The boy is about 16 years old and I just like the color of his jacket and skateboard and his sense of going somewhere.</p> <figure> <picture> <source class="responsive-img-srcset" srcset="/assets/img/2025-05-11-asia/taiwan/boy-with-skateboard-480.webp 480w,/assets/img/2025-05-11-asia/taiwan/boy-with-skateboard-800.webp 800w,/assets/img/2025-05-11-asia/taiwan/boy-with-skateboard-1400.webp 1400w," type="image/webp" sizes="95vw"/> <img src="/assets/img/2025-05-11-asia/taiwan/boy-with-skateboard.jpg" width="100%" height="auto" style=" object-fit: contain; max-height: 50vh; " data-zoomable="" loading="lazy" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"/> </picture> </figure> <p>NTU has a uniquely blue track field (guess which <a href="https://www.google.com/search?sca_esv=c148ee2db824e130&amp;rlz=1C5CHFA_enUS1093US1093&amp;sxsrf=AHTn8zpfqtbGrOqIA-g2-duWi3KxgGacng:1747012027601&amp;q=%E5%8C%97%E8%88%AA+%E8%93%9D%E8%B7%91%E9%81%93&amp;udm=2&amp;fbs=ABzOT_CWdhQLP1FcmU5B0fn3xuWpA-dk4wpBWOGsoR7DG5zJBsxayPSIAqObp_AgjkUGqekYoUzDaOcDDjQfK4KpR2OIjj43mhrQBsMJgHY2LSx-SUj4wz68xSZ8iYTfqgrdxb3MJvHOMODdIcpti-xYMckL_DuO7Mno3LlWlsnznPPjfINcnPSb3s0mY1_Udv3xmGYGwDe_3zR2JNQT7OndwaUM5c3nJw&amp;sa=X&amp;ved=2ahUKEwjPhMv63pyNAxWbv4kEHW7yM5EQtKgLegQIEBAB&amp;biw=1296&amp;bih=1294&amp;dpr=2">other university</a> also has a blue one?). I guess I’m old, universities now give me a feeling of youth that’s almost foreign, and I felt the same when I went back to CMU a couple of years ago.</p> <figure> <picture> <source class="responsive-img-srcset" srcset="/assets/img/2025-05-11-asia/taiwan/ntu-1-480.webp 480w,/assets/img/2025-05-11-asia/taiwan/ntu-1-800.webp 800w,/assets/img/2025-05-11-asia/taiwan/ntu-1-1400.webp 1400w," type="image/webp" sizes="95vw"/> <img src="/assets/img/2025-05-11-asia/taiwan/ntu-1.jpg" width="100%" height="auto" style=" object-fit: contain; max-height: 50vh; " data-zoomable="" loading="lazy" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"/> </picture> </figure> <figure> <picture> <source class="responsive-img-srcset" srcset="/assets/img/2025-05-11-asia/taiwan/ntu-2-480.webp 480w,/assets/img/2025-05-11-asia/taiwan/ntu-2-800.webp 800w,/assets/img/2025-05-11-asia/taiwan/ntu-2-1400.webp 1400w," type="image/webp" sizes="95vw"/> <img src="/assets/img/2025-05-11-asia/taiwan/ntu-2.jpg" width="100%" height="auto" style=" object-fit: contain; max-height: 50vh; " data-zoomable="" loading="lazy" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"/> </picture> </figure> <p>Another favorite photo from this trip. Taipei has a noticably older population. This old lady was just collecting cardboard boxes on <a href="https://maps.app.goo.gl/HB38SB9mM3n816HG7">迪化街</a></p> <figure> <picture> <source class="responsive-img-srcset" srcset="/assets/img/2025-05-11-asia/taiwan/old-lady-480.webp 480w,/assets/img/2025-05-11-asia/taiwan/old-lady-800.webp 800w,/assets/img/2025-05-11-asia/taiwan/old-lady-1400.webp 1400w," type="image/webp" sizes="95vw"/> <img src="/assets/img/2025-05-11-asia/taiwan/old-lady.jpg" width="100%" height="auto" style=" object-fit: contain; max-height: 50vh; " data-zoomable="" loading="lazy" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"/> </picture> </figure> <p>The streets in Taipei are old and vibrant at the same time.</p> <figure> <picture> <source class="responsive-img-srcset" srcset="/assets/img/2025-05-11-asia/taiwan/street-1-480.webp 480w,/assets/img/2025-05-11-asia/taiwan/street-1-800.webp 800w,/assets/img/2025-05-11-asia/taiwan/street-1-1400.webp 1400w," type="image/webp" sizes="95vw"/> <img src="/assets/img/2025-05-11-asia/taiwan/street-1.jpg" width="100%" height="auto" style=" object-fit: contain; max-height: 50vh; " data-zoomable="" loading="lazy" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"/> </picture> </figure> <figure> <picture> <source class="responsive-img-srcset" srcset="/assets/img/2025-05-11-asia/taiwan/street-2-480.webp 480w,/assets/img/2025-05-11-asia/taiwan/street-2-800.webp 800w,/assets/img/2025-05-11-asia/taiwan/street-2-1400.webp 1400w," type="image/webp" sizes="95vw"/> <img src="/assets/img/2025-05-11-asia/taiwan/street-2.jpg" width="100%" height="auto" style=" object-fit: contain; max-height: 50vh; " data-zoomable="" loading="lazy" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"/> </picture> </figure> <h2 id="shanghai">Shanghai</h2> <p>Behold, the manificant BUND.