Why I Joined AI
A year and a half ago I was homeless.
I was doing all of these things simultaneously:
- Full time student at UC Berkeley
- Training with the Marine Corps Officer Training Program
- Teaching juggling to kids for Prescott Circus Theater
- Selling alpaca wool ponchos from Ecuador
- Working as an executive assistant for a Lawrence Labs researcher and also for an adult film star (I remember setting up a project management board for her and getting a notification when she added “edit the BJ video” as one of the tasks on it, assigned to someone else not to me by the way. I also helped try to find the venue for her wedding.)
- Side projects such as consulting for a startup using prizes to incentivize research.
My Navy friend showing off his poncho in the attic I lived in
My four paid jobs and one side hustle were enough to pay for my rent in Berkeley where I split an attic with a roommate, but as my adopted family in Ecuador struggled to find work during the pandemic, I tried to pick one or the other and ended up failing at both.
It reached a point where they were eating one meal a day and I was sleeping on friends’ couches. I wasn’t able to take care of myself and my loved ones at the same time, and I felt crushed and powerless. I realized I needed to make a change.
I dropped out after finishing my summer military training in San Diego in 2022 where I commanded 30 Marine Corps midshipmen in Camp Pendleton, drove a nuclear submarine in the Pacific Ocean, and watched dolphins jumping around the hull when we surfaced. The next week, I was back on a friend’s couch with nothing but a backpack and an urgent need for entrepreneurship.
I pursued work opportunities and was rejected from the ones I wanted. As a backup plan I considered restarting my performing business or even street performing in Denver again where I made $300 a day juggling, but that would defeat the reason why I came to the Bay Area in the first place to become involved with tech projects so I could find scalable ways to make a change in the world.
Starting The Job
Thankfully, the Lawrence Labs researcher I worked for re-introduced me to a senior researcher at OpenAI. That senior researcher was co-founding a start-up and needed a team of data labelers for reinforcement learning with human feedback to train a large language model to create text summaries and answer search engine questions.
He gave me the task set which basically consisted of receiving a commonly searched question on Reddit such as “how does the immortal jellyfish never die?” Then the data labeler searches for articles relevant to answering the question, highlights the portions of the articles that are relevant, writes an answer in their own words, and then cites the highlighted portions from the articles showing where they derived their answer from.
In other words, before Bing copilot was released, we were also building that.
Example:
The co-founders of the startup went almost radio silent as they became insanely busy with the seed fundraise. So I had a job and a budget and immense autonomy to accomplish it. Other than a co-founder telling me that if I wanted to paint boomboxes bright pink and blast music in Berkeley as a recruiting strategy he would bankroll it, I didn’t have much instruction.
I began where any good employee does, with a spreadsheet and with time tracking. I mapped out places for recruiting, created templates, pay schedules, HR systems, protocols for firing people and for quality control, and by the time the co-founders returned from raising 8 figures I had built a functional team of thirty people from scratch. They hired me full time as a founding member.
Over time, I eventually hired five sub-managers reporting to me who managed a team that at its largest reached 70 people at once.
AI Ideas
Our main product was "MultiFlow" which had the end goal of creating a design studio for AI automation workflows kind of like an AI Zapier (also before Zapier released its own AI automations.)
For example, you might drag and drop a text block with the instructions “Take this question about a privacy policy: [input]”
Then drag and drop a website scraper box that can take a URL as input
Then drag and drop a box that has the instructions, “Based on the privacy policy document shown at [input URL], please provide an answer to this question: [input question]” Then add another box that says, "please translate that answer to be sarcastic and sassy."
I wish we were open source because many people have asked me to use this platform and it was incredibly powerful for creating backends for AI apps.
At an SF AI hackathon where we won a pogo stick for "most useless invention" with the creation of a chatbot that sarcastically explains privacy policy documents to you
In this hackathon app we used the MultiFlow backend and someone took this blurry image during our demo. We asked whether Tesla stores data and turned on the NSFW option
The real potential I saw was in being able to save a whole workflow as a “custom block” which essentially means modular programming with chunks inside of chunks that could be shown in a much more visual way for no-code app development.
We had workflows that could write a blog post off of a one word input and generate images for it, too. You could imagine that you could save that whole workflow as a customer block titled “Create blog,” then add that custom block within a workflow that creates a blog post, generates a podcast episode, writes some tweets, integrates with those various platforms and automatically posts it on a certain schedule, maybe even generates a progression of topics to blog about and thus create a fully automated media production. This took minutes to set up within our workflow designer.
I liked the idea of making this a marketplace where workflow designers list their automations on our platform which may be used as a backend for multiple apps… or the workflow can be run within our platform. Then people pay per use of the workflow and the revenue is split between the workflow designers and the platform, and potentially there is also a revenue share with the data labelers who helped contribute to the function of the AI.
This made me think about ways to “mine AI” in which we expanded our data labeling service.
In that structure, data labelers are “investing” their data passively or actively using their time doing tasks training an AI model. The more that AI is used the more revenue they make long-term instead of simply being fired once they’ve contributed enough data points to be automated away.
