# Centum Rapid Test FY2 S3 V2 # Plan ## Enterprise ### Thesis Developers are telling people that Generative AI models can do anything for them. But coming up with those anythings to ask is a stressful friction. Others are focusing on giving great answers in response to specific questions, but I think what people really want is getting great topics to think about when they have time and curiosity to think about anything. If we use our Generative AI to curate topics for each person, then we will be able to give people the joy of discovery and become a trusted information distributor. Whereas Centum is a simple idea with a great business opportunity, we still expect that we will be able to successfully take then defend our position in the market, because operating Centum successfully demands not just doing many things correctly, but also not doing many things correctly. I am confident that we have what it takes to be the single best team to be operating this business in our society. ### Criteria 24.06.03: If Centum reaches 100 daily active visitors, then we will note that people desire Centum. - Metric: Daily Active Visitor - 24.06.03: visited 5+ out of 7 days AND viewed 10+ card - 24.06.01: visited 5+ out of 7 days ## Product ### Centum Minimum Yesterday, I spent yet another entire day working on Centum. Even during breaks, I called my parents, then my girlfriend, and presented to them what I am thinking. I ended the day and got to bed at 11PM, but I was still thinking about Centum for 3 more hours. Then at 2AM, I got struck by a sudden idea, jumped out of bed, and ran to my desk to draw out a wireframe in Figma. It took me mere minutes. Then I had in front of me, Centum Minimum, the distillation of all the joy of learning in the purest possible form. Centum Minimum is a single page, infinite scroll, content site for the intelligentsia that presents interesting information about interesting topics in information cards that are generated in response to the interactions of the visitor. Enter junwon.com. Then the parade of cards begins. One logo at the top, then only cards, some opened, others folded, and ads in between to finance the business operation. Any feature that must be in the site but not on the front page will be hidden inside the logo which is a button. ### Wireframe ![Centum Minimum 1.png](Centum%20Rapid%20Test%20FY2%20S3%20V2/Centum_Minimum_1.png) ![Centum Minimum 2.png](Centum%20Rapid%20Test%20FY2%20S3%20V2/Centum_Minimum_2.png) ![Centum Minimum 3.png](Centum%20Rapid%20Test%20FY2%20S3%20V2/Centum_Minimum_3.png) Visitor is presented with interesting information immediately. Some cards. Interactions, such as which cards the visitor decides to read, skip, open, and fold, informs AI what is interesting and not interesting to the visitor. Advertisements is an appropriate monetization method for the infinite scroll content feed format, but we will try to test monthly membership fee. ## Technology ### Specification - Infinite Scroll - Rapid Science Enable rapid experiments in infinite scroll. ### Technology Stack - AI Models: OpenAI GPT, Google Gemini, Anthropic Claude, Meta LlaMA, Mistral. - For shell and server, Tamagui and Supabase. - Other tools include LangChain. - Assessed or Considered but Discarded Microsoft PlayWright, LangFlow. ## Distribution - Post interesting facts about interesting topics on 3P social media. Get topics from me, Namu Wiki, Wikipedia, 3P social media, and our Junwon app. - WiFi SSID at Places and Events such as Stanford Shopping Center and Stanford Commencement Ceremony. - Community Forums on the internet such as the Reddit community for skincare. ## Finance - In which ways can we earn income? - What are all the costs? - Which incomes and which costs grow at which pace with respect to the scale (based on the metric “daily active visitors”)? # Execution ### Catalog Topics ### Version Log - 2.6.8: 24.06.01: Topic Discovery Client App Shell, Server Database, AI Contents. - History Started developing on 24.05.25. Published version on 24.06.01. Developed on 1X MacBook, 2X Windows computers with NVIDIA RTX GPUs, Samsung Galaxy S24U, Samsung Galaxy Fold Z5, iPhone 13 Pro Max, iPhone SE, 10-inch E-Ink reader, 6-inch E-ink reader. - 2.6.9: 24.06.08: Topic Elaboration - Upgrade Centum AI for 2.6.9 - Bullet Points - Emoticon Bullet Points - Add Intelligence Engine Google Analytics, Vercel Analytics, Statsig. Cut Microsoft Clarity due to console warning about tracking cookie. Chat with Junwon. - Elaboration - Server Database - Contents AI - Elaboration - CUT features - Topic Generation: Until we reach the first 10,000 topics, curate manually. - Front Page Discovery Content Cards Curation: Until we reach 10,000 DAU, curate manually. - 2.6.10: 24.06.12: Topic Links - Links - CUT: Prefer links to topics in catalog Considered preferring topics that exist in catalog (e.g. If catalog has Stanford but not Oxford, then cards on Harvard topic to prefer referencing and linking to Stanford). But decided to not do that. This is to 1) prefer more naturally related links, and 2) save cost (not only our money but also visitor time) by preventing another database query. - 2.6.11: 24.06.16: Topic