</p> <figure> <picture> <source class="responsive-img-srcset" srcset="/assets/img/2025-05-11-asia/shanghai/bund-1-480.webp 480w,/assets/img/2025-05-11-asia/shanghai/bund-1-800.webp 800w,/assets/img/2025-05-11-asia/shanghai/bund-1-1400.webp 1400w," type="image/webp" sizes="95vw"/> <img src="/assets/img/2025-05-11-asia/shanghai/bund-1.jpg" width="100%" height="auto" style=" object-fit: contain; max-height: 50vh; " data-zoomable="" loading="lazy" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"/> </picture> </figure> <figure> <picture> <source class="responsive-img-srcset" srcset="/assets/img/2025-05-11-asia/shanghai/bund-2-480.webp 480w,/assets/img/2025-05-11-asia/shanghai/bund-2-800.webp 800w,/assets/img/2025-05-11-asia/shanghai/bund-2-1400.webp 1400w," type="image/webp" sizes="95vw"/> <img src="/assets/img/2025-05-11-asia/shanghai/bund-2.jpg" width="100%" height="auto" style=" object-fit: contain; max-height: 50vh; " data-zoomable="" loading="lazy" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"/> </picture> </figure> <h2 id="japan">Japan</h2> <p>Setonaikai, the inner harbor of Hiroshima with numerous islands, is the highlight of this trip to Japan.</p> <figure> <picture> <source class="responsive-img-srcset" srcset="/assets/img/2025-05-11-asia/japan/belvedere-setoda-480.webp 480w,/assets/img/2025-05-11-asia/japan/belvedere-setoda-800.webp 800w,/assets/img/2025-05-11-asia/japan/belvedere-setoda-1400.webp 1400w," type="image/webp" sizes="95vw"/> <img src="/assets/img/2025-05-11-asia/japan/belvedere-setoda.jpg" width="100%" height="auto" style=" object-fit: contain; max-height: 50vh; " data-zoomable="" loading="lazy" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"/> </picture> <figcaption class="caption"><p>Belvedere Setoda by 川上 喜三郎, it’s like a modern torri that teleports you to a different world.</p> </figcaption> </figure> <figure> <picture> <source class="responsive-img-srcset" srcset="/assets/img/2025-05-11-asia/japan/bikes-480.webp 480w,/assets/img/2025-05-11-asia/japan/bikes-800.webp 800w,/assets/img/2025-05-11-asia/japan/bikes-1400.webp 1400w," type="image/webp" sizes="95vw"/> <img src="/assets/img/2025-05-11-asia/japan/bikes.jpg" width="100%" height="auto" style=" object-fit: contain; max-height: 50vh; " data-zoomable="" loading="lazy" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"/> </picture> <figcaption class="caption"><p>Mopads are definitely popular on the Ikuchi Island.</p> </figcaption> </figure> <figure> <picture> <source class="responsive-img-srcset" srcset="/assets/img/2025-05-11-asia/japan/guardrail-by-the-sea-480.webp 480w,/assets/img/2025-05-11-asia/japan/guardrail-by-the-sea-800.webp 800w,/assets/img/2025-05-11-asia/japan/guardrail-by-the-sea-1400.webp 1400w," type="image/webp" sizes="95vw"/> <img src="/assets/img/2025-05-11-asia/japan/guardrail-by-the-sea.jpg" width="100%" height="auto" style=" object-fit: contain; max-height: 50vh; " data-zoomable="" loading="lazy" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"/> </picture> <figcaption class="caption"><p>Guardrail by the sea, I’m just admiring the colors.</p> </figcaption> </figure> <figure> <picture> <source class="responsive-img-srcset" srcset="/assets/img/2025-05-11-asia/japan/hill-of-hope-480.webp 480w,/assets/img/2025-05-11-asia/japan/hill-of-hope-800.webp 800w,/assets/img/2025-05-11-asia/japan/hill-of-hope-1400.webp 1400w," type="image/webp" sizes="95vw"/> <img src="/assets/img/2025-05-11-asia/japan/hill-of-hope.jpg" width="100%" height="auto" style=" object-fit: contain; max-height: 50vh; " data-zoomable="" loading="lazy" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"/> </picture> <figcaption class="caption"><p><a href="https://maps.app.goo.gl/T2TRbmiPizSZeF3V6">Hill of Hope</a>, of the wildest things I’ve seen in my life with tons of Italian marble moved to this island to build this sculpture. Not to mention that it is actually located in a temple <a href="https://maps.app.goo.gl/iZpqKSWERakydp4ZA">耕三寺</a> that’s also wild itself.</p> </figcaption> </figure> <figure> <picture> <source class="responsive-img-srcset" srcset="/assets/img/2025-05-11-asia/japan/onomichi-kids-480.webp 480w,/assets/img/2025-05-11-asia/japan/onomichi-kids-800.webp 800w,/assets/img/2025-05-11-asia/japan/onomichi-kids-1400.webp 1400w," type="image/webp" sizes="95vw"/> <img src="/assets/img/2025-05-11-asia/japan/onomichi-kids.jpg" width="100%" height="auto" style=" object-fit: contain; max-height: 50vh; " data-zoomable="" loading="lazy" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"/> </picture> <figcaption class="caption"><p>Middle school kids having a field trip in Onomichi. They just found out I was taking a photo of them while they are taking pictures of themselves.