Delphi Dialogues with Grass Founder: We Are the Secret Data Suppliers of the AI Era

"No matter what kind of AI system the future holds, it will need access to information."

Compiled by: Foresight News

Guest: Andrej Radonjic, Co-founder and CEO of Wyn Labs, Head of Grass

Host: Delphi Digital

In October 2026, Delphi Digital's podcast interviewed Andrej Radonjic, co-founder and CEO of Wyn Labs. As the head of the Grass project, Andrej elaborated on how Grass evolved from a small bandwidth-sharing project into a secret data supplier serving cutting-edge AI labs like OpenAI and Anthropic. Currently, the Grass network has millions of nodes and has delivered over 1 billion GB of web data to AI labs. In this conversation, Andrej disclosed for the first time the complete business structure of Grass, the two major markets of training data and inference data, the huge opportunities in robotic video data, and the newly released Live Context Retrieval product suite.

Key Insights

• "From the perspective of the internet, it looks like ordinary users are browsing normally. But in reality, it’s a massive data infrastructure platform feeding nutrients to the forefront of AI."
• "Honestly, most of the internet designed for humans is difficult for AI agents to access. Websites' default policy is to block all data center IPs directly."
• "Currently, training data and inference data are two separate things, but in the future, they will merge into one. The final model will continuously receive real-time information like a fire hose."
• "What robotic companies want most is to let the model watch a billion videos and learn how to use tools and do chores by itself."
• "The real moat is: for every dollar you invest in infrastructure, the economic output created is greater than that of competitors."
• "Grass's ultimate vision is to become the medium through which AI accesses information. No matter what kind of AI system the future holds, it will need access to information — and Grass is that pipeline."

The following is a compiled summary of the podcast dialogue (content has been edited):

Grass is Not a VPN, but a Secret Data Supplier for AI

Host: First, explain to our listeners what Grass is and how you got into this field.

Andrej: Grass is a network composed of millions of nodes, where users voluntarily join and share their bandwidth in exchange for rewards. On the other end, various data infrastructures use the bandwidth of these millions of users to access web data on a large scale.

How did I get into this field? It was quite random. I studied engineering physics in Canada, focusing on nuclear engineering and photonics. When I graduated, most nuclear engineering jobs were in remote areas of Canada, and in my twenties, I didn’t want to move there. Then I realized that many skills I learned in physics could be applied in finance, so I worked in finance for a few years before returning to pursue a PhD in computational physics.

When the pandemic started, I was studying applied mathematics for my PhD, modeling complex physical equations. During the pandemic, I had a lot of free time and started doing various odd projects. One of them was an arbitrage bot — finding price differences across different websites to make profits. This required a lot of web scraping, and that’s how I fell into the field of web crawling. A few years later, I met two co-founders, and after discussing our experiences, we all felt this was a clear opportunity, and we’ve been working on it ever since.

Why It’s Difficult for Agents to Browse the Web Today

Host: What did you see that others didn’t? Why did you decide to pursue this?

Andrej: The biggest signal was the pandemic. During the pandemic, most businesses moved online, and many companies obtained intelligence — such as price intelligence or even government calculations of CPI — in ways that previously involved sending people with notepads to stores to jot down prices. With stores closed during the pandemic, everyone was forced to spend money online.

Web infrastructure became increasingly advanced, but at the same time, the demand for web intelligence exploded. Many websites realized: we need to make information public so that people can discover us and buy our products. However, we want to build a walled garden that only allows consumers to access it, blocking all data center IPs.

The demand for automated web data infrastructure surged at that time. We felt that this trend wouldn’t stop — as machine learning algorithms became stronger, it would spread across the entire internet.

Host: So what can AI agents not do now? What is becoming increasingly difficult to access?

Andrej: To be honest, most of the internet designed for humans is difficult for AI agents to access, and this situation will persist for a long time.

There are many reasons for this. Many websites block AI crawlers for various reasons — either out of fear of competitors scraping data or concern that AI training will use their content. But regardless of the reason, they find it hard to distinguish who is accessing, so the default policy is a blanket ban: block all data center IPs.

