The Meat Token: A Unified Theory of Biological Computation

We introduce the Meat Token (MT), a discrete unit of biological cognitive expenditure that serves as the natural dual to the silicon token employed in modern large language models. Where the silicon token is cheap, fast, and emotionally inert, the meat token is expensive, slow, and frequently overwhelmed by feelings. We formalize the conversion ratio between the two, present the MEAT-bench evaluation suite, and demonstrate that on a wide class of physical-world tasks the meat token remains, against all expectation, dominant.

The dominant assumption in contemporary discourse holds that human cognition will be subsumed by silicon. We reject this framing. The human brain remains the only known general-purpose reasoner that can fix a toilet, comfort a grieving relative, and parallel-park, all within the same hour, on roughly 280 kilocalories.

We define the meat token as the minimum increment of focused biological attention. One meat token corresponds, by convention, to the cognitive load of recalling a coworker's name during an elevator encounter. All other tasks are calibrated against this baseline.

A sentence is tokenized into cuts. The phrase “please bring me a sandwich” decomposes, under our standard tokenizer, into four cuts: [BRISKET / CHUCK / RIBEYE / SHANK]. Cuts are not commutative: rearranging them changes the dish.

We evaluated leading frontier models against a single moderately-rested human on the following tasks. Lower meat-token counts indicate greater human efficiency. Lower silicon-token counts indicate greater LLM efficiency. The two are not the same axis.

We observe that human cognition exhibits significant marbling — interleaved layers of fat (intuition) and lean (deliberation) that cannot be cleanly separated. Attempts to remove the fat produce dry, gristly reasoning of the sort one encounters in corporate town halls. Our recommendation is to leave the marbling intact.

Silicon and meat denominate effort in incompatible currencies — the kilowatt-hour and the kilocalorie. We establish a fixed peg between them, the Calorie Exchange Rate, and demonstrate that any AI workload may be losslessly restated as the human metabolic effort it represents.

A corollary, the Maillard Principle, follows: every token an LLM emits to spare a human effort is itself paid for in heat, and that heat is exactly convertible to a sandwich the machine has, in effect, eaten on the human's behalf.

The meat token is not obsolete. On any task that requires moving an atom, comforting a mammal, or correctly interpreting a sigh, it remains orders of magnitude cheaper than its silicon counterpart. Future work will examine dry-aging as a fine-tuning analogue.