Mass Low-Cost Systems

The mass deployment of small satellites in low Earth orbit is revolutionising defence, communication, and various other sectors. It strikes me that we’re witnessing the beginnings of a similar paradigm shift that could soon apply to swarms of humanoid robots, drones, and autonomous vehicles.

A clear parallel can be drawn with the evolution of military aviation: while expensive aircraft like the F-16 have historically dominated, companies like SpaceX are proving that cost-effective, mass-produced systems—such as reusable rockets—can achieve strategic goals at scale. The same principle is likely to apply to UAVs and autonomous vehicles, where high quantities of relatively low-cost units will prove more effective and versatile than a few expensive, specialised models.

This differentiation will shape both industry and strategy, with profound implications for sovereignty and security. The ability to mass-produce and deploy scalable systems will redefine technological dominance, offering distributed networks of machines that balance cost, efficiency, and adaptability. Whether in space, on the ground, or in the air, this dual approach to innovation feels like the cornerstone of the next wave of technological evolution.

Knowledge and Data Dominance

In the early days of the Internet, we unintentionally created a vast, open reservoir of information accessible to anyone. This led to the development of reasoning and language models trained on Internet-scale data, democratising AI to a degree. However, the current landscape is shifting: private companies are now aggregating their own datasets, focusing on embodied AI and niche applications, creating silos of proprietary data.

This shift raises concerns about data monopolisation. A few dominant actors are amassing immense amounts of personal and task-specific data, granting them unprecedented power to build models tailored to specific modalities and tasks. For instance, ChatGPT’s potential access to vast user interactions could enable it to model not only language but also people’s thought processes, preferences, and reasoning patterns. Such capabilities might enhance reasoning and decision-making but also consolidate knowledge dominance in the hands of a few, shaping the flow of innovation, privacy, and even societal behaviour.

Training Approaches

When it comes to training AI systems, there are three main approaches for data and knowledge capture, each with its own trade-offs:

• Physical World Training: Training systems in the real world provides the most authentic and precise data, but it comes with significant challenges. Real-world deployment is expensive, risky, and slow, especially when dealing with robotics.
• Simulation-Based Training: Simulated environments overcome many of these limitations. While initially expensive and complex to build, simulations offer low ongoing costs, exceptional speed, and unparalleled flexibility.
• Internet-Scale Data Training: Leveraging internet-scale video data provides access to a vast array of real-world scenarios. Such data is particularly useful for understanding behaviours and decision processes but lacks the fine-grained physical details needed for comprehensive understanding.

The future of AI training will likely involve a combination of these approaches. Pre-training on large-scale datasets will create generalisable systems, while fine-tuning in specific environments will optimise agents for desired behaviours. The challenge lies in balancing the cost, scalability, and generalisability of these approaches to develop AI that can adapt effectively to real-world demands.