Hey everyone! This is an ML Chess variant I built when going through fast ai. It made it to the top of the Hacker News front page and generated a ton of great discussion from the ML community.
It all started here with the learnings from this course so I thought I’d share the story in the hope of inspiring new fast ai learners. If you’re into Chess, ML, math or puzzle games, I think you’d enjoy this read:
I first got inspired by Jeremy’s advice in his great book to go deeper and deeper on one main ML project end to end, instead of continuing to split my focus across a dozen half-baked ML demos where I was barely scratching the surface of a problem space.
After powering through tenaciously for numerous sleepless months, I was humbled to see that what started as a scribble on a piece of paper is now a live ML product being actively used and loved by 10,000+ people in 300 cities around the world.
As you’ll see from the writeup and the thoughtful input it got on HN, this problem space can be equally well suited for several different approaches @jeremy introduces in fast.ai, from Tabular to DL, CNNs, NLP, LLMs, to RL, and so on. I try out several of these with examples in the project writeup - if you think of more suitable alternatives, please do chime in!
For those currently taking fast.ai, I hope the story of Echo Chess inspires you to keep building, stay curious and power through the countless iterations. There are very few things as enjoyable as seeing the power of ML improve something you’ve built with your own two hands.
Happy to answer any questions on your mind. All feedback from the community is welcome.
I recently started fast.ai and also building something from the course. Thanks for sharing your inspiring experience. Being on top of hacker news and 10,000+ users is very impressive! Congrats and thanks for sharing!
Thank you @Chuhao, I’m glad you found the writeup helpful! It’s definitely been a journey Excited to hear more about what you’re building if you’re willing to share. What stage in your project are you at right now - and how can we help as a community?
I built a MVP to help Youtubers to predict the click-through rate of thumbnail. You can try it at oldfatboy.com.
I aim to build a Shopify for creating personal deep learning models instead of websites. A user-friendly interface for Youtubers to build a customised AI model fine-tuned from their own data.
It was first built for my wife. She is a YouTuber with 100k subscribers. And this fine-tuned model really increased her click-through rate! Therefore, I’m thinking of delivering the power of AI to the end user. As Jeremy mentioned in the lecture, deep learning itself is very powerful now. It’s about how to let people who have pains know how to use it.
I like that you have a customizable version for youtubers to train on their own channel data here. Curious if you considered using an API to grab the Youtube studio data automatically for users instead of them emailing you their zip files manually. Are you using Zapier to automate these user requests in the current workflow?
It’s interesting to see that even the basic ‘above average’ vs ‘below average’ CTR class predictions have already helped your wife improve her thumbnail CTRs on her live channel. This feels like a promising problem area to explore further - might be a better fit for a regression approach instead of classification FYI.
Btw Lingzhi’s YT channel looks really cool! PhD bio, expert crocheter, soon-to-be deep learning practitioner. That’s a great combo.
P.S. super minor note: there may be a typo on the homepage you linked ‘it can predict if the CTR’
It’s my personal website. The blogging platform is powered by a European startup called Hyvor. They’re still pretty new and are a little rough around the edges but the customer service has been great so far. I think the founder replies to all questions himself. I chose them because of: (1) flexibility, (2) support, and (3) ease of integration with their snappy commenting tool.
Congrats on the launch. I’ve tried the product and it really is somewhat addictive. Got to level 3 haha.
I totally agree with what you say about going 0 to 1 in a smaller project rather than having a dozen half baked projects.
Personally, I built Hotcheck, with some of the tools and knowledge we’ve seen on the course.
And to my surprise, it went viral on an AI blog and now is being used in 40+ countries.
It’s obviously still a work in progress, but I’d like to ask you what strategies have you been using to distribute your product once a minimal working mvp was ready?
Thank you @santy, that’s great to hear. Congrats on your launch as well!
Growth has been entirely through word of mouth so far. Iterating quickly with the early adopters to go from MVP to a proper v1.0 was key. Have you been able to establish an early community for your product?
I’d say it’s a very important step for capturing user feedback and facilitating organic growth.
Good job reaching level 3 If you’re in the mood for a challenge, feel free to try out level 9+ where you get a squad of multiple pieces to coordinate - or sacrifice.
You can always skip to any level by using the Level dropdown at the top: https://echochess.com/
Ok got it, word of mouth plus generating an early community. I think the product has the word of mouth component as it has spread virally around the world.
But I hadn’t figured out the strategy of building a small community around it, it’s a clever one. Any systems you recommend most (building a twitter account v mailing list v maybe sth else i still haven’t learnt?)
Sure thing. Depending on what kind of ML product you’re building, setting up a Discord server might be a great idea too. There are a lot of community options that might work based on the user personas you’re targeting. What user segment would you say is the ideal fit for your product?
Oh sorry, i didn’t get this notification until today. From what i’ve seen in the last couple of days, many young people/influencers use it before they post something on social media, many people told me they enjoy using it to figure out which is their ‘best’ picture.
In that sense, I guess I should set up a discord server and create a community of young people/influencers in it.
Interesting. A good rule of thumb is usually to go where your users currently are - as in embracing the space in which the users’ activity and conversations are already happening, or where most of your target user base is likely to find you organically.
From what you’re describing*, instagram or tiktok might be particularly relevant to the community you’re after.
*Caveat: I’m personally not active on either of these platforms, so take this advice with a huge grain of salt. At the end of the day, the only feedback that matters is the one coming from your actual users
Hey @cyborgafterall Sorry for the very late reply as I was back in China for some family commitments past weeks.
Thanks very much for kindly reviewing my MVP and pointing out the typo . Zapier is a brilliant suggestion! The process of getting data from YouTubers will definitely need to be more automatic! My MVP is too minimal
Love your discussion with @santy about getting users and building a user community. That’s exactly what is in my mind now. How to get more YouTubers other than my wife to give me feedback on this project? I tried to cold email YouTubers but it doesn’t work well. After reading a blog from Paul Graham, Do things not scale, I decided to attend networking events to recruit my user manually. And I got my first ever user!! He is a YouTuber in Sydney with 150k subscribers. People is surprisingly willing to try my product if I meet him/her in person. Now, I’m going to scrape data manually from his channel, make a customised model and send a personal link (probably oldfatboy.com/his-chanel-name) to him.
I have a crazy idea of organising the first-ever Australia YouTuber Summit in Sydney to let more YouTubers come to meet me. My plan is to make a summit website first and some agenda with fake speakers. See if anyone would sign up for ticket pre-sale. If there is traction, I’ll really invite these speakers.
I want to share with you another exciting news that happened to me last weekend. I went to meet Jeremy as the organiser of Sydney Fast.ai Meetup!! How great is that! He is an amazingly friendly and remarkable person to talk with!
I also met two brilliant fast.ai alumni like you at the meetup with Jeremy and one of them is interested in working together on the current project. I feel it would be great to have a team. Did you build your remarkable Echo Chess just on your own? That would be very impressive considering you also need to do word-of-mouth promotion. Another question is How did you manage to make the top of Hacker News? I tried to make a post but doesn’t work well.
Another idea. I’m thinking about organising an online meetup for fast.ai people to share their projects and learn from each other. Because I feel very rewarded for my Sydney Meetup where members shared their machine-learning projects and helped each other. Would be very cool to meet more fascinated Fast.ai people like you online. I believe many new fast.ai students like me will also find your project inspire their imagination!