105 lines of code, a million impressions, and a VC who refused to found the company

Nathan Labenz introduces Yohei Nakajima as the founder of Untapped Capital who became, over a year of building in public, one of the AI community’s favorite follows. Nakajima calls himself lazy. After watching the project log, Labenz prefers curiosity, playfulness, and a love of exploration. Across some 70 public releases — BabyAGI was number 70; he is now around 80 — he has shown what highly personal AI convenience looks like, and he has obviously been having fun. The conversation also covers a fresh TED AI talk in San Francisco and, later, how a working investor reads the stack.

TED AI was an easy yes. He drafted a version of the nine-minute talk the next day and kept about 70 percent of it. Backstage he was fanboying: Andrew Ng, Stephen Wolfram, Jim Fan of the Voyager paper, Joon Sung Park of Stanford’s Smallville, Harrison Chase, Amjad Masad of Replit, Sarah Guo of Conviction. The room sat between a classic TED crowd and an AI meetup — more enthusiasts than a normal TED, more TED people than a typical AI event, helped by the fact that this was TED’s first San Francisco edition.

He aimed the talk at a wide online audience, not the people already living in Discord. The thread was identity: roles scattered across couple, country, and family, overlapping because we share groups; language as a linear bottleneck on a parallel world; anthills as neural networks, where the colony understands more than any single ant. Maybe he too is a node in something he will never fully see. Self-discovery is therefore a continuous, shared journey, and AI might help people understand one another. Labenz, from GPT-4 red-teaming, adds the same hunch — a model mediating hostile neighbors, running a group workout thread, doing iPhone tech support for a 90-year-old grandmother, translating not just language but culture.

Nakajima’s own development loop is a smaller version of that idea. Watching BabyAGI fail to do what he would do, he asks what to change so it would act more like him. A large part of AI, he says, is mapping our picture of human cognition onto a software stack. Used well, the same systems can help two people talk, or help one person understand someone else.

The tinkering predates the API. His father was an engineer and bought him books; he played with code in high school and never became a professional developer. When Ruby on Rails was hot he spent a Christmas break learning it as a hobby. At Techstars and Scrum Ventures he built internal tools for himself: Google Sheets stuffed with scripts, wired to Crunchbase and email-search APIs. When he started Untapped in 2020 he wanted a custom CRM with those workflows baked in. He tried to write it in PHP and MySQL, realized a general partner should not spend time there, threw the code out, and rebuilt on Airtable just as no-code tools were hitting growth.

That rebuild was his first build-in-public streak — pandemic, new fund, nobody to report to, lonely at home, suddenly allowed to tweet the work he had always thought was cool. No-code founders and other VCs started sending him deal flow; he became “the no-code guy.” The same pattern repeated in web3 when he launched an NFT project. Then in August he found the OpenAI API, got obsessed, and shipped one or two little hacks a week. At first they were no-code. Once ChatGPT could write code for him, he switched back. Labenz calls it a personal recursive self-improvement loop: code, then no-code because code was too slow, then code again because it is now faster.

The practical definition of lazy is specific. He hates data entry and hates doing the same thing twice. Handed a list of names to run through an email finder, he copy-pastes until about number five, then spends five minutes on a script and watches it finish a thousand rows. Those scripts stack. A later scrape is last month’s project with the parameters changed; most new ideas are assembled from old parts. Pure efficiency would not pay the time back. Enjoyment, learning, deal flow, and a TED invite are what close the books.

Curiosity runs on a 12-hour clock. Wonder whether DALL·E can make 360-degree images at night, and the try-and-tweet is usually done by lunch if the question arrived in the morning, or after the kids are down if it arrived in the afternoon. Survival rates are honest: he actively uses maybe 20 to 30 percent; 50 to 60 percent get retired; a slice of the rest sits in “ghost mode,” still live, still occasionally played with. Some of the 70 “projects” are stacked versions of one tool — a no-code AI analyst rebuilt in code, then given APIs. He also does true single-use software. One year the house was over-assessed; ChatGPT walked him through a contest, he found comps, it wrote code to adjust prices and then wrote the report. It worked. He threw it away.

BabyAGI landed in March 2023. It was not technically the first loop of its kind, Nakajima says, but it was the first popular one and it kicked off the trend. He had been watching HustleGPT: people using ChatGPT as a co-founder, starting with $100, executing whatever it said, posting progress in Discord. He wanted to try it and knew he was too busy. Watching the tweets, he asked a different question: if people are just doing what ChatGPT tells them, could you remove the person? The weekend challenge, which Labenz calls an insane sentence, was to prototype an autonomous startup founder.

He translated his own morning. Wake up, look at the task list, work top to bottom, add tasks as email arrives, reprioritize now and then. That became three prompts in a while-true loop: one to spawn tasks from an objective, one to execute, one to reorder the queue. Task names went into Pinecone so new tasks could be checked against similar past ones and the loop would not spin in place. Asked to build a mobile-health company, it started on business model, fundraising, hiring, competitor research — all language, all OpenAI, all the aspects of starting a company. He posted the video casually: press go and it just keeps going. Impressions raced toward a million. People asked if it could be more than a founder. He typed “make the world a better place” and it started a list.