</p> </figcaption> </figure> <figure> <picture> <source class="responsive-img-srcset" srcset="/assets/img/2025-05-11-asia/japan/setoda-lemon-480.webp 480w,/assets/img/2025-05-11-asia/japan/setoda-lemon-800.webp 800w,/assets/img/2025-05-11-asia/japan/setoda-lemon-1400.webp 1400w," type="image/webp" sizes="95vw"/> <img src="/assets/img/2025-05-11-asia/japan/setoda-lemon.jpg" width="100%" height="auto" style=" object-fit: contain; max-height: 50vh; " data-zoomable="" loading="lazy" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"/> </picture> <figcaption class="caption"><p>Ikuchi island is the largest lemon origin in Japan. They are delicious. And they are proud of it.</p> </figcaption> </figure> <figure> <picture> <source class="responsive-img-srcset" srcset="/assets/img/2025-05-11-asia/japan/setoda-main-street-480.webp 480w,/assets/img/2025-05-11-asia/japan/setoda-main-street-800.webp 800w,/assets/img/2025-05-11-asia/japan/setoda-main-street-1400.webp 1400w," type="image/webp" sizes="95vw"/> <img src="/assets/img/2025-05-11-asia/japan/setoda-main-street.jpg" width="100%" height="auto" style=" object-fit: contain; max-height: 50vh; " data-zoomable="" loading="lazy" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"/> </picture> <figcaption class="caption"><p>Main street on Ikuchi Island. Absolutely nothing opens on Wednesday.</p> </figcaption> </figure>]]></content><author><name></name></author><category term="photo"/><category term="photo"/><summary type="html"><![CDATA[Taiwan Cute foo dog in from of the 大稻埕慈聖宮, there is a pork trotter noodle shop that’s incredibly dilicious and nofrill.]]></summary></entry><entry><title type="html">Reflection on finding a new job after failed at being a founder</title><link href="https://kyleygao.com/blog/2025/job-search/" rel="alternate" type="text/html" title="Reflection on finding a new job after failed at being a founder"/><published>2025-03-17T00:00:00+00:00</published><updated>2025-03-17T00:00:00+00:00</updated><id>https://kyleygao.com/blog/2025/job-search</id><content type="html" xml:base="https://kyleygao.com/blog/2025/job-search/"><![CDATA[<p>After a <a href="/blog/2025/retro-on-my-first-founding-attempt/">failed attempt to start a company</a>, I was, for the 4th time, active in the job market with 8.5 years of experience under my belt. The search took 5 weeks, during which I chatted with 18 companies, received 3 offers, 3 rejections, was ghosted by 1, and withdrew from the rest at various stages. In terms of outcome, I’ll be joining Anthropic to be a founding member of its Claude Growth team.</p> <p>In this post, I’d like to document my thought process during my job search. I believe this will be valuable to me, if not to anyone else, in the future.</p> <h2 id="what-made-this-job-search-unique">What Made This Job Search Unique</h2> <p>This job search was special in 2 aspects:</p> <ul> <li> <p>Thanks to generative AI, the entire tech industry and software engineering jobs (in their broad sense, encompassing MLEs, DS, and others) are experiencing an extremely rapid paradigm shift. ChatGPT has been around for two years, DeepSeek was hailed as the AI Sputnik moment for the US, and predictions like <a href="https://www.businessinsider.com/anthropic-ceo-ai-90-percent-code-3-to-6-months-2025-3">“AI will write 90% of code in a few months”</a> have been circulated.</p> </li> <li> <p>I recently spent a year as a startup founder, which didn’t turn out as desired. Before becoming a founder, I had understood that such an experience shouldn’t negatively affect my career. However, I didn’t fully comprehend what kind of impact it might have, if not negative. Surprisingly, it led me to join a growth team, which is a significant departure from my previous experience as a backend engineer.</p> </li> </ul> <h2 id="the-impact-of-ai">The Impact of AI</h2> <p>Thanks to DeepSeek, I’ve had some <a href="/blog/2025/one-month-after-deepseek-r1/">thoughts on the trajectory of AI</a>: I believe that the scaling law will persist for some time, even though it may no longer solely rely on scaling pre-training. Consequently, we will likely soon have 2X-5X AI capabilities, if not 10X-100X. Meanwhile, AI’s transformative power runs so deep and wide that forecasting specific developments remains elusive – the only certainty is that we face profound and revolutionary changes ahead.