Ten Shots at Building AI

One weekend in the fall of 2022, I built my first AI powered side project, a game I wanted to be known as the "Wordle of Generative AI". It used an early version of Stability AI's Dreambooth API, and was a hit among the friends and anonymous internet people who tried it out. One friend even begged me to keep making more puzzles after the 12 original days were up. I happily extended it, integrated it with DALLE, and relaunched it in 2023 and 2024. That project kicked me into the AI rabbit hole.

Over the past 2+ years, I've built and publicly launched 10 individual AI powered apps (many more if you count single day hackathons) with vastly different goals, technologies, and levels of success. Some were for fun or experimentation, and 5 were serious attempts at launching a real business. As I was gradutating MIT Sloan a few months ago, I went through the repos, the dbs, the emails, and my own notes to make sense of it all. Among those 10 apps are tens of thousands of users, thousands in revenue, many painful decisions to move on, and a ton of fun along the way.

These years have been exciting but chaotic for me: I quit my job, I got engaged, we moved to Boston, I went to business school, we had our wedding, and all around us AI was taking over the world. I wanted to take a moment to reflect on each project, how it came to be, and what it led to next. It's always easier to connect the dots looking backwards.

Read along to understand what I've been up to, what I'm most excited about, and probably learn a thing or two.

The 12 Days of AI: 12daysofai.com
LowTech AI: lowtech.ai
Modelle: modelle.ai
Max Compute Co: maxcompute.co
Promptly: (chrome webstore)
Build AI: trybuild.ai
Mail Kiwi: mailkiwi.com
The Turing Fest: theturingfest.com
Front Row Fantasy: frontrowvibes.com
"Async Inference Co": domain taken over by current project

One liner: The Wordle of Generative AI
Inspiration: In 2022 I was experimenting with the stable diffusion API with my friend Charlie Durbin in NYC. It was clumsy and expensive to use through Dream Booth, but in a single sitting we managed to get a webapp with text-image working, these were the days before Cursor. The next weekend, inspired by hearing the Wordle creator sold his project to the New York Times after making a fun game for his girlfriend, I decided to turn our starter project into a game too. The point was to create a newly possible game format by using text to image and share the magic of Gen AI. The game still proudly proclaims "You're about to use generative AI"! Crazy way to phrase things these days.

The launch: I built an initial version in late November 2022, but decided to theme it around Christmas and launch with the theme of "The 12 Days of AI". I shared around with friends and on product hunt, and got hundreds of users!

The game was free to use, but I eventually added in the ability to purchase stickers of the images you generated during gameplay. Not one person bought a sticker, and I eventually removed the buy sticker functionality.

Status: This game lives on and you can still play today! I haven't touched the core gameplay, ux, or styling of the app, but I have slightly updated it for 2023 and 2024. You can still give it a try today.

Play 12 Days of AI

One liner: Dead simple AI tools for non-tech savvy people
Inspiration: The day open AI released the gpt-3.5 API, I wanted to make something with it. A few weeks earlier, a former coworker of mine had started an AI newsletter during the initial ChatGPT craze and ran a "prompt of the day" every edition. I thought that including the text prompt alone was too many steps: copy the prompt, open up ChatGPT, paste it in, delete the templated portions of the prompt, write in your information, and then get a result. To make it simpler, I built a simple way to use templated prompts immediately from the newsletter: click the prompt, land on a webpage with text boxes for the areas to enter, enter in your info, and get a result!

After running it for a few newsletter in March and getting positive feedback, I put that project on the backburner as things accelerated in my life. I got accepted into MIT business school, I turned 28, I got an offer to join a startup as a technical cofounder, I decided to go to business school, and I resigned as the CTO at Sandhill Markets. With a few months free before school, I picked LowTech AI back up to pursue a bigger opportunity than newsletter tech: helping more people in my life take advantage of LLMs. By June of 2023 I was using ChatGPT all the time, but no one around me was! With easy-to-use templated prompts I wanted to allow more people to experience the magic.

