Beyond brain-inspired compute: engineering persistent intelligence

Frontier compute is often discussed in terms of hardware: greater processing power, specialised accelerators, neuromorphic chips and architectures capable of performing computation more efficiently. But biology suggests another frontier. What if we look beyond reproducing the efficiency of biological systems and investigate the computational mechanisms that allow them to learn, adapt and maintain behaviour over time?

At Qognetix, this question is shaping our work on Persistent Intelligence: exploring whether biologically grounded computational mechanisms can provide capabilities that are difficult to achieve through prediction-led AI architectures alone.

From biological inspiration to biological mechanisms

Biological intelligence is not produced by isolated predictions. It emerges from systems that maintain internal state, adapt through experience and continuously interact with their environment. Mechanisms including neural dynamics, plasticity, sensorimotor learning and internal regulation contribute to behaviour over time.

Our research explores how these principles can become computational mechanisms in their own right. Rather than using biology simply as inspiration for more efficient computing, Qognetix is developing SIBINE, a biologically grounded framework for investigating persistent neural dynamics and adaptive behaviour. The objective is not to reproduce a brain, but to determine which biological mechanisms are computationally useful for building intelligence that persists, learns and adapts through ongoing interaction.

Digital twins as experimental environments for intelligence

Digital twins provide a powerful way to investigate these questions. They are commonly used to model physical systems, test scenarios and validate behaviour before real-world deployment. We are exploring an additional role: using digital twins as controlled environments that expose an intelligence to challenges it cannot yet solve.

Within Qognetix, this creates a development loop. A biologically grounded system encounters a behavioural challenge inside a digital environment; its response reveals missing or inadequate capabilities; biological mechanisms can then be developed or refined before the system is re-tested under controlled conditions.

The digital twin therefore becomes more than a validation environment. It becomes an experimental instrument for discovering which computational capabilities the intelligence itself requires.

What DAKTARi exposed

We recently tested this approach using DAKTARi, Qognetix's digital driving environment. The initial benchmark was valuable precisely because the system did not improve as expected. Instead, the experiment exposed a missing sensorimotor learning capability within SIBINE.

Rather than hard-coding better driving behaviour, we introduced a generic, biologically inspired cerebellar forward-model layer and supporting mechanisms. Subsequent testing showed prediction error reducing through experience, providing evidence that the new mechanism was learning from the relationship between action and sensory consequence.

That does not yet demonstrate retained driving competence, which remains the subject of further controlled testing. More importantly, the experiment demonstrated the value of the methodology: a behavioural environment exposed a biological capability gap that conventional performance benchmarking alone might simply have recorded as failure.

An engineering loop resembling biological adaptation

There are parallels with biological adaptation: environmental challenges create pressures that expose capabilities an organism needs to function effectively. We are not attempting to reproduce evolution, but to use increasingly complex environments to help identify which biological mechanisms are computationally valuable.

Driving exposes particular sensorimotor requirements. Robotics, industrial systems, energy networks or other environments should present very different challenges. As the range of digital twins expands, the same methodology could systematically expose different deficiencies and guide the development of increasingly capable persistent intelligence.

More capable intelligence creates an execution problem

As intelligence becomes more persistent and adaptive, however, capability alone is not enough. Systems that retain state, learn through interaction and influence real environments create an execution challenge: how do we ensure that what they can do remains bounded, observable and subject to intervention?

This leads to Qognetix's second development track, Trusted Execution. It focuses on the runtime mechanisms required to constrain actions, preserve evidence of how decisions developed, support replay and provide intervention pathways when behaviour moves outside acceptable boundaries.

These requirements are not unique to biologically grounded intelligence. They increasingly apply wherever AI-generated decisions move beyond prediction and begin influencing consequential systems in the real world.

The convergence: Persistent Intelligence + Trusted Execution

These two tracks address complementary questions. Persistent Intelligence asks how computational systems might develop and retain useful capabilities through ongoing interaction. Trusted Execution asks how increasingly capable systems can operate within explicit, inspectable and enforceable boundaries.

For Qognetix, developing them together is important. More persistent intelligence should not require accepting less control, while stronger governance should not depend on preventing systems from adapting.

This suggests a broader requirement for frontier compute: capability and governability need to advance together. The challenge is not only to create new forms of intelligence, but to build the execution infrastructure needed to trust them as they move from controlled experimentation towards real-world use.

What this means for UK frontier compute

The UK's frontier-compute opportunity extends beyond increasing compute capacity or developing new processors. It also includes the architectures, experimental environments and execution infrastructure that determine how new forms of computation can be developed and safely applied.

Digital twins could play an important role in that ecosystem, providing controlled environments in which emerging computational architectures encounter increasingly realistic challenges before consequential real-world deployment.

Our work at Qognetix is exploring one part of that frontier by connecting biologically grounded computation, persistent intelligence and governed execution. The next frontier may not simply be computing faster or more efficiently, but learning how to engineer computational systems that can persist, adapt and act while remaining governable as they do so.

Nic Windley is Co-founder and CEO of Qognetix, a UK deep-tech company developing new infrastructure for persistent and trusted intelligence. Qognetix operates across two complementary development tracks: Persistent Intelligence, exploring biologically grounded computational mechanisms for adaptive intelligence, and Trusted Execution, focused on the runtime infrastructure required to keep increasingly capable AI systems bounded, observable and governable as they operate in the real world.

Technology and Innovation programme activities

techUK bring members, industry stakeholders, and UK Government together to champion emerging technologies as an integral part of the UK economy. We help to create an environment where innovation can flourish, helping our members to build relationships, showcase their technology, and grow their business. Visit the programme page here.