</p> <p>DeepSeek has made the GPT-4/4o equivalent generation of models commodities. These models are already incredibly capable and their costs are rapidly decreasing. This has already prompted foundational model companies to invest more in the application layer (OpenAI’s Operator and Deep Research, Anthropic’s Claude Code and MCP). I predict that 2025 will be a year of significant growth for AI applications, benefiting both consumers and enterprises.</p> <p>Because of these hypotheses, I had 2 expectations for my ideal job:</p> <ol> <li>It must be close to the source and center of AI impact, which means companies that create frontier models (Google, Meta, OpenAI, Anthropic) or companies that develop products that otherwise wouldn’t exist without AI (Perplexity, Cursor, and all sorts of agent companies).</li> <li>I wanted my role to be closer to product development, ideally with direct visibility of users, sales, marketing, in addition to engineering.</li> </ol> <h2 id="how-being-an-ex-founder-shaped-my-path">How Being an Ex-Founder Shaped My Path</h2> <p>For Anthropic, I applied for a role in the Labs team, which is responsible for prototyping new applications of Claude models, again, as an investment into the application layer. Both MCP and Claude Code were invented by the Labs team. It sounded like a great match given my interests in exploring AI applications and my recent experience as a founder, which involved quickly prototyping and validating new ideas.</p> <p>For some reason, by the time my interview got to the hiring manager stage, I was instead being evaluated as an engineer for the Claude Code team, which didn’t seem to be a good fit given my lack of experience in developer tooling. Then, to my surprise, it was suggested that I could be a good fit for the Claude Growth team, particularly because of my experience as a founder.</p> <p>I’ll admit—I initially approached the Growth team opportunity with skepticism. My misconception painted growth engineering as a peripheral function focused on superficial product tweaks rather than core innovation. But deeper research completely transformed my perspective. Growth sits at the powerful crossroads of product strategy, engineering, and marketing—precisely the nexus I needed to explore. The strategic skillset developed in this role addresses exactly what I found myself lacking as a founder. While I plan to write a dedicated post on why growth teams are crucial (especially for consumer applications), I’m now convinced this role provides the perfect foundation for my future entrepreneurial ambitions.</p> <p>After the decision, a friend asked whether I regret not staying in the deep learning research field; otherwise, I might be able to work directly on the models as a researcher. It’s a question I asked myself several times, and I think the answer is no. Well, sure, I may be able to have higher pay if I had kept doing deep learning research, but the fact that I chose not to do a PhD at CMU and to have left the research field and moved closer to product shows where my interests really lie. Plus, how could anyone, except the visionary/lucky few people, anticipate the ChatGPT moment? What if Gen AI didn’t break through? My long-term goal, as it has been clear for the last 5 years, is still founding my own company, and I’d like to think I’m on the right track.</p> <h2 id="additional-reflections">Additional Reflections</h2> <h3 id="support-network">Support Network</h3> <p>I received numerous helps and supports from my network, including past colleagues, friends, recruiters, investors, not only for the job search itself but also for emotional support, backchanneling, and for just having intellectually interesting conversations. I’ve never been the best at extending and maintaining networks, but it’s reassuring to know that people are willing to help out just because I was trying to be the best version of myself.</p> <h3 id="the-evolution-of-interviews">The Evolution of Interviews</h3> <p>As I have become more senior, I can feel that I’m assessed A LOT more on behavioral rounds, more on the system design rounds, and a bit less on the coding rounds. Most coding and design questions are practical and relevant to the hiring role. I didn’t do a ton of preparation, so I was glad that I was able to solve/answer most of them purely based on my own experience. It pays off to stay hands-on and to stay curious about how things work.