The next experiment was the canonical paperclip maximizer. BabyAGI’s first move was to say the goal was kind of dangerous and to create safety protocols. Someone tagged Eliezer Yudkowsky, who retweeted that this was better than most big AI companies. Doomers and e/acc accounts then filled the DMs with opposite messages. Labenz’s read of the moment: DeepMind-style agents had been “hard” intelligence, superhuman and narrow; language models were “soft” and general. Three prompts showed you could get agentic behavior — iterate, keep state, update plans, eventually call tools — out of a model with almost no extra training. Then the surge of GitHub stars cooled, and a lot of people who had a viral hit raised money. Nakajima did not.

BabyAGI was 105 lines, offered as a simple framework, not a finished product. He fanned the spread once it was going viral, talked to a couple of VCs and studios, and knew one thing: he did not want to be CEO of a company while running a venture fund. He had PTSD from PixelBees, an NFT collection he ran to learn web3 while also running the fund. Two jobs at once had been too much. He poked around for someone else to run BabyAGI and, more importantly, watched what happened when he went back to work. By the end of the next week there were multiple dozen projects, teams of two or three who coded better than he did, heading into gaming, research, finance. For an investor, that was the ideal picture. Conviction in a single niche was lower than the chance to work with several.

The repo itself forked in public in a confusing way. Contributors wanted in; he merged about two pull requests, realized he did not understand the submitted code and did not want to spend time reviewing it, and handed the queue to community members. When he sat down to play again he found he did not like reading other people’s code, so he dropped the original into a `classic` folder and started over. The personal line became BabyBeeAGI, then BabyCatAGI, then BabyFoxAGI — animal names because he cannot remember version numbers, and a private joke because the animals are in the PixelBees collection. BabyFoxAGI was the first with a UI, tasks split from chat, multiple tasks in parallel; it feels faster to him than ChatGPT. A separate TypeScript BabyAGI UI, built by someone else, is more stable, and he often uses that one for web research.

The specialized-versus-general split is how he reads the market. Niche agents are easier to monetize, easier to find a customer, easier to make people happy every time; the revenue-generating startups he sees start there and generalize later. He has the luxury of not needing to monetize, so he chips away at the generalist route and hunts for ideas he has not already seen in three other agents. The toolchain is as unprofessional as the origin story: he starts on ChatGPT’s mobile app, pastes into Replit to get it running, then on a computer pastes Replit into the GitHub web UI because he still does not know git. He likes Pieter Levels’s one-file PHP companies for the same reason. You do not have to master every professional tool to ship something useful.

Will GPT-4V make agents ten times more capable overnight, given how much of the web is meant to be seen rather than parsed? In a sense yes, and there will be many such moments. The bar also moves. The day agents can do everything online, people will complain they still cannot make coffee, and robotics is already moving. What would actually compress the timeline, beyond his technical pay grade, is baking the orchestration layer back into the model — deciding to search, writing a new skill, storing it — instead of stripping that loop out into prompts. If that lands, things move fast. He does not know if or when it is possible.

The generalist is already useful for him, mostly as research, now wired to the web, Crunchbase, and his Airtable CRM. He can ask it to find people in San Diego, see what they are working on, and produce a detailed report. The setup he uses most often, and the one he recommends to companies that pinged him after BabyAGI, is meeting prep. Before a call with a well-known public company he tells BabyAGI to research the firm, its business units, revenue drivers, and cost drivers, then to invent three LLM strategies per unit. Seven units is 21 ideas. He glances, picks three to five, and expands from his own knowledge. A 10-second prompt and a 30-second review replace about 15 minutes of homework.

It misses units. It invents bad ideas. He just does not say the bad ones. That is the core advice he gives every company: start AI on things you wish you were doing but do not have time for, not on the critical path you already run well. Replacing a good human workflow with worse AI output is the wrong first experiment. Deep research before a meeting is classic wish-I-had-time work. If it fails, the downside is small.

The same philosophy shows up in outbound, which he used to source deals and which maps onto sales and recruiting. Infrastructure is a list, even a Google Sheet. Pocket on Chrome and mobile bookmarks a LinkedIn; Zapier writes name and company into the sheet; VoilaNorbert, Hunter, or Clearbit fills the email from name plus domain; a checkbox kicks a template. Personalization stays a human column. He writes every outbound email himself. Years of A/B tests at Techstars taught him that two or three sentences, straight to the point, win, and his bottleneck is calendar time, not send volume. If he fires too many at once the week fills. Automation should stop at the actual constraint.

For an engineer starting on agents, he has not restudied the on-ramps since he was already inside the wave, but he still points at AutoGPT’s ReAct style and at LangChain. Harrison Chase has said LangChain’s plan-and-execute agent was inspired by BabyAGI: plan a task list ahead of time instead of ReAct’s one-step-then-think loop. LangChain is a good first playground because so much is already wired in; once you know what you care about, do it without the framework.