I was exploring ideas in federated learning, homomorphic encryption, or ZK data contributions as well (I write about this here: https://www.noahchonlee.com/post/how-to-decentralize-ai)
On the legal side, there were also ideas to make every LLM into its own LLC with revenue sharing agreements or even equity splits shared with data labelers.
It was also looking likely that Europe would enact data privacy laws preventing AI companies from scraping data from the Internet for training which implied that whoever had a privacy-preserving way for people to consent to contribute data to these AIs would be well poised to take the AI market.
At the core there was an idea of automating away online work and becoming an AI mechanical turk. If someone dragged a block into the workflow that we could not automatically complete yet, then we would have our human data labelers do the task manually enough times until eventually the AI could imitate them and then forever afterwards the action within the block could be automated and not need a human doing it. We also looked into ways to integrate machine learning over mouse clicks and keyboard strokes so the AI can truly do anything on a computer a human can do. Even back then there was a startup that had used trial and error to teach an AI to order a burger online by simply telling it to do so and it could move the mouse around and make the clicks to make the order.
Automating Away Jobs
I didn’t figure it out fast enough for my liking.
I had hired over 100 people during my year there and I would call with them and hear them thank me for being able to spend more time with their kids because they can work flexible hours at home, live in a better home environment in their own apartment because now they have a good wage, attend university thanks to this, etc. I felt amazing hearing these things.
But our business model was literally to employ people to teach an AI to do a task until the AI could do it well enough that we no longer needed humans to do it, then recombine these automated tasks into even more complex configurations that anyone could easily create.
One co-founder preferred a B2B model using this as a platform for businesses to use internally whereas I preferred the idea of empowering anyone to build with AI by doing a B2C model.
When I found out one potential client was looking at how they could use our automations to fire 300 employees in the Philippines, I realized I would personally be automating away jobs faster than I was creating new ones, that I did not have a clear enough path forward to quickly distribute revenue or equity to data labelers long-term, and didn’t not have enough control of the company to be able to influence it to direct revenues towards charitable or UBI programs.
Despite the 6 figure income (which was only 60k after taxes in California and is mostly eaten up by SF cost of living) and the respect I received and the powerful network and even feeling like I was in the midst of the most exciting possible industry in the best possible place to be involved at the exact time when the industry was still small enough I felt I could actually know a significant chunk of the network… I didn’t feel right staying there.
So I left.
AI → Web3
As you can imagine, ideas of tokenizing revenue shares of an AI or showing onchain when data is being accessed came up during my time searching for ways to decentralize AI. Increasingly, web3 seemed to be lagging behind AI while also containing within it the seeds of the exact solutions which were increasingly being needed.I also felt the urgency increasing for ways to verify authorship of a piece of online content in the face of deepfakes as multiple founder friends of mine were raising millions creating deepfake apps.
They were fun ways to add a filter of a celebrity’s face and voice and get bought out by Snapchat for example, but it also seemed clear to me that this would lead to a world where an unethical political party could get themselves elected through using deepfake videos in Facebook in Brazil, for example. If I unintentionally support a technology that leads to someone like Bolsonaro being re-elected and the burning of the Amazon and the deaths of more of my friends there, I would never be able to forgive myself.
Some ideas for addressing deepfakes are written here: https://www.noahchonlee.com/post/how-blockchain-can-save-the-internet-from-deepfakes
I am concerned about the dangers of AI based on personal experiences. And considering these concerns, it seemed reasonable to work in blockchain technologies that could address those concerns rather than continuing to work in AI. After all, the biggest AI safety orgs OpenAI and Anthropic are now exactly the orgs that scare the most people as they advance AI the fastest. That sure backfired. Plus, I couldn’t stand the insufferable ego, elitism, and the self-flagellating guilt that spills over onto bystanders of various people in the AI safety movement who believed they needed to be saviors of the world.
There was also the consideration that I was leaving the place where all the money was flowing into. I left AI in the bull market to join web3 in the bear market while most everyone was heading in the opposite direction.
I turned down AI jobs and felt it difficult to extricate myself at times and take the risk that I might once again run out of funds and not be able to take care of my loved ones. So my choice seemed highly uncertain and potentially quite dumb at the time.
I also see the potential for AI and still keep my toes dipped in AI a tad:
- I did a little recruiting work on the side for a couple AI companies
- Helped build this AI voice for a river presented to the Prime Minister of Montenegro (airiverchat.com)
- Built 7 AI apps in 7 days with Dylan Wu
- Taught some AI classes in Syria
- Spoke during the AI week in Zuzalu
- We will also likely incorporate AI translation into a couple platforms I am involved with building
But other than that I have mostly been involved in web3 nowadays. There are certain questions web3 helps with that drew it to me such as: In the face of centralized control and automation brought by AI, how do we easily send funds across the world to decentralize wealth potentially through UBI?
My first time ever using crypto was to send funds to a friend in Turkish occupied Syria because it was the only way to do so, and I found it was so much simpler than trying to do international bank transfers.