</p> <h3 id="location-and-remote-working">Location and Remote Working</h3> <p>I was quite tired of working from home for 5 years and was particularly looking for companies with a physical office in NYC and with required in-person time. I noticed many people and companies have the same preference, especially for overachieving people who view their career as a high priority. NYC has a good amount of options for both VC-backed startups and big companies, despite still being second to the Bay Area and maybe Seattle too. Not surprisingly, most NYC startups are in the financial service adjacent fields, but there are also some good ones in sales tech and cybersecurity.</p>]]></content><author><name></name></author><category term="article"/><category term="tech"/><category term="startup"/><summary type="html"><![CDATA[After a failed attempt to start a company, I was, for the 4th time, active in the job market with 8.5 years of experience under my belt. The search took 5 weeks, during which I chatted with 18 companies, received 3 offers, 3 rejections, was ghosted by 1, and withdrew from the rest at various stages. In terms of outcome, I’ll be joining Anthropic to be a founding member of its Claude Growth team.]]></summary></entry><entry><title type="html">Retro on my first founding attempt</title><link href="https://kyleygao.com/blog/2025/retro-on-my-first-founding-attempt/" rel="alternate" type="text/html" title="Retro on my first founding attempt"/><published>2025-02-27T00:00:00+00:00</published><updated>2025-02-27T00:00:00+00:00</updated><id>https://kyleygao.com/blog/2025/retro-on-my-first-founding-attempt</id><content type="html" xml:base="https://kyleygao.com/blog/2025/retro-on-my-first-founding-attempt/"><![CDATA[<p>In January 2024, I quit my job at Abnormal Security and co-founded a startup as CTO.  Having worked on it for a year, I left the company in Jan 2025.  In this post, I’m not going to mention names and specifics, given that as of writing, the company and the product still exist, being operated by one of the cofounders.  (Obviously those who are curious can easily trace to the specifics.)</p> <p>Between having a question mark on my resilience when things get hard and admitting the sunk cost in favor of moving on to better opportunities, given my judgement on the trajectory of the company, I chose the latter.  This decision made a deep cut in my identity and created some emotional turmoil, which even had collateral damage on my family, but I’ll focus on the more factual and technical aspects in this retro, leaving the emotional aspects for the future with more time for reflection.</p> <p>While there are numerous details that could be mentioned, I’m going to focus on the highest level.  Because, let’s say we need to go from A to B separated by long distance, the high-level choices determine whether you walk, drive a car, or fly an airplane, whereas the small details are just optimizations like whether you are driving a Toyota or Ferrari.</p> <h2 id="background">Background</h2> <ul> <li>I met the other 2 cofounders through a VC program that tries to capture pre-idea, pre-cofounder opportunities.  One of them was PM and she took the CEO role, I took CTO, and the other one was a UI/UX designer and was responsible for design and user study.  We are all in different cities and weren’t willing to relocate for this project, so we’ve been fully remote for the whole time.</li> <li>The problem we tried to solve was web browser tab overload as well as the deeper needs behind the symptom.</li> <li>We created an AI assistant as a Chrome extension that helps users to stay organized with their tabs. </li> <li>By Feb 2025, we’d gotten 10k downloads on Chrome Web Store, ~40% of which came from paid ads over &gt;10 months, ~50% came from 1 month of being highlighted in the home page of Chrome Web Store, and the rest ~10% from various other channels.</li> <li>We adopted a freemium and subscription model with some naively packaged paid features.  It resulted in some okay but not great free-to-paid conversion rate.</li> <li>Cash burn was primarily spent on salaries, and secondly on Meta ads.  The tech infrastructure costs pretty little even including the LLM cost, which could be paid by the minimum revenue generated.</li> <li>We had the long-term plan to become a cross-platform and cross-app organization tool.  Think organizing browser tabs from Chrome on your laptop and Safari on your iPhone, and think a folder where you can visually organize files, calendar events, messages, and web pages related to a specific project or topic.  Even better, most of the grunt organization work is done automatically by AI.