The launch: After the initial newsletter launch in march, I re-launched the project in June of 2023. I put us on product hunt (but botched it due to incorrectly planning the date), announced it on Twitter, shared it with friends, and wrote blog posts about the product. I even brought on 10 interns to write content and boost SEO for the site. At MIT in the fall, I hosted a prompt-a-thon with the help of some classmates Clyde, Priya, and Will.

Status: By the end of 2023, I stepped away from the project as my main focus. I struggled to imagine how this could grow to anything more than a wrapper, and I knew I had limited time to focus on building and launching a company during this grad school window. Two years later LowTech AI still runs and has great SEO leading to roughly 100 new signups daily. Even better, some pay a monthly subscription to get better models and unlimited generations! Overall, this has been an incredible project to learn deeply about the UX patterns around building AI products, building growth via SEO, and supporting a passively revenue generating product. Try out LowTech AI today.

Try LowTech AI

One liner: The best model for you is you.
Inspiration: After LowTech AI, I partnered with Tim Valecenti at MIT to build something. We weren't sure exactly, but we had become friends and white boarded a few ideas to try. We decided we wanted to work on something that was newly possible with AI, leading AI companies like OpenAI wouldn't build themselves, was B2B in nature, we could prototype quickly, and would only improve as AI got better. We were industry agnostic and talked about ideas from construction, to children's toys, to healthcare.

We eventually landed on building Modelle, a virtual try on experience powered by stable diffusion. I imagined a world where you load up a clothing website and all of the models are you! That's the vision we were going after, and we set off to build our prototype and find our market. We talked with retailers from Lululemon to Congo Clothing Company, and hacked together the first version of our prototype on a Friday in December 2023.

The launch: We shared among our friends, potential customers, and even walked around the Prudential Center demoing with shoppers. Thanks to everyone who tried us out!

Status: After about 2 months, we decided not to pursue Modelle for a few reasons, most of all for me being that I didn't want to dive deep into the fashion e-commerce space. We were also fighting an uphill battle: days after we demoed our prototype with shoppers at the Prue, Alibaba and Amazon independently open sourced their own AI models for virtual try on. Competitors were popping up, including at least two companies in the YC winter 2024 batch. And in the short term, it was far too expensive to generate images at scale for cheap enough. On top of our motivation, our competition, and our poor unit economics, retailers didn't want to pay us for the service unless we proved it increased their sales, and shoppers cared much more about the "fit" of the clothing than the style of the clothing with their skin tone and face.

When inference costs really fall and these models improve, I'm sure there will be several companies in this space that transform fashion and e-commerce, but it won't be me!

Check out Modelle (site live, but inactive)

One liner: LLM optimization for enterprises
Inspiration: I made the difficult decision to go to business school for a few reasons. First, I wanted to move away from building fintech webapps and become the CEO of a startup. Second, I saw the efficacy of GitHub copilot and wanted to move away from building software as my core competency. Finally, I wanted to shift my career focus towards the only industries I thought mattered on any meaningful timescale, AI and energy. MIT seemed to be the only university for an engineer at heart with a top business school program, and I'm glad I made the leap.

In January of 2024, MIT's Schwarzman College of Computing released a writing competition on the future of computing, and I took that as an opportunity to dive deep on the intersection of AI and energy. My essay, "Compute Will Flood the World", didn't win a cash prize, but it did spark what would become Max Compute Co. At this point I'd built a few AI projects and had done research on AI's energy consumption so I knew two things to be true: using AI models was extremely expensive, but also extremely resource intensive. But not all models are created equally, and the rise of small models led to my building MCC.

Our product was a middleware proxy for developers to route LLM requests to the smallest, cheapest, and most environmentally friendly but best model to solve a user's query. It was great for companies to save money, reduce latency, and improve reliability, and as a side benefit was great for the planet to save energy.

After planning out heuristics for building this sort of router and starting to talk with new companies, we built the prototype at an AGI House hackathon on MIT campus and came in 2nd. We won 5k in free Lambda credits!

