The Community Is the Dataset: Why We're Backing a Small Robot Model
Every founder I know is having the same argument about AI right now: do we need the biggest model, or do we need our own data? In robotics, that argument just got settled by the published evidence — and the answer is why Pirin is backing zk0, the open federated-learning project I've been building for community-trained robot models.
The short version: for a robot arm, more data beats more parameters — it isn't close.
What the evidence actually says
SmolVLA, the 450M-parameter model zk0 trains, was scaled up 5× by its own authors. On the standard LIBERO benchmark that bought +1.45 points. Pretraining the same model on a community corpus of 481 small datasets — ordinary LeRobot users' demonstrations — bought +26.6 points on real hardware. Same weights, better data.
It isn't one paper. A 7B model (OpenVLA) gained +18.8 points from a better training recipe alone, with zero new parameters. A 0.9B model (X-VLA) outscored every 7B entry on the same benchmark. And the ICLR data-scaling study found generalization follows a power law in the number of different environments and objects — variety, not volume per task and not model size.
Even Tesla's decade is the same story told at industrial scale: years of FSD progress on a modest 2019 inference chip, powered by billions of fleet miles and training examples scaled from 1 million to 10 million — with the bigger model generations arriving only afterwards, to polish the long tail.
We published the full, referenced evidence review — including the claims we checked and corrected, like the "10 billion km of labeled footage" line Tesla never actually made — on zk0's site: Data Beats Parameters.
Why this is a founder story, not just a robotics story
If data is the scarce asset, then whoever organizes the data wins. zk0's bet is that the data already exists — in the homes and workshops of the thousands of people who bought a $300 open-source robot arm and recorded demonstrations they can't share, because the footage is their kitchen, their kids, their business. Those datasets are stranded. Uploading them to a public hub is a non-starter; that's exactly why they're valuable and exactly why they stay unused.
zk0's answer is federation: the data never leaves the owner's machine. Only model updates travel. The community is the dataset — and unlike a model checkpoint, a contributing community compounds.
That's also why zk0 is now actively recruiting, on the strength of that evidence:
- Data contributors — SO-100/SO-101 owners whose private episodes can train the shared model without being uploaded anywhere.
- Compute contributors — anyone with a consumer GPU in the 10 GB+ VRAM class (a 3060 12 GB, a 3090, a 4070) willing to run a node. The model is small on purpose: the LeRobot hardware guide puts fine-tuning at roughly 10–16 GB of VRAM.
A hands-on workshop series for new contributors is forming around the same idea — bring an arm, leave with a live node.
The Pirin takeaway
Founders keep asking how to compete with labs that have 100× their compute. The robotics evidence points at the honest answer: you probably don't need their model. You need data that is yours, diverse, and organized — and a community with a reason to contribute it. That's a founder-shaped problem, not a GPU-shaped one.
Read the evidence, then decide if you want in: Data Beats Parameters — zk0. Background on the project itself: Decentralized Robotics AI: Building zk0.bot.