Search - Search - 2.6.12: 24.06.22: Centum for All - Separate out Centum Writer and Centum Editor. - Publish mobile apps for iOS and Android. - Visitor Guide Card ### Technology Documentation [Technology Documentation FY2 S3](Technology%20Documentation%20FY2%20S3%20339b3baa2d464b3cbe406a6cbd81fbca.md) ### Academic Research [Academic Research in Centum RT FY2 S3 V2](Centum%20Rapid%20Test%20FY2%20S3%20V2/Academic%20Research%20in%20Centum%20RT%20FY2%20S3%20V2%20ec60d232bef24e068d8e6b2e44c9443f.md) ### Estimations - Catalog Size It takes a person 22 minutes to study 5000 words. To provide enough contents for that person on Centum for 1H per day for a year, Centum needs 550K words (assuming that every word is chosen to be studied and not skipped). If we prepare 1K words per topic, we need 550 topics. This number is very small. In comparison, we already have 2K topics that are presentable in production Centum. The cost is also trivial. Already at our current level, we can generate contents for 1 topic (approximately 1K words) in 2 minutes, which means we can generate contents for 5K topics in a week, and this generation is currently getting done just on my laptop in my home. The production database storing the 2K topics in production is also only 150MB. On a positive note, this means cost of satisfying people is affordable. On the other hand, this means we need to keep making our product better and better, because anyone can come in and challenge us even after we take a position in the market. Overall, I am optimistic about this finding. Another way to estimate size of catalog to generate is referencing Wikipedia. Wikipedia has 6.8M articles, 4B words. New topics added to Wikipedia peaked 2 decades ago which was 5 years after it was founded, and about 5% of Wikipedia articles have been updated each year. This means Wikipedia is already a superset of topics that people take interest in. Wikpedia is also verbose for ordinary consumers. I think 1M topics with 1B words will be plenty to satisfy most cases, so we can plan to prepare that much, then let community address the remaining cases. Generative AI price varies widely in orders of magnitude, but using a magnitude I consider to be reasonable ($1 for 1M tokens), we will be able to prepare the baseline catalog with an upfront capital investment of $10K. We can easily add capital efficiency by targeting specific scenarios (e.g. food recipes, traveling in California) such that we only need a subset of the topics (e.g. 10K topics for $100), then finance growth using income from the first version. # Review Centum Rapid Test FY2 S3 V2 was a great success. Centum in production grew to become a great product that I am enjoying using everyday and that people around me are revisiting repeatedly without any reminders from me. Friends are messaging me with unsolicited feedback for the product, describing what they like and what more they want, and this phenomenon gives me confidence in Centum. I was also able to dream what I shall grow Centum to be in the future. In this section, I will review the insights I got while working on Centum for the past several weeks. These insights are directly shaping Product Plan in [Season Plan FY2 S4](../FY2%20S4%20(24%2007%2001-24%2009%2022)/Season%20Plan%20FY2%20S4%20b19e5dc1d4c640e7812c74b713377060.md) which is starting at the exact moment at which this test (Centum Rapid Test FY2 S3 V2) is ending. Centum has already become the best tool for going down the wiki rabbit hole. Centum visitors get fed interesting facts on interesting topics as soon as they just open the app and start scrolling. On Centum, visitors can relax and enjoy, unlike in chats with Generative AI models (such as ChatGPT), where consumers are asked to start and guide the conversation. Visitors can also choose to start and guide the rabbit hole journeys on Centum using its search feature and the links feature. Most importantly, Centum was made for wiki rabbit hole sessions. Its dark mode look is easy on the eyes at night in bed. Its topics are selected to be interesting. And Centum writes its contents to be fun (unlike Wikipedia which writes in boring difficult ways). Centum not only attracts the smartest people in the population, it also makes its customers become the smartest people at social events. During conversations, people who have been on Centum will know interesting facts on many different topics. I predict that Centum will also grow virally. People love to share what they know with other people, so Centum customers will feel compelled to share what they know during conversations with other people, and the impressed audience will ask them how they know such interesting information, and they will introduce the audience to Centum. Centum will be a positive power that elevates the intelligence of all of humanity. ### Competitive Advantages - Unlike Wikipedia, Centum is fun and Centum can generate contents for any topics. - Unlike Pinterest or Instagram or YouTube which are uncomfortable to look at in darkness, Centum is comfortable to view in bed. - Unlike ChatGPT and other Generative AI chat assistants which push to customers the frustrating cost of coming up