</li> </ul> <h2 id="what-went-well">What went well</h2> <ul> <li>I’d been wanting to start something for a few years and had been intentionally preparing for it, e.g. going to conferences and joining communities that attract startup people.  I connected with the guy who introduced me to the VC program by talking to another random guy at a Web3 conference.</li> <li>Working on a consumer product is super fun.  Your user is your buyer (as opposed to the decision-maker in B2B SaaS).  The feedback loop is super fast; I can fix a bug for some user while having a call with them, and I can get feedback on a new feature as soon as it’s shipped.  No matter what you do, there will be users who sing praises and who curse at you – it’s interesting to read the feedback and to strategize whether and how to respond.</li> <li>I was able to develop and operate the entire tech stack (shown below) with some help from a junior full-stack engineer, as I wasn’t very familiar with frontend development back then.  At the end, I became pretty fluent with the frontend part as well.  I was pretty proud that there was no incident during my 3-week trip back to China, and the product seems to be running okay without me now that I’ve left the company for a month.</li> </ul> <div class="row justify-content-center"> <div class="col-md-10"> <figure> <picture> <source class="responsive-img-srcset" srcset="/assets/img/skipper-architecture-480.webp 480w,/assets/img/skipper-architecture-800.webp 800w,/assets/img/skipper-architecture-1400.webp 1400w," type="image/webp" sizes="95vw"/> <img src="/assets/img/skipper-architecture.png" class="img-fluid rounded z-depth-1" width="100%" height="auto" style=" object-fit: contain; max-height: 50vh; " loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"/> </picture> </figure> </div> </div> <ul> <li>Besides the frontend development, it was also rewarding to be able to have hands-on experience on some stuff that I’d only had some superficial understanding of. Some of these experiences were actually helpful in my recent job search interviews. <ul> <li>Operate a Postgres instance with RDS: scale up instance type and storage, create a multi-az failover instance for near-zero downtime upgrades, deal with metrics like IOPS/EBSIOBalance/EBSByteBalance, create read-replica, snapshot, backup, snapshot exports for analytics, etc., etc.</li> <li>Integrate and use a full-fledged product engineering platform including feature flag, A/B tests, product analytics, session recordings, etc.  We primarily used Posthog, which has all of the features and saved me a ton of time.</li> <li>Set up an observability stack with self-managed Prometheus and Grafana, plus CloudWatch and PagerDuty.</li> <li>Payment and CRM integration with Stripe and HubSpot for things like product marketing emails, influencer referral program, and of course, self-service subscription management.  These work areas are not intellectually difficult but very detail-oriented.  The hard thing is to remember how things are set up and not mess them up in future changes.</li> <li>Growth oriented product features like sharing, link unfurling, etc.</li> </ul> </li> </ul> <h2 id="what-didnt-go-well">What didn’t go well</h2> <ul> <li>Overall, in the hindsight, the decision to jump into this project was reckless. It was a <a href="https://en.wikipedia.org/wiki/Charlie_Munger#Lollapalooza_effect">lollapalooza effect</a> of multiple psychological biases. For example, the preparation I’d been doing created an inconsistancy avoidance bias. There were mere-association bias related to the background of my cofounders and the origin of the project (from a CMU lab). Not to mention the social proof bias from all the startup related social media posts.</li> <li>Not having a previous relationship with the cofounders themselves is already a big risk.  That plus being fully remote is almost guaranteed to fail.  You can even argue that the fact we didn’t want to relocate for the company was a sign of a lack of commitment.  Next time around, I’d spend more time pressure testing the cofounders no matter how good they look on paper or unless I’ve personally worked with them (and invite the same rigor on myself) and insist on being co-located to start with.  When the shits hit the fan, the most important thing is that the cofounders are still willing to stick together, which is a very strong requirement if you just compare it with other types of relationships like dating vs marriage.