If your agent is running on AWS, every website can detect that it’s an AWS IP. Your agent and a competitor’s price-scraping bot look the same to the website. The default response is to block them directly. We see this situation becoming more common.

This applies not only to agents but also to standard large model usage scenarios, even to the process of training large models themselves — when you need to scrape the web for training data, you encounter the same issues. To obtain the best training data, you need to access as much information as possible, then filter and select the relevant information to include in the model weights.

Training Data vs Inference Data: AI Needs Two Different Types of Information

Host: Your business now includes both training data and inference data. Can you talk about the differences between these two markets?

Andrej: Currently, AI model training works like this: a training cycle takes a long time. Most of the time is spent converting data into a format suitable for training. Then you generate more data based on the model's output, which is called reinforcement learning, and adjust the model based on that data.

However, once the model is trained and deployed in the real world, its knowledge is capped within seconds and cannot continue to learn in real-time. Although continuous learning is being researched, it has not yet been solved.

So there are currently two completely different processes:

• Training process: Scraping massive amounts of internet content to create very useful pre-training datasets. The scale is large, sometimes targeting specific languages, types of files, or specific content for directed scraping.
• Inference process: After the model is trained, it needs to be able to look things up online. You can deploy the model on AWS, but if it’s in a data center, it won’t be able to see much content — for example, content that can only be accessed from specific countries. To access large-scale real-time information, you must route traffic through a network like Grass.

We are very excited about a new emerging field, which is inference data. The core idea is that models need to access contextual information in real-time. When you are using a model, or when a model is called by another model, to make the model useful, it needs to access the latest real-world information.

In the coming years, as continuous learning frameworks mature and can run at scale, these two processes will merge into one. The model will continuously receive real-time information from the entire internet, constantly updating.

Letting AI Watch a Billion Videos to Learn on Its Own

Host: You mentioned that you have indexed over 10 billion public web videos. How are robotic companies using this data?

Andrej: The exact number may be outdated; we have now indexed over 100 billion. I say “indexed” because we don’t blindly download every video — video as a medium is too dense, and storage and processing are expensive. We do targeted scraping.

Robots and robotic data are particularly interesting fields right now. Let me explain the basic logic of machine learning research: if you know what to optimize and the data format is suitable, you will be happy doing machine learning. You throw a bunch of data at familiar algorithms, and they can quickly learn from it.

Now, robotic researchers' ideal state is to let the model watch millions of videos (to be precise, tens of billions), listen to millions of podcasts, and read trillions of words of internet data, then learn how to act and generalize from that information. For example, if I’ve seen a person using a lawnmower, then as a robot, I know how to do it. I’ve seen how they walk, how they move their arms, how they handle tools, and after watching millions of times, as long as the algorithm and settings are right, I will know how to do it myself.

The model from the embodied intelligence company Generalist AI has excited everyone because it may be the first to prove that a model can learn to perform actions just by watching a large number of videos — for example, picking up a specific colored ball and placing it in a cup. They even demonstrated in real-time: a person performed this action, and after watching, the model replicated the same action exactly.

Since then, another embodied intelligence company, Skild AI, has also published papers proving that this behavior can be scaled using web videos. So the demand for web videos is growing very rapidly now.

Live Context Retrieval: Simplifying Anti-Scraping Engineering into an API

Host: You recently released the Live Context Retrieval (LCR) product suite. Can you elaborate on that?

Andrej: Yes, this is our newly released Contents API, which is the first in the entire product suite. There will be a Search product later, and we will also add multimodal capabilities — such as audio listening, image reading, and semantic understanding.

The coolest part of the Contents API is that it was originally a tool for our internal use. The team needed to collect various modalities of training data on a large scale, and sometimes just using residential IP networks wasn’t enough — you need to handle TLS fingerprints, efficiently render browsers on servers, and constantly change request patterns, or else the website will block you. (Note: TLS is the encryption channel between the browser and the website when browsing the internet, ensuring that the data you transmit cannot be intercepted and confirming that the other party is the real website.) To efficiently collect training data internally, our engineers built a very user-friendly web content API that everyone can call directly without having to write crawlers from scratch each time.

