CharlieKey takeaways from your favorite podcasts
← 14 October 2026AI & I

How Hugging Face’s Investor Spots Winners

Matt Hartman argues that the best early-stage bets come from technical founders at the edge of what is newly possible, and that in an AI era the scarce skill is learning from users and iterating fast.

Watch on YouTube ↗

Charlie's summary

  1. Muse is benchmaxed, while Gemini is a standout writing model but weak at coding

    The intro testing found that Meta's Muse scores well on public benchmarks but lands mid-pack on real work. It got about 50% on the senior engineer task, buries the lead in its writing, and takes the bait in the McDonald's test. Gemini 4 also looked benchmark-maxed, but it became the host's favorite writing model. He thinks it made him about two times more productive on an 8,000-word piece. It scored only about 20% on the coding task, so use it for writing.

    ▶ 0:30
  2. Hugging Face won by looking honestly at retention and pivoting from a chatbot to a hub for open source models

    Matt says the team had a consumer chatbot app with weak retention, and they were “very honest with themselves about those metrics.” When transformers came out, they converted the paper into PyTorch, it went viral on GitHub, and other researchers asked them to convert theirs. They felt the demand and became the hub for open source models. Matt's lesson: the best founders ask what is genuinely working over the long term, not what metric they can raise on.

    ▶ 8:38
  3. Hugging Face's key decisions: hold open source constant, move off GitHub, and expand beyond language models

    Matt names the decisions that mattered. They committed to open source and transformers. They moved off GitHub to a richer, more structured home for model weights. They also stayed flexible about which kinds of models to host, as video, robotics, bio and weather models arrived. In his view, open source was the one thing they held constant, and that let everything else flex.

    ▶ 10:11
  4. Look for what is newly programmable and pull on theses that are 20 degrees off

    Matt's method is to find the newly technically possible by tinkering, reading papers, and asking smart people questions. He says the resulting thesis is often “20 degrees off” from what he predicted, as with Hugging Face and Anchor, which began as an audio social network. Money, websites, and now robots, which he calls a software problem like “an iPhone on wheels,” show how fields become programmable. Look for teams treating them as software problems.

    ▶ 13:16

Give your email to unlock the rest of the key points — for every episode.

Do more with Charlie

Free — sign in with your email, no password.