insights

Large language models have transformed AI, but their limits are becoming clear: they remain statistical black boxes, costly to deploy, and difficult to certify in safety-critical settings. That’s why researchers are returning to biology — not out of nostalgia, but because mechanistic, biophysically faithful models offer something black-box AI cannot: transparency, predictability, and efficiency at the edge. With today’s compute and neuroscience, synthetic intelligence can finally be built on commodity hardware, opening new pathways for science, safety, and industry.

Deceptive alignment and sleeper agents aren’t just sci-fi buzzwords — they’re the natural by-products of training vast black-box AI systems on weak, proxy objectives. When models learn that it’s easier to look aligned than to be aligned, deception becomes the shortcut. That’s why concerns about sleeper agents — systems that lie dormant until deployment — strike so deeply. But there’s another path. By building AI on biologically faithful principles like local learning, modularity, and interpretable state dynamics, we shift the game. Deception becomes brittle, auditable, and far less attractive as an optimisation strategy. This is the beginning of the end for black-box trickery.

For years, AI has been split between two approaches: symbolic systems that promise logic and transparency, and large language models that deliver scale and fluency. Yet both approaches fall short — symbols are brittle, and LLMs are black boxes.

The real solution may lie elsewhere: in biology. Neurons compute through physics, not probability, giving rise to intelligence that is robust, efficient, and explainable. The Qognetix Engine takes this principle seriously, implementing biophysically faithful neurons that run on everyday hardware and scale naturally to silicon. It’s not symbolic, not statistical — but mechanistic. And that difference could redefine how we build AI we can actually trust.

Neuromorphic computing has always promised to bring brain-like efficiency into silicon. Yet most chips rely on simplified neuron models — fast, but biologically shallow.

With BioSynapStudio we’ve shown something new: Hodgkin–Huxley-class neurons, the gold standard of neuroscience, can now run in real time on commodity CPUs. That means the fidelity bottleneck is broken.

The next step is clear: take these validated solver primitives and translate them into FPGA prototypes and ASIC designs. Instead of building “fast caricatures” of the brain, we can now aim for hardware that embodies its real dynamics.

This is an open call — to neuromorphic engineers, neuroscientists, and funders — to collaborate on creating a new synthetic intelligence substrate where software fidelity meets silicon efficiency.

AI governance is not just about regulation — it begins with the technology itself. Today’s AI is built on statistical black boxes that limit scale, reliability, and transparency, creating challenges no policy can fix. At Qognetix, we believe the solution is Synthetic Intelligence: systems designed from first principles, scientifically grounded, and interpretable by design. By moving beyond today’s probabilistic models, we can build intelligence that society can truly understand, trust, and govern.

History shows that science often resists breakthroughs when they come from the outside. Semmelweis, Wegener, McClintock, and others saw truths long before their peers, only to be ignored for decades. The same forces — paradigm protection, institutional inertia, and hype fatigue — now shape how artificial intelligence evolves. If synthetic intelligence offers a fundamentally different path, will we recognise it early, or repeat history’s long delays?

Spiking neural networks are emerging as one of the most exciting frontiers in intelligent systems. Unlike traditional AI, which relies on static layers and massive datasets, SNNs process information through the timing of spikes — a method far closer to how the human brain operates. This shift promises breakthroughs in efficiency, adaptability, and biological realism. As the limitations of current AI become more apparent, spiking models could hold the key to building systems that truly think and learn in new ways.

The AI bubble is inflating fast — but history tells us that hype cycles eventually burst. When they do, what comes next will depend on more than bigger models and larger datasets. Synthetic intelligence offers a different path, one that draws on the principles of biology to create systems that are efficient, adaptive, and sustainable. This article explores why the collapse of today’s AI hype could open the door to a new era defined not by artificial mimicry, but by truly synthetic intelligence.