Artificial intelligence has reached astonishing heights. Large language models (LLMs) and diffusion systems now dominate the landscape, powering everything from chatbots to image generation. Their success is undeniable — they are extraordinary tools for recognising patterns and producing useful behaviour at scale.
But cracks are showing. These systems remain statistical black boxes: difficult to interpret, costly to deploy outside data centres, and nearly impossible to certify in safety-critical environments. As regulators, researchers, and engineers ask tougher questions about reliability and explainability, the shortcomings of today’s AI approaches are becoming harder to ignore.
This is why we are returning to biology. Not out of nostalgia, but out of necessity. Biologically faithful models offer something that black-box AI cannot: predictable behaviour, causal transparency, and a pathway to hardware-friendly, certifiable systems. Advances in computing and neuroscience now make it possible to build such models on ordinary machines — something that simply wasn’t practical in past decades.
The central question isn’t whether biology can “beat” AI. It’s whether we have the right tools for the right problems. For large-scale prediction, LLMs remain unmatched. But for interpretability, certifiability, and efficient edge deployment, biological modelling is the toolset we’ve been missing — until now.
What “Synthetic Intelligence” Meant Then
When people hear “synthetic intelligence” today, it can sound like something entirely new. But the idea has been around for decades — only, what was attempted in the 1980s and 1990s bore little resemblance to true biological modelling.
Back then, three main approaches carried the “synthetic intelligence” label:
- Symbolic AI and Expert Systems.
These were built on rules, logic trees, and formal languages like Prolog or Lisp. They could solve narrow, well-defined problems but were brittle — once conditions shifted, their rigid rule sets collapsed. - Early Connectionism.
Basic perceptrons and primitive backpropagation networks made headlines, but they operated with abstract math nodes rather than any real biological fidelity. They lacked ion channels, dendrites, plasticity — the physics of neurons was absent. - Neuromorphic Chips.
Experimental analogue VLSI circuits mimicked spikes in very crude ways. They were clever engineering exercises but simplified neurons so drastically that they could not serve as faithful models of brain dynamics.
These systems failed not because “biology doesn’t work,” but because biology wasn’t really being modelled at all. The attempts relied on shortcuts, limited by the tools and compute of their era. When those shortcuts hit their limits, the field entered the so-called “AI winters.”
What “Synthetic Intelligence” Means Now
Today, the phrase “synthetic intelligence” carries an entirely different meaning. Rather than symbolic shortcuts or crude approximations, the focus is on biological fidelity — building models that operate according to the same mechanistic rules as real neurons.
Our approach is grounded at the Hodgkin–Huxley (HH) level, where ion channels, gating variables, and compartments define the physics of spiking. That fidelity unlocks behaviours impossible to capture in the abstract systems of the past.
Key features include:
- Biophysical Fidelity.
Models operate at the HH ion-channel level, not abstract math nodes. They capture dendritic processing, compartmental dynamics, and conductance-based spiking. - Beyond Simple Spikes.
Plasticity, pruning, hormones, and even neurogenesis can be incorporated, enabling adaptive behaviours rooted in biology. - Persistent Memory.
Drive-based hippocampal memory allows persistence beyond transient weight changes — offering mechanisms closer to real cognition than backprop nets. - Validation-First.
Every claim is benchmarked against canonical ODE solutions and electrophysiological standards. This ensures results are reproducible and evidence-based, not speculative.
This is not “old neuromorphic hype.” It is a fundamentally new foundation, made possible by advances in neuroscience, software toolchains, and commodity compute. Where earlier attempts only hinted at biology, modern synthetic intelligence is biology — finally feasible at scale and on everyday hardware.
Why Go Back to Biology Now
The obvious question is: why revisit biology at this point in AI’s evolution? The answer is that the problem set has shifted, the tools have matured, and the demands on intelligence systems are no longer the same as they were when large-scale machine learning first took over.
1. Different Problem Set
Large language models excel at scale: predicting patterns, generating fluent text, synthesising images. But they don’t provide mechanistic explanations, certifiable safety, or efficient execution at the edge. When LLMs were first breaking through, nobody cared about those things — the question was: can we get useful behaviour at scale? Now that these models dominate, their cracks are visible.
2. Practical Maturity of Tools and Compute
In the 1990s and even 2000s, running HH-class simulations required supercomputers or bespoke neuromorphic chips. Today, commodity multi-core CPUs and GPUs, combined with modern compilers and solver techniques, make it possible to run faithful models on laptops or modest servers. The hardware finally matches the biology.
3. Demand for Interpretability and Certification
Regulators, safety-critical industries, and explainability researchers are asking for models that can be reasoned about, tested, and certified. Black-box predictors don’t meet that bar. Biology-rooted models are inherently mechanistic: every parameter corresponds to something measurable, perturbable, and auditable.
4. Complement, Not Replace
This isn’t a zero-sum game. Statistical models remain the best tool for scale-based pattern recognition. Biophysically faithful substrates are the right tool for causal reasoning, certifiable behaviour, and safe deployment. The future isn’t biology or AI — it’s biology and AI, each applied where it fits best.
How We Differ From Other Neuromorphic Projects
It’s true that spiking neural networks and neuromorphic chips are not new ideas. Many groups have explored them over the past two decades, often with impressive efficiency. But most of those efforts have relied on simplified neuron models and black-box optimisation. That approach trades away biological realism for speed — useful in some contexts, but inadequate where fidelity, transparency, and reproducibility are critical.
Our focus is different:
- Fidelity First.
We model at the Hodgkin–Huxley, multi-compartment, dendritic level. That means physics-based neurons, not just spike events or integrate-and-fire abstractions. - Commodity Performance.
