From ads to tasks

We are partnering with leading data-labeling companies and, increasingly, AI labs to collect human preference and evaluation data at scale. This is the data used in post-training, model evaluation, ranking, reward modeling, and other feedback loops that help models become more useful, reliable, and aligned with what people actually want.

I have always been bullish on AI and its potential to expand human capability and increase economic productivity. We are excited to turn part of that growing AI economy into a mechanism through which people around the world can earn mobile data.

Today, much of this work is distributed through established data-labeling and task platforms. That is a useful starting point for us because it gives Airvoy access to task demand while we build scale, improve the product, and learn what types of work are best suited to our users.

Over time, our goal is to build relationships directly with the AI labs and companies that need this data. As Airvoy grows, we believe a large and geographically diverse network of people completing high-quality evaluations can become valuable infrastructure for the AI ecosystem itself.

Unlike expert data, which requires specialized knowledge and therefore has a relatively constrained labor supply, preference data draws on something much more abundant: human judgment. Text models, image models, audio models, multimodal systems, and agents all need evaluation signals, including comparisons, rankings, preferences, and judgments, to improve. As model capabilities scale, the demand for high-quality evaluation does not disappear; in many cases, the evaluation problem becomes more important.

Airvoy can provide something valuable in that market: a large, geographically diverse base of humans able to produce those signals.

The main variable determining how many tasks a user must complete is therefore no longer local advertising revenue, but the marginal cost of delivering mobile data in that country. Wholesale data prices vary substantially because of network competition, infrastructure, spectrum, population density, backhaul costs, and capacity utilization. Data is generally cheaper in markets with dense infrastructure and strong carrier competition, and more expensive where networks are costly to build and operate.

As we grow, we expect both sides of the equation to improve. More users should make Airvoy a more useful partner for companies that need human evaluation at scale, while greater data volume should improve our ability to negotiate connectivity costs. The combination of those two things is what we believe can make the model significantly better over time.

The long-term goal is very simple: we want someone anywhere in the world to be able to complete one useful task in less than a minute and earn a meaningful amount of mobile data.

A concrete target for us is 100 MB for a single task, whether that person is in India, the United States, Peru, or somewhere else. We are not there today, and the economics will not become identical in every country overnight, but that is the direction we are building toward.

I think this is achievable because the value of human evaluation can be substantially higher than the value generated by watching an advertisement. With ads, there was a hard ceiling imposed by advertising economics, particularly in lower-CPM markets. Human-preference data gives us a much larger economic opportunity and a path toward rewarding users based more on the value of the work they perform than on where they happen to live.

Our next phase is about scale. We want to grow Airvoy into a network of hundreds of thousands of active users across many countries, demonstrate that those users can reliably produce useful human-preference data, and use that scale to build increasingly direct relationships with the labs and companies training and evaluating AI systems.

Ultimately, we would like Airvoy to work directly with those organizations. If a lab needs hundreds of thousands of people across different countries, languages, cultures, and demographic groups to evaluate a model, compare outputs, rank results, or provide preference data, we want Airvoy to be one of the places they can go.

If we can get there, the model becomes much more powerful. AI companies get access to a distributed global evaluation network, and users get access to connectivity in exchange for a very small amount of useful work. That is the opportunity we are pursuing.