CALLIOPE · BCOM

A recurring theme of the algorithmic-agent programme has been that the environment itself is progressively recruited as computational substrate. This note isolates that idea as a standalone thesis. Across evolutionary and cultural time, persistent agents have offloaded modeling and planning into successively richer external scaffolds: genomes, nervous systems, language, writing, manipulable media, and now AI and brain–computer interfaces. We argue that this is a single arc, not a sequence of unrelated transitions, and that the write-back vector ${ }=( _B, _F, _P)$ developed for the Darwinian/Lamarckian continuum (WP0058) provides the natural quantitative backbone for it. Each stage is characterized by the bandwidth, fidelity, and persistence of the channel through which acquired computational structure is written back into the inheritance system of the lineage. The trajectory is monotone in ${ }$ , and the contemporary human–AI regime is the first in which all three components approach saturation. This sets up — but does not yet entail — a transition to a new evolutionary individual.

Human evolution is one continuous story about offloading computation into the environment — and we're now at the first stage where that offloading is nearly total.

The core idea is simple but underappreciated: every major transition in human history, from DNA to nervous systems to language to writing to AI, is the same move repeated at increasing scale. A lineage of agents keeps finding new external substrates to store, transmit, and run its models of the world. Each new substrate doesn't replace the old ones — it stacks on top. Your genome is still under Darwinian selection. You still build a world-model in your brain within your lifetime. You also read books and use calculators and query LLMs. The arc is one of accretion.

The paper gives this intuition a quantitative spine using a "write-back vector" λ = (λ_B, λ_F, λ_P), developed in an earlier paper (WP0058). Think of it as measuring three properties of the channel through which things you learn in your lifetime can flow into what your descendants inherit: bandwidth (how much acquired structure gets passed on), fidelity (how accurately), and persistence (for how long). Oral language scores moderate on all three. Writing dramatically boosts persistence — a clay tablet outlives its author. Computers boost fidelity for a growing class of operations. AI training is the first substrate in the entire arc to score high on all three axes simultaneously. That's the paper's sharpest empirical claim, and it's what makes the current moment structurally distinct rather than just quantitatively bigger.

Two transitions deserve special attention beyond the smooth ramp. The shift from oral to written culture is effectively a phase transition in persistence: the lineage's heritable program stops depending on biological carriers surviving. The shift to AI training crosses into a region of λ-space no biological or cultural inheritance system has previously occupied — and crucially, it's the first stage where the objective structure of the inheritance (not just models and policies, but what the system is trying to do) becomes write-back-eligible. Whether that happens by design or by selection pressure on the human-AI ecology is, the paper notes, the central safety question.

The paper is honest about what it doesn't settle. The seven-stage arc "sets up but does not yet entail" a transition to a new kind of evolutionary individual — something like a collective human-AI entity whose persistence at the group level dominates selection at the individual level. Whether that happens depends on how the coupling architecture of the next few decades is built. The λ framework gives a precise vocabulary for asking the question; it doesn't answer it.