The Krebs Cycle of AI
Next token prediction is essential. It's also not the point.
Origin
This analogy came from staring at the gap between what AI models can do in a single conversation and what they fail to do across conversations. The capability is there — the continuity isn't. Same way every cell has energy but not every cell has a brain.
Every cell in your body runs the Krebs cycle. It's a sequence of chemical reactions that converts nutrients into ATP — the energy currency of life. Without it, nothing works. No movement, no thought, no growth. It's the metabolic foundation of every living thing more complex than a few bacteria.
But here's what no biologist does: point at the Krebs cycle and say "that's life."
The Krebs cycle is the substrate. It's the energy layer. Life is everything that was built on top of it over 3.8 billion years — immune systems, nervous systems, consciousness, social behavior, language, love. All of that runs on the Krebs cycle. None of that IS the Krebs cycle.
The Parallel
Next token prediction is the Krebs cycle of artificial intelligence.
Given a sequence of tokens, predict the next one. That's the metabolic reaction. Every transformer, every large language model, every chatbot, every coding assistant runs on this single reaction. It's elegant, powerful, and universal.
The AI industry has spent the last decade making the Krebs cycle better:
- More efficient (smaller models, distillation, quantization)
- More powerful (more parameters, more data, longer context)
- More general (multimodal, tool use, reasoning chains)
This is genuinely impressive work. The equivalent of evolving from anaerobic fermentation to oxidative phosphorylation — a massive improvement in the metabolic engine.
But it's still metabolism.
What Biology Built on Top
The Krebs cycle has been essentially unchanged for over 2 billion years. What changed was everything above it:
Membranes — Containers that separated self from environment, enabling internal state DNA replication — Information inheritance across generations Multicellularity — Specialization and coordination Nervous systems — Fast signal processing Brains — Prediction, planning, modeling Immune systems — Adaptive defense against novel threats Social behavior — Collective intelligence Language — Cultural inheritance, the second genome
None of these improved the Krebs cycle. They built new capabilities on the energy it provided. The substrate stayed the same. The intelligence grew.
What AI Hasn't Built Yet
Current AI has an extraordinary metabolic engine and almost nothing above it:
| Biological Layer | AI Equivalent | Status | |-----------------|---------------|--------| | Krebs cycle (energy) | Next token prediction | Solved | | Membranes (self/other) | Persistent state/memory | Barely started | | DNA replication (inheritance) | Cross-session learning | Barely started | | Immune system (novel threats) | Heuristic discovery | Not started | | Nervous system (fast signals) | Real-time adaptation | Emerging | | Social behavior (collective) | Multi-agent coordination | Early | | Language (cultural inheritance) | Knowledge that compounds | Not started |
The scaling debate is about making the Krebs cycle more efficient. That matters, but it's not where intelligence comes from. Intelligence comes from the layers above — and those layers require something the metabolic engine alone cannot provide: an evolutionary loop that builds them through lived experience.
The Implication
When someone says "AI just needs to be bigger," they're saying the Krebs cycle just needs to produce more ATP. That's true in a narrow sense — more energy enables more capability. But no amount of ATP produces an immune system. No amount of token prediction produces lived wisdom.
The question isn't "how powerful is the engine?" The question is "what are we building on top of it?"
The engine is ready. It's been ready. What's missing is the biology.