The Animal Kingdom of AIs

For three years, we’ve been promised one AGI to rule them all. Scale transformers far enough and you get universal intelligence.

But intelligence has never worked that way.

Nature Doesn’t Build One Mind

Birds mastered flight and navigation. Dolphins evolved sonar and social coordination. Humans specialized in symbols and tools.

Each species found its cognitive niche. Each generalizes broadly within that niche. None does everything.

The AI Kingdom Is Already Here

We’re witnessing the birth and evolution of distinct cognitive species:

• CNNs for perception and vision
• LLMs claimed language, open-ended dialogue, and (perhaps) code
• Neuro-Symbolic models handle conversational task execution with guarantees
• GNNs excel at graph reasoning, molecular design, and protein folding

Each architecture creates a different kind of generalizing machine. LLMs can write poetry but can’t guarantee they’ll book the right flight. Task-execution models can complete procedures reliably but can’t create art. Different architectures, different capabilities, different species.

The “Scaling Laws Are Physics” Myth

The idea that eight researchers at Google discovered the AI computation in 2017? That scaling laws are laws of physics?

Picture humanity in 25,000 AD. Or intelligent civilizations elsewhere. Will they all be scaling transformers? Of course not. They’ll have architectures built on physics we haven’t discovered, computational substrates we can’t imagine.

Which AI?

“AI will replace doctors.” Which AI? The diagnostic imager? The drug discovery system? The surgical robot? The patient interaction agent?

CNNs can’t hold conversations. LLMs can’t control robotic arms. Task agents can’t fold proteins. Different jobs need different species.

When someone asks “What do you think about AI?” the only answer is “Which one?”

The Animal Kingdom  

The future isn’t one superintelligent overlord. It’s an entire animal kingdom: zebras and giraffes, eagles and dolphins, each optimized for its own practical domain. Some species will emerge, others will go extinct. They’ll compete, coexist, and create ecologies we can’t conceive of.

Perhaps one day a species will emerge capable of learning across task families, much as humans eventually arose from nature’s diverse animal kingdom, bridging cognitive niches that had been separate for millions of years. But nothing in our current trajectory suggests that scaling today’s architectures will produce such a system.

The question isn’t whether AGI will arrive. It’s which new species of intelligence we’ll discover next, and what computational breakthrough will birth it.