</li> <li>If you have disruptive technology (e.g. ChatGPT), you can create a market that didn’t exist before (AI chatbot).  Very few companies are fortunate enough to own a technology like ChatGPT, and in contrast, they have to make sure the market already exists – the best signal for that is there are already many big companies working on it.  Of course, you also need a novel solution (e.g. Notion for notes/docs), a better execution (Uber among tons of other shared rides attempts), or better distribution (Teams vs. Slack).  For our product, there were only a few super small companies doing similar things, a bad sign!</li> <li>GTM, sales, distribution, growth, however you call it, is super important.  I’d argue that it’s more important than the product, unless you have a disruptive product like ChatGPT.  And within growth, for a consumer product, organic growth is more important than paid growth.  The ads platforms today are pretty expensive; you are doing great if you can get &lt;$10 CAC, but that is still a lot for the unit economics to make sense.  So you have to do as much organic growth as possible to spread the cost of paid marketing.  If you can organically grow as many users as your paid channels do, you can halve the CAC for the marketing spend.  This is why you see all the founders (and Elon Musk) shit post on social media – it doesn’t even matter what their personal image is as long as their posts create enough traffic to their products.</li> <li>Given above, you can’t have all the cofounders working on product development.  Sure, at the very beginning, everyone works on the product, but at some point, the growth, along with other responsibilities like investor communication, becomes a full-time job.  In my previous jobs, I didn’t understand why the CEO could be so remote to product development, but now I understand.  They are doing important work that only they could do (assuming there’s a competent cofounder leading product development).</li> <li>If something is business-critical, one of the founders has to personally work on it, even if they didn’t have previous experience with it, before delegating to external help or hires.  Non-founders almost never care as much as founders and have as much context as the founders do.  On the flip side, founders can’t effectively delegate and ensure quality without doing it themselves.</li> </ul>]]></content><author><name></name></author><category term="article"/><category term="tech"/><category term="startup"/><summary type="html"><![CDATA[In January 2024, I quit my job at Abnormal Security and co-founded a startup as CTO.  Having worked on it for a year, I left the company in Jan 2025.  In this post, I’m not going to mention names and specifics, given that as of writing, the company and the product still exist, being operated by one of the cofounders.  (Obviously those who are curious can easily trace to the specifics.)]]></summary></entry><entry><title type="html">DeepSeek R1发布一个月后的想法</title><link href="https://kyleygao.com/blog/2025/one-month-after-deepseek-r1/" rel="alternate" type="text/html" title="DeepSeek R1发布一个月后的想法"/><published>2025-02-16T00:00:00+00:00</published><updated>2025-02-16T00:00:00+00:00</updated><id>https://kyleygao.com/blog/2025/one-month-after-deepseek-r1</id><content type="html" xml:base="https://kyleygao.com/blog/2025/one-month-after-deepseek-r1/"><![CDATA[<blockquote> <p>DeepSeek R1发布快一个月了，各方观点都发表的差不多了，写写我的想法</p> </blockquote> <p>DeepSeek-V3和R1发布，颠覆了美国科技圈三个认知：</p> <ol> <li>中国做不出独立创新性成果</li> <li>Scaling law需要指数级的增加算力</li> <li>前沿大模型一定要是闭源的</li> </ol> <p>关于第一点，只能说average美国人对美国以外的世界了解还是有限。前几天HackerNews有一篇<a href="https://news.ycombinator.com/item?id=42871371">关于DeepSeek的一个Clickhouse数据库节点泄露的帖子</a>，文章里有一些数据库schema的截图（比如<code class="language-plaintext highlighter-rouge">log_stream</code>, <code class="language-plaintext highlighter-rouge">metrics_stream</code>)，下边有这么一条<a href="https://news.ycombinator.com/item?id=42872867">评论</a>让我啼笑皆非。中国还是经济落后的第三世界啊，怎么可能会读写英语，怎么可能作出创新的大模型。</p> <div class="row justify-content-center"> <div class="col-md-6"> <figure> <picture> <source class="responsive-img-srcset" srcset="/assets/img/one-month-after-deepseek-r1-01-480.webp 480w,/assets/img/one-month-after-deepseek-r1-01-800.webp 800w,/assets/img/one-month-after-deepseek-r1-01-1400.webp 1400w," type="image/webp" sizes="95vw"/> <img src="/assets/img/one-month-after-deepseek-r1-01.png" class="img-fluid rounded z-depth-1" width="100%" height="auto" style=" object-fit: contain; max-height: 50vh; " loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"/> </picture> </figure> </div> </div> <p>吐槽过第一点，第二点和第三点究竟是否应该被推翻，我觉得是有待论证的。