Just Grass
Since the beginning of this year, Grass has quietly shipped some of its most important infrastructure updates since launch. Most of this has flown under the radar — intentionally. We’ve always trusted that the right people would pay attention when it mattered.
Beyond staying focused, we’re very tuned in to the fact that in our industry, escape velocity is the best moat you can ask for. We’ve had a great track record of spotting obvious opportunities before they’re obvious, sometimes before incumbents even realize they exist. Constantly telegraphing our direction would have killed this advantage.
This feels like the right time to share more — and to kick off a new blog.
Grass Mobile
Grass Mobile is live, and it’s working better than we expected.
It’s extremely difficult for websites to block mobile IPs. Most mobile traffic passes through large carrier-grade NATs, meaning thousands of users share a single IP address of a cell tower. To block one user, you’d need to block the entire tower. Even when blocks happen, mobile IPs in the U.S. rotate at least seven times per hour. This makes them resilient, fast-cycling, and a critical component of our real-time data infrastructure.
We launched the app quietly, distributing a limited set of sign-up codes without any announcement. Almost overnight, over 50,000 unique devices connected from all over the world.
The atmosphere is electric.
We believe mobile will be the form factor that takes Grass to the scale necessary for live context retrieval. It’s also the most likely path for any future human-in-the-loop work. If we ever need mass quantities of Grass data labeled by humans, it probably happens here.
The Grasshopper
We recently opened the waitlist for our dedicated hardware unit, the Grasshopper.
It’s designed for high-throughput contributions to the Grass network; this is particularly important for web-scale multimodal data collection, as well as the ability for many low-latency connections to concurrently proxy through individual nodes. When we tested the Grasshopper prototype, it was capable of 3–4 megabytes per second of sustained scraping.
This alone is powerful, especially when compared to the Grass desktop nodes, of which the top 10% are limited to 250–750 kilobytes per second. But we’re also thinking about what other kinds of compute might eventually move to the edge: annotation, segmentation, fully hosted browsers for agentic workflows…
No promises yet. But the direction is clear.
It’s also a full-circle moment. Grass started as a Chrome extension with nothing more than a waitlist and a vague landing page in April 2023. There was no product. Just an idea that a decentralized data protocol might be possible, and a small group of people who believed in the vision.
It’s incredible to see the excitement still going strong, and even more incredible that the same community has stuck around the whole way.
Video Search
What started as in-house tooling for our custom curation work is quickly becoming its own product.
Companies will be able to find any moment, object, or action that exists in any video on the public web by searching for what’s happening in the scene, not just in subtitles or metadata tags. The core tech is built to scale to billions of videos, but we’ve limited the closed beta to a small subset for now. AI labs and researchers are already using it, and early demand from other sectors (content creators, adtech, defense) has surprised us.
Each frame is annotated with machine-generated labels, forming a dense layer of synthetic data that can be searched and retrieved via API. The interface we’ve demoed is just a front-end to help people understand what’s possible. The real value is in the annotations themselves, and the ability to query them at scale.
We’re only scratching the surface of what this can become. Some people have expected this to be a breakout product on day one. That’s not how foundational infrastructure works. We’re building something that gets more valuable with scale, and we’re still early in that curve.
What We Care About
There's been some speculation about our stance on the token (why we haven't "marketed" it more aggressively, or said more publicly about its trajectory).
The reality is simple: we're focused on building infrastructure for the next several decades.
We’ve been grateful (and honestly, energized) by how vocal and engaged the community has been. It’s clear that people care deeply about what we’re building. From day one, we’ve done our best to show that we care just as much. We:
- Turned down centralized exchanges that asked for considerable amounts of the token's float (this ensured that those who contributed to the Grass network would be the only ones to own it on day one)
- Refused to let locked tokenholders stake (we were among the first to challenge the status quo by doing this)
- Transferred all of Wynd Labs’ intellectual property to the Grass Foundation before launch
- Ensured that all commercial contracts face the Grass Foundation, and not Wynd Labs (there has never been an "equity" entity that directly benefits from the protocol's commercial success; the token is the protocol's only relevant asset)
We’re also investing heavily in physical infrastructure to reduce our reliance on cloud services. Our datacenter buildout will materially lower costs over time. These are savings that directly benefit the network in the long run.
None of this is flashy, and we’re OK with that. Short-term attention is not what we’re optimizing for. We’re trying to build something that will last forever.
Where This Is Going
Nine months ago, we made a bet that multimodal data collection would be one of the few remaining places where scaling laws still applied. That bet was right.
Since then, demand for vast amounts of diverse video, image, and audio datasets has exploded.
At the same time, the public web is getting harder to access. In the past several months, most of the major platforms hosting multimodal data have started aggressively blocking datacenter IP addresses. This has quietly positioned Grass in a very unique spot.
Even the best-resourced labs (including OpenAI) have said they’re still years away from running their own internet-scale web crawls. They keep getting blocked.
We’ve always believed the solution is a decentralized network that can’t be blocked in the first place.
Right now, nearly all of the demand we see is being driven by training, and there will always be a place for Grass as a supplier of training data for all kinds of ML workflows. We’ve always believed, however, that the end-game is live context retrieval. As more compute cycles eventually move from training processes to inference (one could argue that if this doesn’t happen, AI has failed), Grass will be used for retrieving the context necessary for models to reason, act, and generate in real time.
We think a meaningful share of how people interact with the internet will soon happen through LLMs instead of directly through websites. Nearly all of these interactions will require live access to the web.
Bing shut down its API recently, and Google never offered one in the first place. This was obviously not done due to technical limitations.
We’re still early, but the right foundations are in place.
Appreciate everyone who’s been with us on this journey.
More soon.