EMBEDDED TECHNOLOGY The shift from gigantic models
to collaborative intelligence The pursuit of artificial intelligence has, for years, been centred on a singular vision: ever-larger models fed by endless streams of data and immense computational power.
By RTI W
e have already witnessed astonishing breakthroughs from these monolithic designs, such as the powerful large language models (LLMs) that have transformed our digital lives.
Nevertheless, an expanding consensus within the AI community, encompassing researchers and industry leaders, is now recognising a fundamental truth: these scaling laws are exhibiting diminishing returns. The time when making models bigger would make them look better is ending. Limits are being hit in the attempt to generate truly novel, high-quality data, and even with more parameters, the gains in reasoning and problem-solving for novel tasks are becoming increasingly marginal. The truth is, though, that the economy is also really tough. Basically, every time there’s a small improvement, it needs a huge amount of investment, and that makes it really hard to keep on growing.
This crucial moment calls for a strategic pivot. The AI field is no longer pursuing the fantasy of a single, all-encompassing “super- brain.” Instead, it is now looking at a more meaningful, and potentially more realistic, idea: the idea of several expert agents working together, or a “Society of Mind.” This isn’t a new idea. Decades ago, Marvin Minsky proposed that intelligence does not stem from a single, perfect principle, but rather from the ‘vast diversity’ of countless simple ‘agents’ working together. What was once considered a philosophical framework is now proving to be uncannily prescient for practical AI design.
Imagine, instead of having one colossal model attempting to perform every task, you could utilise a collaborative network of specialised AI agents. Each agent would be an expert in its field, whether that be visual processing, language generation, logical reasoning or interacting with the physical world. This ‘mixture of experts’ architecture, in which a gating mechanism routes tasks to the most appropriate specialist, would allow for far more efficient use of computing power
40 JUNE 2026 | ELECTRONICS FOR ENGINEERS
and data, potentially unlocking capabilities that monolithic models simply cannot achieve.
However, these multi-agent systems can only thrive if they have a fast and reliable way to share knowledge. This requires an explicit communication fabric, or nervous system, to coordinate and collaborate. With this communication layer in place, the agents can collectively tackle complex, multi-step tasks with greater accuracy and reliability, and improve the quality of their output compared to any single model. This approach also introduces crucial benefits such as redundancy and robustness: if one agent falters, the others can compensate, making the system as a whole more resilient. Most importantly, the right communication layer can make the decision-making process not only traceable, but also data-centric, making debugging and alignment with human intent far easier.
The future of AI: Physical Intelligence It is not just about digital reasoning; it is also about acting within the physical world. This involves real-time decision-making in environments such as autonomous vehicles, surgical robotics and power grids. Physical intelligence redefines AI’s role, transforming it from a mere observer to an active participant in complex systems. Realising
this potential requires collaborative, multi- agent architectures underpinned by robust communication infrastructure. To achieve this vision, we must move beyond ever-larger models towards more collaborative and specialised architectures. Multi-agent, society-of-mind approaches are gaining rapid traction, with networks of expert systems working together to solve complex problems. However, orchestrating such diverse agents effectively requires a proven communication backbone that already connects safety-critical systems today.
RTI Connext: Introducing AI into existing systems
RTI Connext communication software is already providing the “nervous system” that connects intelligent components to the physical world. This software is currently being used in over a million autonomous vehicles, NASA’s launch and recovery system, and air traffic control across Canada and the North Atlantic. Its proven track record in high-stakes, real-time environments makes RTI uniquely suited to helping teams embrace a ‘drop-in AI’ approach. Essentially, RTI Connext software allows teams to use the data model as the main interface, enabling new AI features to be integrated consistently and flexibly without the need to modify existing components or carry out costly redesigns.
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46 |
Page 47 |
Page 48 |
Page 49 |
Page 50