</p> <h3 id="需要指数级的增加计算资源吗">需要指数级的增加计算资源吗</h3> <p>首先，对计算资源的崇拜的共识是过去十年甚至三十年里被反复验证的，强化学习先驱Rich Sutton在2019年的短文<a href="https://www.cs.utexas.edu/~eunsol/courses/data/bitter_lesson.pdf">The bitter lesson</a>里就简明扼要的指出AI能力的质变几乎每次都来自于更多的计算，而不是通过优化带来的量变。</p> <p>从另一个角度来讲，既然scaling law还没有显著的见顶，那买显卡堆计算就主要是一个经济项目。只要ROI说得通，就有金主愿意砸钱。相对的，DeepSeek V3/R1主要解决的是在给定计算上限（被美国卡脖子）的条件下，怎么最大化训练和推理效率的优化问题。做优化是一个科研项目，科研项目的不确定性更大，对人才的要求更高(DeepSeek的研究员牛逼！)。科研问题还不能证伪，如果一个团队投入了大量人力物力还是没做出成果，也不能说这个成果是做不成的。在OpenAI，Anthropic，Google的研究员们都是经过过去二十年的学术熏陶，一篇一篇paper做到现在的，难道不明白有很多优化可以做？只不过大家形成了共识，先把买显卡堆计算这条路推到极点，发现能发现的所有<a href="https://arxiv.org/abs/2206.07682">emergent abilities</a>，然后再回头做优化也来得及。所以五千亿美元的stargate还是会进行下去，Nvidia的股票还能再长。</p> <p>当然了，这是从纯科研的角度想问题。DeepSeek厉害的一点是让模型的推理阶段成本更低，需要的硬件资源更少，正投资本家们的下怀，所以我们才看到微软和阿里这些自己有自研大模型的公司争先恐后的部署DeepSeek，以及最近的<a href="https://mp.weixin.qq.com/s/V8Hri4WSVJWgp8IcKxlIvw">微信接入DeepSeek</a>。这个事情通过Jevons Paradox反过来又会继续刺激需求端的暴增，继续利好Nvidia。</p> <h3 id="前研大模型要开源吗">前研大模型要开源吗</h3> <p>先说结论，我认为现有模式的开源模型（只给模型参数，不给训练数据和过程代码），主要的价值来自于营销推广，长期的商业护城河还有点观察。（这里我们只讲商业策略，不讲安全性问题，不管是不是真心实意，OAI和Anthropic选择闭源都说是为了安全）。</p> <p>DeepSeek这一波操作，一下子破了圈，手机app冲上北美app store第一名，各大云厂商提供deepseek的api，产品接入，全民皆知。是极致的市场营销。这其中最重要的，我认为是增进了品牌价值和一部分消费者端的粘性，但可能没有构成很大的护城河。</p> <p>我为什么这么说？如果几个月后任何一个其他的公司发布了一个超越V3/R1能力并且成本更低的大模型，会发生什么：</p> <ul> <li>企业端 <ul> <li>公司会集体切换到新的模型，如果是开源模型就自己部署，如果是闭源模型就上api</li> <li>已经微调过模型的企业会基于新的模型重新做微调。</li> </ul> </li> <li>消费端 <ul> <li>一些用户因为已经用了deepseek，会懒得切换到新的应用，但也有一些用户会马上切换</li> <li>模型wrapper应用要么会把新的模型加入到可选模型里，要么就在后台直接切换到新的模型 这体现出一个问题，就是大模型之间没有适配性差异，从一个模型切换到另外一个模型，对企业和对消费者的成本都非常低。</li> </ul> </li> </ul> <p>让我们来想想传统软件开源的优势是什么？</p> <ul> <li>一是让更多的人参与到软件的改进中，比如说PyTorch开源后获得了学术界和业界非常多的贡献。</li> <li>二是通过接口和行为的依赖让对用户进行锁定，然后在免费开源核心的基础上提供收费服务（主要是managed cloud hosting为形式）以获取利益，这个商业模式在过去十年重复太多遍了。典型的例子比如Gitlab, ElasticSearch, MongoDB, Kafka (Confluent)等等。这个模式成功的核心在于，一旦企业使用了一个开源项目一段时间，会对这个项目的借口和行为形成依赖，导致迁移到新解决方案的成本非常高，竞争对手必须要提供超出迁移成本的优越解决方案才有机会说服企业作出改变。可以说是一个非常聪明的商业模式。 反过来看，开源大模型是否能利用这两个开源的优势还很难讲：</li> <li>一，只开源参数不开源代码和数据，社群是没有办法帮助优化大模型本身的。微调过的开源模型可能可以通过huggingface这样的平台造成一些复用，但还是一样的问题，社群怎么来对开源微调模型做贡献呢？</li> <li>二，大模型之间的迁移成本非常低，模型上侧接口都是一样的，下侧的部署和计算资源的调配也差不多，所以<strong>在模型行为差不多的时候</strong>迁移成本非常低。另外，怎么在免费提供的开源模型参数上开发收费的服务好像还没有什么人在尝试，是一个未知的问题。唯一知道的是像api调用这样的服务，通过云厂（火山）或者专业的模型服务厂（硅基流动）比模型厂自己更靠谱。</li> </ul> <p>这里好像有一个突破口，有两个部分，非常难，但是可能可行：</p> <ol> <li>保持模型在某一个特殊的能力上一直领先</li> <li>围绕这个模型能力开发传统软件方案，和模型深度纵向集成</li> <li>这两个部分都开源 比如说，我的模型编程能力特别强（比如Sonnet），在迭代中我要一直保持它的这个能力行业第一，开源。然后基于这个能力，开发一个AI编程平台（类似Gitlab，Graphite，Cursor，等等），也开源。然后就可以在这个平台上设计收费功能，利用传统开源软件的优势了。</li> </ol>]]></content><author><name></name></author><category term="article"/><category term="tech"/><category term="ai"/><category term="china"/><summary type="html"><![CDATA[DeepSeek R1发布快一个月了，各方观点都发表的差不多了，写写我的想法]]></summary></entry><entry><title type="html">Photo - Mina Perhonen Kyoto (皆川明京都)</title><link href="https://kyleygao.com/blog/2024/photo-minaperhonen-kyoto/" rel="alternate" type="text/html" title="Photo - Mina Perhonen Kyoto (皆川明京都)"/><published>2024-11-28T00:00:00+00:00</published><updated>2024-11-28T00:00:00+00:00</updated><id>https://kyleygao.com/blog/2024/photo-minaperhonen-kyoto</id><content type="html" xml:base="https://kyleygao.com/blog/2024/photo-minaperhonen-kyoto/"><![CDATA[<p>皆川明在京都的旗舰店坐落在一个很有历史感的建筑里面，我站在里面有一种非常熟悉的感 觉。想起来跟大连的一些日本建筑有类似的调性，好像一下穿越到了大连中山医院那些长长 的走廊和楼梯间里。</p> <p>Mina Perhonen’s store in Kyoto is located in a historical building that caused some nostalgic feelings. It reminded me about the Japanese architectures in Dalian, especially the long corridors and staircases in the Zhongshan Hospital.</p> <div class="row justify-content-center"> <div class="col-md-6"> <figure> <picture> <source class="responsive-img-srcset" srcset="/assets/img/mina-perhonen-kyoto-480.webp 480w,/assets/img/mina-perhonen-kyoto-800.webp 800w,/assets/img/mina-perhonen-kyoto-1400.webp 1400w," type="image/webp" sizes="95vw"/> <img src="/assets/img/mina-perhonen-kyoto.jpg" class="img-fluid rounded z-depth-1" width="100%" height="auto" style=" object-fit: contain; max-height: 50vh; " loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"/> </picture> </figure> </div> </div>]]></content><author><name></name></author><category term="photo"/><category term="photo"/><summary type="html"><![CDATA[皆川明在京都的旗舰店坐落在一个很有历史感的建筑里面，我站在里面有一种非常熟悉的感 觉。想起来跟大连的一些日本建筑有类似的调性，好像一下穿越到了大连中山医院那些长长 的走廊和楼梯间里。]]></summary></entry></feed>