The engine is designed to run on ordinary hardware: laptops and modest servers, not only HPC clusters or bespoke neuromorphic silicon. - Deterministic Hardware Mapping.
Solver primitives are built to translate cleanly into finite-state machine representations. This makes FPGA or ASIC prototyping straightforward while preserving causal semantics. - Validation-Driven.
Every model is benchmarked against canonical ODE solutions and electrophysiological standards, ensuring claims are evidence-based and reproducible. - Tooling for Reproducibility.
With BioSynapStudio, the IDE, we unify model authoring, experiment setup, visualisation, and benchmarking. Collaborators don’t just get results — they get a workflow that ensures repeatability.
Where others prioritise efficiency through simplification, we prioritise faithfulness plus reproducibility plus hardware-friendly semantics. This solves a different class of problems: scientific validation, safety-critical control, and certifiable behaviour in constrained compute environments.
Why Biological Models Yield Transparency
One of the strongest arguments for returning to biology is that it naturally delivers transparency. Unlike statistical black boxes, biophysically faithful models are mechanistic: every parameter has meaning, every behaviour has a causal trace.
Mechanistic Parameters
Conductances, ion channels, dendritic compartments, gating variables — each corresponds to a measurable physical element. When you adjust one, you know exactly what it represents, and you can compare it directly to electrophysiological data.
Causal Structure
Behaviour emerges from local, well-defined rules: synaptic weights, time constants, and plasticity dynamics. Interventions are traceable, so when the system changes, you can explain why.
Fewer Hidden Degrees of Freedom
LLMs contain millions of embedding dimensions that are mathematically powerful but semantically opaque. Biophysical models, by contrast, operate with a limited set of biologically grounded variables, reducing interpretability gaps.
Ablation and Provenance
With biological models, you can ablate a synapse, ion channel, or compartment and observe principled, interpretable effects. Provenance of behaviour can be traced to explicit circuits or dynamics, not buried in a statistical soup.
Hardware Mapping That Preserves Semantics
Because solver primitives are designed as deterministic state machines, execution on silicon preserves causal order. This means that what you test in software can be audited, certified, and trusted when implemented in hardware.
In short, transparency is baked into the substrate, not bolted on after the fact. This makes biology-rooted synthetic intelligence far better suited for domains where auditability and certifiability are non-negotiable.
Market Perspective
The scientific community has always been the proving ground for new computational paradigms. That is where we are starting: by providing researchers and labs with a platform that reduces costs, improves reproducibility, and enables mechanistic fidelity without the need for HPC budgets. But the opportunity doesn’t end there.
Today: Research and Education
- Unified Control Centre.
Scientists currently juggle NEURON, NEST, Brian2, and MATLAB — each with its own quirks, licensing, and learning curves. BioSynapStudio aims to unify those workflows in a single environment that delivers HH-class fidelity on commodity hardware. - Lower Costs.
Running on laptops and modest servers reduces HPC spend and reliance on expensive MATLAB toolboxes. This directly lowers costs for labs and departments. - Accessible Licensing.
University-wide licensing models, similar to MATLAB’s academic deals, can make biologically faithful simulation broadly available for teaching and research.
Tomorrow: Adjacent Research Markets
- Pharma and Biotech.
Neuron models can be used for drug discovery, neuro-modulation, and toxicology screening. - Neurotechnology and Brain–Machine Interfaces.
Startups developing implants or interfaces need faithful simulation tools to prototype safely and efficiently. - Educational Platforms.
Deploying synthetic intelligence models across universities can seed the next generation of computational neuroscience research.
Longer-Term: Commercial and Industrial Applications
- Safety-Critical Systems.
Aviation, automotive, and defence require certifiable control systems. Biophysical models provide interpretability and auditability where black-box AI falls short. - Edge AI and Embedded Systems.
In devices where LLMs are too large and power-hungry, biologically faithful SNNs can provide efficient, transparent intelligence. - Neuromorphic Hardware Vendors.
Hardware companies building the next generation of chips need “killer apps” that validate their silicon. Faithful synthetic models are the perfect match.
The trajectory is clear: start small with research and education, expand into adjacent markets that need biological fidelity, and ultimately address billion-dollar industries where safety, efficiency, and certification are essential.
Conclusion
It is easy to dismiss biologically inspired intelligence by pointing to past failures. The professor’s argument — that synthetic intelligence was tried in the 1980s and 1990s and didn’t work — confuses categories. What failed back then were symbolic systems, crude perceptrons, and simplified analogue chips. They weren’t biological fidelity, and they collapsed under their own shortcuts.
Today the context is different. We now have:
- Decades of neuroscience that map real ion channels, dendrites, and plasticity mechanisms.
- Commodity compute powerful enough to run Hodgkin–Huxley–class models outside of supercomputers.
- A growing demand for interpretability, certifiability, and safe deployment — requirements that black-box AI cannot meet.
This isn’t about nostalgia. It’s about pragmatism. LLMs solved large-scale pattern prediction, but they left critical gaps: mechanistic explanations, auditability, and efficient edge deployment. Biologically faithful synthetic intelligence fills those gaps, complementing rather than replacing statistical AI.
The path forward is clear: validation first, traction in research, then expansion into adjacent markets, and eventually the large-scale industrial applications that demand safety and certification. The early science market may be small, but it is fertile ground for building credibility, generating revenue, and proving that this new foundation is not only viable but necessary.
Biology was not ready to be modelled in the 1980s. Today, it is. And the reasons for going back to it now are stronger than ever.