The launch: The biggest moment for Max Compute Co. was the MIT 100k pitch competition. We'd been waitlisted from MIT Delta V, rejected by YC after 2 interviews, and denied from a number of other programs and pitch competitions as we tried to get our company off the ground. But from hundreds of applicants, we were lucky to pitch among 20 semifinalist companies for a chance to win 100k. We were more lucky to win and advance to the finals. Then in front of a huge crowd at MIT, Tim and I pitched MCC together and won the 100k prize. It was a hugely validating win, and an incredible intro to the Boston VC community.

Status: Unfortunately for Max Compute Co., the AI landscape had shifted below us. By the time we won the 100k prize, 5+ competitors had launched. Potential enterprise customers either wanted to partner with a single AI company directly, not route between several. An open source solution looked like the winner. And most importantly, Tim and I had different priorities. Despite some positive, external validation from the MIT 100k, we didn't see a way to make this business work and decided to part ways.

Check out Max Compute Co. (site live, but inactive)

One liner: Grammarly for AI Prompts
Inspiration: In March of 2024, I'd helped run the MIT Sloan AI conference, the second of its kind and the first held at the MIT Media lab. As part of that conference, I helped run a session with Sabrina, an applied AI team lead at OpenAI. A few weeks after the conference, I met up with Sabrina in San Francisco and she shared her idea for a "grammarly for prompting". After building LowTech AI, I loved the idea immediately, and during a cab ride to Palo Alto I coded up a chrome plugin prototype (no code complete, all natural). We passively worked on the idea for several weeks, culminating in a weekend when Sabrina came up to Boston from NYC for a full day hackathon.

The launch: We tested with ourselves and a few friends but never published it more widely. Overall, Sabrina got busy with work at OpenAI, and I moved on to my next project. It's still an idea for the future.

Status: Hibernation, potentially very possible now with WebGPU and tiny models!

Check out Promptly (chrome webstore)

One liner: An inexpensive public GPU cloud for AI training powered by renewables.
Inspiration: Atomic VC is a VC fund and startup accelerator behind breakout companies like hims & hers, Exowatt, andElly. The firm experiments with ideas, tries hard to kill them, and takes on the most promising opportunities to found, fund, and find a team to run. Around the time we were debating next steps for Max Compute Co., Atomic reached out to me about becoming a founder of a new company they were exploring called Build AI.

Taking advantage of low energy prices in remote areas and the massive demand for GPUs for training, Build AI aimed to build modular data centers were energy was cheap but intermittent. Success would mean cheaper training for companies, AI infrastructure powered by renewables, and a perfect merging of my interests and strengths on a world scale problem. In May I moved to NYC to work on Build AI with the intention of dropping out of MIT if the business was taking off. I was ready to be all in.

The launch: At Build AI, I took on the role of a non-technical founder and focused on A) hiring and managing the build of our cloud offering and B) pushing for sales by talking with customers, anyone who would get on the phone with us and talk about AI training. Without actually having built out AI modular data centers, we launched a cloud platform in August and onboarded customers using rented GPUs.

Status: It became clear over the months I was working on Build AI that the core assumptions of this company were wrong. A win for entrepreneurial experimentation, but another failure for me. In part with my recommendation, Atomic decided to shut down the company. A few key reasons we decided to shut down:

For a training cluster with top of the line interconnected hardware, the price of energy didn't meaningfully lower the cost of GPU/hr pricing. Training requires the best hardware (H100s at the time) intfinibanded together, and the hardware alone for building a cluster is the only cost that matters over the lifetime of the asset. Earlier financial models for the company had discounted how expensive the GPU hardware and networking would be. From a user perspective, you get a cloud that is slightly cheaper, but has sizeable tradeoffs (intermittency, higher latency).

We realized that only about 40-50 teams in the world are consistently training on more than 40 GPUs at a time. If you're working close to the AI space, you can probably name everyone of them, and at reasonable scale they all want to build and own their own hardware. Companies like SF compute Co have stepped in to cater to startups for bursty training, and companies of larger scale are building their own clusters. The vast majority of training is smaller finetuning jobs.

Small companies train on free credits. Larger company train on hardware they own. The largest companies build data centers.

Atomic would need to invest 30M+ for a single POC modular data center to prove the model before we scaled. They were doubtful it was worth the shot.

