Gemini 4 Argon Can Finally Compete With Claude and ChatGPT

Summary · 5 min read instead of 51 min

Matt Hartman is the founder and managing partner of Factorial Capital, a pre-seed and seed fund he started in 2023 and runs as a solo GP with founders he has backed as venture partners. Before that he spent nearly a decade as a partner at Betaworks in New York, where he was part of the team that wrote the first check into Hugging Face when it was a chatbot app for teenagers. Hartman calls Hugging Face probably the largest place for open-source AI, and Nvidia bought it a month before this episode for $13 billion.

Matt Hartman argues for backing technical founders: he assumes every early-stage company will pivot, and teams whose pivot foot is the technology rather than a market can follow it into new areas, as Hugging Face moved from a teen chatbot with weak retention to the largest home for open-source AI. He says AI has cut the loop from user feedback to a new version from two months to perhaps two days, so turning user behavior into product insight is now exponentially more valuable, and he screens for founders with earned insights. Factorial runs as a solo GP fund whose venture partners are founders he has backed, and its portfolio has become its referral network. Building Ghost Pepper, a local voice-to-text app, taught him that vibe-coded tools need upkeep and gave him new respect for SaaS. His current theses include developer infrastructure, frontier non-language models for robotics and science, and AI apps built by teams that can go fully down the stack.

Minute by minute

1. 5:03

   Hugging Face pivoted from a teen chatbot. Hartman calls Hugging Face a marketplace between AI researchers and the developers who use their work. Betaworks backed it while exploring conversational software after the Word2Vec paper showed words could become numbers that carry meaning; the first product was a chat app. It had many users but weak retention, and when the team converted Transformers to PyTorch, it went viral on GitHub, researchers asked them to convert their papers, and the work moved to the Hugging Face hub.

2. 9:50

   Three decisions set Hugging Face's path. Asked by Natalia Quintero for the pivotal decisions, Hartman first credits the team for judging itself on retention rather than raising on whatever top-line metric it could. He then names open-sourcing Transformers, moving model weights off GitHub to a richer, more structured home for researchers, and expanding beyond language to video, robotics, bio and weather models while always holding open source constant.

3. 12:45

   Theses start with the newly possible. Hartman says his goal is to find the "newly technically possible" by tinkering, reading papers and questioning smart people, and his thesis is often "20 degrees off" from what a company becomes. Around 2014 he noticed nothing competed for his ears and spent a year seeking a Twitter for audio. Anchor, whose team knew why people dislike hearing their recorded voice, began as an audio social network and became the most popular tool for creating podcasts by serving creators, not listeners.

4. 18:07

   Technical founders can pivot further. Hartman says he always assumes an early-stage company will pivot, so he asks where it could go. Using basketball's "pivot foot", he says real estate executives building AI for real estate will likely stay in that market, while a technical team may find the same shape of data in the car industry. The most successful companies he has been involved with kept the pivot foot on technology, not a vertical.

5. 21:07

   Portfolio founders became the deal source. Hartman set up Factorial around 2023 as a solo GP fund with venture partners, such as Hugging Face CEO Clem and a Venmo cofounder, to help him make high-quality decisions quickly on fast-moving technology. Over time the portfolio itself became the referral network: Clem led him to a team that introduced Patronus, which then introduced Causal Labs. He describes it as a "flywheel or a family tree" and notes Factorial began in Hugging Face's office.

6. 27:06

   Faster iteration makes product insight key. Hartman wonders whether companies like Factory and Modal grew so fast because AI speeds up iteration: a loop from user feedback to a new version that took two months might now take two days. Hugging Face took 10 years to reach its position; these took only a few. Whatever the job is called, he says the ability to see how people use a product and improve it is "exponentially more valuable", though distribution also matters more as building gets easier and noisier.

7. 31:01

   He screens for earned insights. Asked how he screens for this, Hartman says founders rarely think they are pivoting, so their ability to iterate matters more than the current product. He looks for an "earned insight", something few people were placed to learn, which buys time before competition arrives. He says companies more often find traction by shipping and cutting everything else, and cites Anchor's separate onboarding for users with and without headphones as care that pays off.

8. 34:18

   Ghost Pepper changed his view of SaaS. Hartman built Ghost Pepper, an open-source voice-to-text app that runs local models so it has no inference cost, using Claude Code and Jesse Vincent's Superpowers skill. It reached number one on Hacker News, and users said it was faster than state-of-the-art products. Maintaining it changed his view of SaaS: he says paid apps aren't that expensive, and what users pay for is a product that keeps improving. Quintero says she dropped her own vibe-coded CRM over upkeep.

9. 39:43

   Transformers can predict more than language. Hartman says much of the opportunity is still infrastructure, such as tools for new kinds of developers and products that maintain all the code being written. He continues to think about frontier non-language models: if Transformers excel at next-token prediction, robot movement is time-series data, and Factorial has invested in a couple of companies betting robot training could change. He cites a Hugging Face hackathon on a child's rare disease, with diagnosis and drug-repurposing tracks.

10. 41:59

    AI apps will be built full stack. Hartman expects AI apps, even consumer or vertical ones, to be built by teams technical enough to go fully down the stack. He cites Instagram, whose square photos, he has read, spared the app from asking the server about orientation and cut load time. AI-native teams, he says, will borrow ideas from research papers, build their own models where needed and use big companies' models elsewhere, and striking that balance will be a real edge.

11. 44:10

    The best founders are playing. Hartman says the best builders he sees are playing, as in the early internet; he tried a text-to-CAD tool to see if he could vibe-make a robot and thinks it is not there yet. He thought AI would be at least as big as mobile and now expects it to be much bigger. Likening the moment to a dark World of Warcraft map revealed as you move, he says even model builders are still exploring, and there are more solutions to existing problems than ever.

12. 47:58

    Pair new sources with older books. Hartman reads X, where he says people now post longer pieces that you can feed into Claude to apply to your investing, along with what is trending on Hugging Face and newsletters such as Ben's Bites and Every. He calls his approach a barbell: the newest material at one end and older, hopefully timeless work at the other, such as a book on the fall of the Roman Empire, since certain things seem to be repeating.