Later, we realized that this product had reached top industry levels, better than most things on the market. So we made it the first product open to the public in the LCR suite.

What is it? We simplified the entire process — managing web requests from global residential proxies, loading browsers, configuring anti-scraping strategies — into a single API call. You just need to provide a URL, and it will quickly return the website content to you. You can also specify which country's perspective to access from, and it can output in Markdown or plain text.

The next release will be Search. Why will Search accelerate the adoption of the entire LCR? Because ordinary agents need to access 5 to 50 web pages each time they go online, and even more during complex reasoning. The Contents API is now suitable for those clients who know which websites they want to monitor. Once Search goes live, the model won’t even need to know where to find content — it will just call the Search API and then call the Contents API to extract the content.

Collaborating with Cutting-Edge AI Labs, Keeping Secrets Like Defense Technology

Host: What’s it like to collaborate with these cutting-edge AI labs? How have their needs changed?

Andrej: It’s super cool, man. No exaggeration, AI may be the technology that defines this generation and even a historical era for humanity. Being able to see how these models improve step by step and help elevate the limits of human civilization is incredibly exciting. For many people on our team, this is the most fulfilling and motivating thing.

Of course, resources are highly concentrated, and these labs are filled with the smartest people. Every conversation is an intellectual challenge, and you get to see a lot of cool things. The hardest part is not being able to share with the world — for example, if we participated in a cool project release, we’ve known it was coming for months because we contributed to it. But we can’t say anything.

I mentioned this a few months ago on Handsome's podcast; we often compare it to defense technology: a defense company helping the government find terrorists clearly won’t broadcast what they’re doing. Because if you say it, others can reverse-engineer it and counter it. The current AI industry is similar — everyone is racing to achieve the best capabilities. If everyone knows what data a certain model company is focusing on, they can reverse-engineer what they are doing and compete with them before they release it.

So we often see which direction the entire AI industry will head in the next 6 to 12 months, but we can’t say.

Grass's Business Model: A Foundation Without Owners + Service Company

Host: You recently released revenue proof, and the market response has been very positive. Can you talk about Grass's structure?

Andrej: Grass's structure is indeed quite unique, and it may seem a bit strange to many clients, but this design has very important reasons — all for the sake of incentive alignment.

The audit report was done by an accounting firm, which verified all procurement orders between Grass and major AI labs and other clients, then reconciled them with the actual cash received. Of course, there is a slight time lag — some invoices take 60 days to be paid after they are issued — but overall, it matches up.

Many people may not realize that the entity in this audit report is the Grass Foundation, not an equity entity. The Grass Foundation is a foundation without owners, and all clients — whether purchasing pre-training data, fine-tuning data, or inference data products — ultimately deal with the Grass Foundation. It is a no-ownership structure, managed by independent directors, and any financial transactions require board approval.

I work at Wyn Labs, which is the service provider for the Grass Foundation. Why is Wyn Labs necessary? Because in the real world, you still need to hire people, pay social security, and buy insurance; you need an equity entity to handle these matters. However, Wyn Labs itself does not own any intellectual property — all intellectual property is transferred to the Grass Foundation. Wyn Labs does not directly face any clients, and any collaboration between Wyn Labs and the Grass Foundation requires approval from the board of the Grass Foundation — and these directors have no economic interests in Wyn Labs. Of course, the Wyn Labs team itself are all token holders.

Future Vision: Becoming the Pipeline for AI to Access Information

Host: Last question. If everything develops according to your vision, what will Grass look like in 5 to 10 years?

Andrej: Look at the current pace of development, exponential progress; no one can predict what the world will look like in 5 years. But Grass's mission is clear: we want to become the way AI accesses information.

Whether it’s agents browsing the web or future more complex AI systems needing to obtain information in better and less biased ways — Grass is that medium. What will it look like in 5 years? Who knows? Maybe by then we will be launching satellites to collect information instead of using millions of IPs around the world as digital satellites to see the world from their respective perspectives.

No matter how the world changes, AI systems will need access to information, whether for training, synthesizing new data, or ultimately becoming a massive continuous learning process. Grass's vision is to be that pipeline for accessing information.