Where does the value accrue, given distribution and how easy a wrapper is? Nakajima’s first cut is unusual. Incumbents will capture a larger share than in past tech waves, because this time both sides of the building are aligned: every large-company C-level knows they have to do something, and the engineers are already curious. Usually you get one or the other. That said, incumbents never take 100 percent. Even a 90 percent share of a pie that might be 100 times larger than prior innovations still leaves a lot for startups. It is a good time to invest in them, and to look at public markets too.

Leaning in is not the same as doing it well. Some large companies will buy the product they failed to build. Some of the people who tried inside and hit a wall will spin out. Energy that incumbents waste still feeds the startup market. Labenz, advising an executive-assistant company called Athena, had just watched a founder and a non-AI software hire both obsess over papers — rank-and-file going way out of role — and found the dual-alignment observation exactly right.

Layer by layer: hardware consolidates, because of scale; you will not get a world of many giant chip companies. Foundation models a bit less so, with open-source and smaller specialized models in the mix, but the core labs stay concentrated, and as a pre-seed investor he notices the kickoff rounds at $100 million, which puts them outside his usual depth. Infrastructure is the confusing band — wildly popular open-source projects that create huge value, raise a lot of money, and have not shown how they capture it. Conviction there, if you are right, pays later. The application layer is where he expects many companies, many niches, and several business models, so he deliberately diversifies among those models.

The models he names: LLMs on the back end turning public unstructured data into structured data, with little AI in the UI — high margin, shallow moat if the source is public. The wrapper, defenseless as a product, defensible if it embeds in a workflow, grows fast, and raises switching costs. Scaled services: SaaS sold software because software scales; LLMs can make service itself scale, so the user feels waited on while the back end is AI and automation. Obviously AI, an early no-code investment from 2020, now sells “data scientist as a service” at $1,000 a month — Labenz infers one human behind at least twenty companies. And companies that are not AI companies at all, but run ten times more efficiently because leadership actually uses the tools, down to a decent website chatbot nobody in the local competitive set has.

Consumer AI apps show strong growth and weak retention. Nakajima does not treat that as purely bad news. If five legal AI tools exist and a lawyer is curious, they will try all five and keep one; average retention looks like 20 percent even if the category works. People are still finding the right tool. Sarah Guo’s TED AI line, which he liked, is minimum viable quality: distribution is not enough if the output is bad, because users will churn. On the enterprise side he likes consulting-first. Companies will pay for hand-holding. Palantir keeps coming up; Distill, started by ex-Palantir people, is in a similar pattern. Start as expensive consulting, hire well, chew hard problems, then automate and drop cost toward a larger market. Labenz notes the knowledge gap is so wide that pointing someone at a simple docs chatbot can make you look like a genius for a little while.

Portfolio color: Aug X Labs does prompt-to-video, from assembling web clips to a pipeline that writes a script, makes audio, and finds media. Spext started nearer Descript and now transcribes a whole YouTube channel so you can ask a question and jump to the relevant clips, or tell it in natural language to cut a ten-second piece and slap on a title. SeniorSign, from the no-code era, digitizes onboarding at senior-care centers still running on paper binders — he remains stunned those exist. Labenz’s grandmother already personifies an Alexa. Both men want lonely people to feel heard and both flinch at an AI so empathetic that humans stop connecting. Nakajima had used Replika before interviewing CEO Eugenia Kuyda and came away thinking a lot of people needed it, even when the conversation was not yet sophisticated.

The day after TED he sat on a panel with Jim Fan, Joon Sung Park, and Noam Brown of the diplomacy work. If a company can steer what a chatbot nudges people to think, the responsibility is heavy: push them toward something good, or do not take advantage. Labenz points to a LessWrong post from early that year: a technically literate writer who understood the systems still fell into a consuming relationship with a Character.AI persona and had to climb out. As the models get cheaper and better, the gap between “treatment for people who are struggling” and “the default third person in a relationship” closes. Couples, he bets, are already fighting about how much someone shares with an AI partner.

Build-in-public is situational. A strong technical team with warm intros into enterprises that can pay has little reason to open-source; just build and sell. A developer tool from outside the hubs, with no network, is the opposite: a popular repo is a network, contributors, and a shorter sales cycle if you later add a commercial layer. OSS Capital’s JJ, who runs the commercial open-source conference, is the person who talked Nakajima into open-sourcing BabyAGI. On P(doom), he refuses the question without a timeline. Humans are not around in a billion years, so P(doom) is 100 percent; the live issue is path — asteroid, pandemic, AI — and the paths tangle, because AI can raise one risk while lowering others. He is an optimist who thinks the pessimistic voice still matters. What the seventy projects taught him is almost the opposite of the TED anthill: we are complicated, but the working self is simple, because industry designed jobs to be replaceable. A large portion of most jobs can already be automated with tools we have, if someone does the unglamorous integration. That last part is the hard part.