We were seeing enormous outlays of capital from existing hyperscalers and GPU clouds like CoreWeave, Lambda, RunPod, and Crusoe.

After all my customer calls, building the platform, researching opportunities, and evaluating what we'd learned, I recommended Build AI shut down. It crushed me, and I dealt with that in the weeks that followed, but it was the right decision.

Check out Build AI

At this point, I returned to Cambridge to finish my MBA at MIT. I learned a ton working on Build AI, but I didn't know exactly what to do or what to build next. Through the chaos of wedding weekends and school, I was lucky to come across the Sundai Club. Sundai is a group of hackers from MIT, Harvard, Northeastern, and the Boston area that get together on Sundays to launch AI powered apps in a single day. Get there at 10am, demo at 9pm.

It's important to note that at this point I, a business school student, was still writing code daily, and around this time switched from working in VS code with Copilot to using Cursor. I didn't expect the change to be so drastic, but wow was I wrong. Cursor is an incredible tool that's made getting into a flow state building software that much easier and more fun.

Through my experience with Github Copilot, switching to Cursor, and these hacks with the Sundai club, I wrote an article about the future of software engineering and the only two skills that matter: taste and tradeoffs.

During Sundai hacks, I've built seven or eight products, but I want to focus on two that I worked on beyond the single day hackathons: Mail Kiwi & The Turing Fest

One liner: Submit a prompt, get a custom postcard in the mail.
Inspiration: This tweet that Kevin tagged me in! I loved it, thought about it for months, and eventually built it. A really fun project.
The launch: I launched on Twitter to lots of positive reactions. Still need to build out what one friend called "Mom mode", send a postcard with a custom message. AKA, the normal way to send a postcard.
Status: Still alive and well. Send yourself a postcard today! My grandma does it every few months. Lob is the best postcard API I could find.

Send yourself a postcard with Mail Kiwi

One liner: Quiplash for LLMs
Inspiration: Honestly, I'm just a big fan of quiplash, and after using convex.dev I figured I had the ability to make this in a single weekend.
The launch: I've played 30+ games with friends all over MIT, starting at a night hosted by Blake Blaze.
Status: Still alive and well. Someday, I hope to host an event at The Turing Tavern in Cambridge MA for groups to play.

Play The Turing Fest

One liner: Fantasy football for music
Inspiration: Working with Blake on a fun new concept.
The launch: Shared with a group of friends at MIT, great feeback so far!
Status: Blake is continuing the work, I think it's a killer project!

Join a league on Front Row Fantasy

One liner: An API for cheap asynchronous inference
Inspiration: My experience at Build AI led to building a plaform that focused on inference not training. When you stop worrying about hyperfast inference, there's a ton of room to optimize. We wanted to enable more economically viable AI use cases with higher latency AI, and make better usage of the GPUs already plugged into our energy infrastructure.
The launch: Built a prototype and launched to a small subset of companies.
Status: It didn't work for a number of reasons, but led to me meeting Hongyin.

One liner: A single editor to manage your atoms (materials science processes) and your bytes (data to inform those processes)
Inspiration: I met, the founder a brilliant engineer, who came to this concept after exploring the materials space to fix what he called a "top three problem in the world".
The launch: This work is in progress, but hasn't publicly launched just yet!

It's been hard at times, especially letting go or parting ways with a collaborator, but it's been extremely rewarding and fun. Business school can be a lot of talk, and while I learned an unbelievable amount and met incredible life long friends and mentors, I couldn't imagine not participating in the AI wave that's happening all around us.

My advice if you're starting out working on something now: you have to go for it. Anytime spent wavering is wasted time. Once you're in the middle of it (can take 2 weeks or 6 months), it becomes very obvious very quickly wether that project is worth your time. You'll see the technical shortcomings, the ways OpenAI could crush you, the impossible sales process in front of you, or the feel the lack of traction. It only works if you do it, and it's usually pretty fun too. If you're lucky, you might just find something that works.

Thanks to everyone who's supported me on this journey. Excited to share more on what's next for me with Subconscious v2 in the near future!