Originally published here.

Phelim Bradley is CEO of Prolific, a technology company building the biggest pool of quality human data in the world – and the ultimate platform to access it. Phelim is also a Venture Partner at Pioneer Fund, and worked in genomic medicine and computational biology before founding Prolific.

Certainly not. We’ll only see continued investment. The vast majority of this is coming from the US – revenue growth there is much faster than in Europe right now.

We’re going to see continued investment in AI hubs around the world. London and Paris have the potential to be these hubs, as do other cities in Europe. But today, a lot of growth comes from the US, and those networks are critical for our next phase.

There are still real technical challenges to solve. The most famous models require enormous amounts of training data, which puts strain on servers and energy resources. There’s a trend towards teaching models to be just as good with much smaller data sets. These use less compute and therefore less data. This will make AI tools faster, more efficient, and much cheaper in the long run. Plus, it’s essential for sustainability.

We’re also seeing shifting demands towards the evaluation step in models. Evaluations are like the unit tests: how do you know that the model hasn’t regressed and still provides quality? So beyond building the models themselves, there’s more focus going into their maintenance and continued performance.

But the biggest change we’re working on involves more investment in domain experts and specialists. Large-scale models like GPT are good for broad tasks but aren't perfect for every use case. When training the latest AI models or conducting research, general data isn't enough, and you need responses from people with specific expertise or characteristics.

We’re going to see models that excel in specific niches, and can truly bring the same level of expertise as experienced employees. That means training models on much more focused data sets. And it also means consolidation – infusing AI into industry tools and using real user data from those, rather than broad internet scraping.

High-quality, targeted data from vetted participants leads to more accurate results and helps catch potential issues early. Companies like Carnegie Mellon and Layer 6 have used targeted participant pools to test and improve their AI models.

No, LLMs will serve as core infrastructure, but we're also seeing a growing ecosystem of smaller, open models and distilled versions. The future will likely have both: large models providing the backbone, while specialized models handle specific tasks. Organizations are recognizing the value of this diversity, selecting models based on actual needs rather than their sheer size and power.

The "hype machine" shapes both public expectations and research directions in AI. But the focus needs to be on responsible development rather than just technological advancement.

The real challenge is making sure AI systems are deployed responsibly, which requires attention to data quality and diversity of feedback and input. Transformative change often comes in waves of innovation that are more focused on practical applications and specific use cases. These may generate less hype, but can have significant impacts if not deployed responsibly.

I like Jeff Bezos’ approach: focus on the things that never change. For us that means the breadth and quality of our audience, and the speed of innovation through our platform. We can always improve the data and find new ways of interpreting it.

But this requires a culture of customer obsession. Because it’s not up to us to decide which data is best, we need to understand and solve customers’ problems, whatever they are.