with topics to chat about, Centum comes up with topics for the customers. - Unlike TikTok, Centum supports visitors to guide the journey actively using search (on top of passive journey using recommended contents). - Unlike Podcasts, Centum supports skipping over uninteresting content. ### Gap Between Done and Planned - Metrics I consider metrics to be important, because I consider learning to be important. However, I did not get to add the metrics I planned to add, because I instead prioritized development of features. At the moment, developing features was more important, because Centum lacked minimum necessary features to be usable for known use cases for me and my immediate contacts, and adding metrics was less important, because Centum was not yet visited by many visitors who are not able to give me feedback directly. The metrics I planned to add are following: - Finance: Revenue - Graceful Examples and Disgraceful Examples - Distribution: Traffic Origin - People: # (and % if applicable) of daily visitors, visitors who visited 5 out of past 7 days, - Time spent per visitor per session, per day - Interactions: # (and % if applicable) of interactions per visitor: visits per day, visited days over past 7 days, cards that are studied, cards that are skipped, link clicks, searches topics - Shell Technology: Page Load Time - AI Technology: Time took to generate contents for each topic - Catalog: # of presentable topics, word count per card, links per card - Topics that are most searched, most studied - Disgrace: Searches ending up in Missing Topic case - Offline Datasets I planned to import information from offline datasets such as Wikipedia and Namu Wiki. This work item fell in priority, because Centum in production, which is using a Generative AI model without any separately indexed offline datasets, demonstrated ability to relay fun and accurate information, instead of suffering greatly from Generative AI “hallucinations”, beating original expectations. I still think importing offline datasets will meaningfully enhance Centum Catalog. I do need to prepare a better justification for investing in this work item. ### Monetization Challenge Originally, we planned to monetize Centum using ads. Currently, we are facing an obstacle. Google AdSense is refusing to display ads on Centum. As rejection reason, Google described Centum as a site with “thin contents”. After searching about the term, I learned that Google fears sites that prepare thin contents on a wide variety of topics, because such sites will be structurally preferred by search engine users and therefore search engine ranking algorithms. This news is bad for us in the short term and great for us in the long term. In the short term, we need to find a different way to monetize Centum (or try again to get ads supply from Google AdSense). In the long term, they are correct in that we are structurally positioned to perform strongly on search engines and be preferred by search engine users. ### Roadmap At this moment, I am planning the following 2 versions to be important upgrades. - 2.9: Centum for Each The current version, 2.6, can be described as Centum for All, because any person can visit junwon.com and get interesting facts about interesting topics the way I intended for all visitors. However, Centum for All is like Wikipedia or New York Times in that it shows the same information to all visitors, including a visitor who has visited us for the first time and a visitor who has visited us every single day, multiple times a day, for the past many weeks. Now, this is clearly a problem. The next version, 2.9, will be Centum for Each. I will add Accounts and Profiles, and Centum will get better and better for each person as each person keeps using Centum (let alone each person using Centum making Centum better for all other people too). While a member is using Centum, Centum will learn more about that member, then curate topics that will be interesting to the member (e.g. “Lavender” for a member who has been interested in flowers), and even generate information tailored to that person (e.g. “Lavender: Tea” for a member who has been interested in teas and opened Lavender topic). Centum for Each will be defensible, because newer entrants with similar technology will still fail to reach parity utility in comparison to Centum for Each which has become better for each member over time. - 3.0: Only on Centum To take our own position in the economy and to make our own contributions to it, in other words, to make Centum be the “Catalog of Humanity”, we need to secure information that people can find only on Centum. For that, we need people, not AI, not technology, but people, to come to us and add information to our catalog. In this version, we will be introducing features that are specifically for the people who will be producers in our information economy. I am also planning to steer Centum to address customers who are homebodies and enjoy shopping on the internet and at malls. Without delay, we will start working on this plan in [Season Plan FY2 S4](../FY2%20S4%20(24%2007%2001-24%2009%2022)/Season%20Plan%20FY2%20S4%20b19e5dc1d4c640e7812c74b713377060.md).