In the face of AI deepfakes, how can we make the Internet trustworthy again? Cryptographic history seems like a solution. Ideas such as government IDs linked to someone's web3 wallet through privacy preserving zero knowledge cryptography onchain are ways to associate a person's online profile with a real identity in order to have trust in the sources of information. These "proof of personhood" IDs are also important for wealth allocation and voting systems to prevent "sybil attacks" where one person has multiple online accounts. As well as this, I found that the public goods ecosystem within web3 was in favor of open source software and experimenting with new economic mechanisms to grow and spread wealth.
The Result
The way I saw it, I wanted to create crowdfunded prizes as a tool that I might use to support all the other projects that need to exist in the world. By using this mechanism, we are now working on coordinating incentives to focus builder energy into creating solutions such as with this prize for a ZK DID: https://www.viaprize.org/prize/1ccb751b-ccbf-4632-b558-e3fac506d773)
And I have a series of prizes in mind for projects used for the network society movement such as an intentional community/grouphouse matchmaker app: https://docs.google.com/document/d/1843dXNPbnXJTfaOHSZf-sMrVIvLngEmQ-5P0m0Wo_jA/edit?usp=sharing
In the midst of so much uncertainty about how things like AI will radically transform the world… it seems best to me to return to the basics of being human. I think we need dunbar sized communities where we feel we belong, work that is meaningful, and we ought to have the right tools to come together and take care of one another such as community mutual funds so any group of friends can start their own Norway style sovereign wealth fund (see on my old blog here: https://themasterfool.substack.com/p/proposal-for-community-mutual-funds)
This is why I have been involved with Zuzalu style pop up villages where a Dunbar sized group gathers to colive and cowork in community. In this way, we can iterate on ways to form communities as we identify real needs and build meaningful projects (that are often web3.)
A year later though, I feel great about this decision. It was a bad choice for ease of access to funds, but a great choice for my long-term relationships. By moving from a bull market industry into a bear market industry, I found the people who truly cared about the technology and the projects they were building and weren’t just chasing the hype or the money. The hype and the cycles shift and the markets move up and down, but I think the people who I have connected with and the relationships made will outlast those cycles.
Now I believe that you should start by finding the right people who you are inspired by and want to be like. I could chase the bulls and flee the bears as they fight their way back and forth, but it's humans and not bears or bulls who I want to spend my life with.
With the viaPrize team at EthIndia in Bangalore
Appreciations
A special thank you to…
Steven Elleman for believing in me even as my life fell apart and I was living on his couch
David Hunt for my favorite job working with Prescott Circus
Kevin O’Brien for buying me a ticket to Dweb Camp and helping me find a tech community I felt aligned with
Alok Singh for the work opportunity and for the introduction to the OpenAI senior researcher
Emma Qian my boss at Multi Technologies for all her patience and guidance as I made many mistakes during my time as Chief Human Resources Officer managing the data labeling team. I am so impressed with her work ethic and was honored to be reporting to her.
Jehan Azad who entrusted me with the viaPrize project he had started and whose initial 3k in funding allowed me to create meta-prizes inviting people to contribute to our code which led to me meeting the team in India who became full time team members.
Kristof and Threefold for funding my first trip to Zanzibar and to Giorgia for offering me a free place to stay in return for fire spinning when I ran out of funds until the Gitcoin grant arrived allowing me to travel to India.
Currently all of my team’s funding for our open source web3 company viaPrize is supported through the public goods funding platform Gitcoin. I am very grateful to the over 1000 individuals who have contributed to us there and to Vitalik for funding the Zuzalu round!
To Nicole Sun for welcoming me to Zuzalu
To Janine Leger for how appreciated I felt while I was there and for hiring me to run the first Zuzalu Gitcoin round at a time when I really needed the funds as I was in Zanzibar
To Gary Sheng for his contributions to the team and inspiring and encouraging conversations
To Timour Kosters for his assistance as I was writing up applications for viaPrize
To Umar Khan for his help and amazing work in all things Gitcoin
To all the donors to the viaPrize campaign for delivering medical supplies in Ukraine and to Rev Miller for helping me source the supplies
To Kara and Niklas for welcoming me to a conference in Prospera where I learned so much about special economic zones and charter cities
To Rachel Shu for gifting me the backpack which is my home, and for being one of the individuals who has shared in so many adventures with me
To Nourshan Husein for being an absolute badass and kindhearted leader while facing the threat of assassination every day, and for offering a place to stay to me and Rachel during our time in Syria
To 4seas and Nicholas Hu for the grant to be at Vitalia and for providing the venue for the first viaPrize hackathon and for the brilliant ideas shared by Link Guan
To Shina Foo for the tea and care all the times I got sick from overworking
To my sisters for being there for me when I need to call and rant
To Dipanshu Singh, Swaraj Bachu, Nithin Varma, and Aryan Tiwari for your brilliant ideas and work building viaPrize and for being people I am inspired by
To my adopted family in Ecuador for showing me that giving your all to take care of others can be a normal and fulfilling way of life
Comments