Where is the AGI?

Transformers have pushed neural nets to a ceiling that already feels like science fiction: language, images, music, video—each benchmark we once called "AI-complete" is now a demo you can run on a laptop. We congratulate ourselves for seeing "sparks" of general intelligence, yet a mosquito—whose entire brain fits inside the eye of a needle—still beats our best vision-language models at real-world scene understanding and our most expensive humanoids at embodied common sense. One insect, a few microwatts of power, no pre-training, no H100 cluster.

The gap is humbling. It suggests that scale and gradient descent alone won't finish the job. We need something closer to continual, on-device learning and memory that re-wires itself as effortlessly as a pupal brain remodeling into an adult. Maybe the next breakthrough isn't a bigger GPU but a new interface that hijacks biology itself—running circuits through living neurons that already know how to sip energy and still out-see, out-fly, and out-adapt us.

Where We Actually Stand

Transformers have indeed revolutionized artificial intelligence. They've given us systems that can understand and generate text, images, audio, and video in a way that looks intelligent and creative. In terms of pattern recognition, reasoning within short contexts, and mimicking human communication — we're near the ceiling of what this architecture can do.

But this is not general intelligence yet. These systems don't truly understand the world; they compress correlations. They don't build persistent world models, have long-term goals, or form memories beyond the context window. The "spark" of AGI is visible, but it's mostly a flicker of approximation rather than cognition.

The Gap: Optimization vs. Innovation

Evolution achieved extreme efficiency — a few milligrams of organic matter running on microwatts can outperform trillion-parameter models in adaptive perception and motor control. The gap isn't just computational efficiency; it's architecture and embodiment.

We can't just scale transformers to AGI. Optimization will help (quantization, sparse activation, neuromorphic chips, etc.), but architectural breakthroughs are needed:

Continuous learning: Models that adapt in real-time without retraining.
Dynamic memory systems: Persistent, self-organized memory like the hippocampus–neocortex interplay.
Embodied cognition: Integration of physical interaction and feedback loops, not pure simulation.
Hierarchical world models: True understanding of causality, not correlation.

The Bio-Hybrid Frontier

Using biological substrates (like insect brains) as computational or training scaffolds — is actually being explored at the edge of "wetware AI." There are startups and labs (like Cortical Labs) that use cultured neuron networks interfaced with silicon. The neurons can learn to perform simple tasks like playing Pong, showing organic adaptability far beyond current deep learning.

If this bio-silicon interface matures, we might see hybrid systems where biological computation provides flexible, energy-efficient intelligence while digital layers handle precision and scale.

The Road to AGI

So, how close are we? We're probably decades away from true AGI — not because we lack compute, but because we don't yet understand intelligence deeply enough to replicate it. The next step likely isn't a bigger model — it's a different kind of model, one that blends continual learning, physical grounding, and memory in ways nature already solved.