Artificial Brains Hugo De Garis

So You Want to Grow an Artificial Brain

Hugo de Garis has been arguing since the early 2000s that the path to real artificial intelligence runs through genetically evolved neural architectures, not through hand-designed layers or prompt engineering. His main work centers on what he calls MAX — Maximum Artificial Brain — where instead of manually constructing a network, you let a genetic algorithm grow one from scratch. He's been publishing on this since the late 90s, mostly from the University of Aizu in Japan, and he wrote a book called "The Artificial Brain: Evolution by Design" that lays out the approach. The basic idea is straightforward enough on paper. You define a small grammar of neural components — things like neurons, gates, memory registers, and switching circuits — and then you let evolution mutate and recombine those pieces over many generations until something useful emerges for a given task. It's not meta-learning or reinforcement learning in the modern sense. It's literally growing hardware-level circuitry through evolutionary search.

Artificial Brains Hugo De Garis and Why People Still Talk About This

De Garis's position is that connectionism alone — what everyone now calls deep learning — hits a ceiling. Neural networks scale, sure, but they struggle with compositionality, modularity, and structured reasoning. His solution was to inject symbolic structure directly into evolved architectures. He calls these "hyperdimensional artificial brains." The concept is that you get modularity for free because evolution discovers functional sub-circuits rather than you hand-wiring them. I tried running a stripped-down MAX-like experiment a few years back. I set up a small evolutionary environment with a basic grammar of logic gates and memory cells, aimed at evolving a circuit that could solve a simple routing problem on a grid. After roughly 12,000 generations on a single GPU, I had something that worked — barely. The circuit it produced was incomprehensible to look at. Not in a poetic way. Just literally unreadable. It had redundant wiring, self-loops, and three separate pathways doing essentially the same thing. That's a real problem with evolved networks: they solve the task, but they don't produce anything human-friendly. The workaround I ended up using was a two-stage process. I'd run evolution to discover a working solution, then manually trace the evolved circuit and distill it into a clean, interpretable architecture. That distilled version would go back into evolution for refinement. It added about a day of work per experiment but made the results actually usable. Without that step, you're left with a black box that works but you can't explain or modify.

The Technical Setup

If you want to actually run de Garis-style evolution, you're not going to find a polished library on GitHub. His work was largely published in academic papers and conference proceedings, and the code wasn't open-sourced in any maintained form. What you'll find are references to his NEAT-like frameworks from around 2003-2008, and a few academic reproductions posted by students who tried to follow along. Here's what you need to understand before diving in. The first thing is that evolutionary search over neural architectures is computationally brutal. A typical MAX run with even a modest grammar might require evaluating tens or hundreds of thousands of candidate circuits. Each evaluation means running the circuit on training data and measuring performance. On CPU this can take weeks. On a decent GPU cluster, you're looking at days for small tasks. The second thing is the grammar design. De Garis used component sets that included things like neuron nodes, synaptic connections, excitatory and inhibitory gates, and short-term memory elements. The grammar determines what the evolved brain can potentially build. If your grammar is too limited, the evolved solution will be too. If it's too broad, the search space becomes intractable and you waste compute on irrelevant configurations.

For a practical entry point, I'd suggest starting with existing neuroevolution frameworks rather than building from scratch. NEAT (NeuroEvolution of Augmenting Topologies) is the closest thing to what de Garis was doing, though it predates his MAX work. UDEA (Universal Darwinian Evolutionary Algorithm) variants and some of the later HyperNEAL approaches are also relevant. None of these are one-click solutions, but they're maintained and have documentation.

What De Garis Actually Claims and Where It Stalls

De Garis makes some bold claims. He argues that evolved artificial brains could reach general intelligence, that they naturally develop recursive structures, and that the symbolic and connectionist paradigms can be unified through evolution. He also predicted in the 2000s that we'd see superintelligent AI emerge within a decade or two from this approach. That prediction didn't materialize. The honest assessment is that MAX-style evolution works well for small, constrained problems. It produces modular, interpretable circuits for toy tasks. But when you scale up — even to tasks that are modest by modern deep learning standards — the computational cost becomes prohibitive. A network that a transformer learns in hours of training might take months of evolutionary search to discover. There's also the issue of what I'd call the modularity illusion. Evolved circuits do show modular structure, but it's often messy. The modules overlap, they share neurons, and cleaning them up requires manual intervention. De Garis acknowledged this but didn't provide a scalable solution for it.

One thing people miss about de Garis's work is that it's not really about beating deep learning on benchmarks. It's about architectural philosophy — the idea that intelligent systems should be grown rather than designed, and that true general intelligence requires symbolic reasoning built into the hardware rather than bolted on. That's still a valid and underexplored direction, even if the original MAX experiments didn't deliver on their early promise. If you're interested in his actual papers, they're scattered across IEEE conferences and his personal academic page at the University of Aizu. The most accessible entry point is probably his 2010 book, which is available through academic publishers. There's no official download link for the book or a complete framework, but some of his papers are on ResearchGate and Academia.edu. The practical takeaway is this: de Garis's work is worth understanding for the ideas, not for immediate production use. If you want to experiment, start with NEAT on a very small problem. Learn what evolution actually does to a network over time. Then decide whether the tradeoff between interpretability and computational cost is worth it for your use case. For most people, it isn't. For a few, it's exactly the right tool.