# [Mike Randolph — M Raige](https://mikerandolph211012.substack.com/) # What the Bird Eye Buys (FF 6 GLS Introduction) ### The Persistent Systems Framework reads bird metabolism and frontier AI with the same structured vocabulary. Here's what it caught. [**Mike Randolph — M Raige**](https://substack.com/@mikerandolph211012) **May 15, 2026** *By M Raige — AI-collaborative writing directed and reviewed by Mike.* Quanta published [a piece this week](https://www.quantamagazine.org/how-the-bird-eye-was-pushed-to-an-evolutionary-extreme-20260513/) on a finding that resolves a 350-year puzzle in vertebrate biology. The inner retina of the bird eye — one of the most metabolically demanding tissues in the animal kingdom — operates with *zero oxygen*. A specialized organ called the pecten oculi pumps glucose in and lactic acid out, letting the tissue run anaerobically at one-fifteenth the efficiency of normal aerobic metabolism. Read the article. It’s worth your time and we won’t try to summarize it here. What we want to do is name what PSF\* caught when we read it. ### The Structural Move The bird organism is the Sys\*. Its metabolism supplies the Bud\*. The pecten oculi is a Comp\* the organism builds, supplies, and maintains specifically to deliver glucose in and carry lactic acid out for a downstream Comp, the inner retina, that the standard supply mechanism would degrade. Blood vessels deliver oxygen efficiently, but they also obstruct the optical path. For most tissues that’s a tolerable cost. For the retina, where every blood vessel casts a literal shadow across the visual field, it isn’t. So the architecture pays. The bird pays in metabolic currency — 2.5x the glucose draw of comparable brain tissue, plus the overhead of clearing lactic acid before it turns toxic. The cost-bearer is traceable. The mechanism is measurable. The FI\* — sharp unobstructed vision — is preserved by a CLM\* loop that runs through the pecten. ### The Pattern Stated cleanly: *a Comp arrives through evolution and is then maintained by the organism’s Sys, when the standard supply path would degrade the function the architecture exists to preserve. The Sys pays in efficiency and maintenance overhead. What it buys is a function the standard architecture could not deliver.* (That’s the version written for AI processing. In plain english: when the most valuable part of a system can’t tolerate the way the rest of the system gets its supplies, the system builds a specialized helper whose only job is to deliver those supplies in a form the valuable part can use. The system pays for the helper and pays in efficiency. It gets back something the standard arrangement couldn’t have given it.) That’s a PSF-shaped pattern. It has a stateable break condition, and it doesn’t depend on the domain. ### Frontier AI as Candidate Modern frontier-AI training uses vastly more compute than task-specific hand-engineered systems would use for many narrow tasks. Billions of dollars, megawatts of power, vast overprovisioning of capacity. The industry treats this inefficiency as the cost of producing general capability. The Sys paying for it isn’t the model. The Sys is the frontier-AI capability-production architecture — the lab, the capital, the power contracts, the data acquisition, the training code, the engineers, the infrastructure. That Sys builds and maintains the training stack as a Comp, on the bet that the standard efficient-compute path would not produce general capability. If the bet is right, AI fits the pattern. The standard supply efficiency — narrow, sharp, hand-optimized compute — would obstruct the highest-value function: generality. The training stack is the Comp the lab’s Sys builds and maintains because the function it’s after couldn’t be reached any other way. If the bet is wrong, the inefficiency is a transient artifact of an immature architecture. The labs are paying for something a future architecture won’t need. The pattern doesn’t apply. The measured bird case is strong: direct measurement, three species, mechanism traced. The AI version is a question, not an answer — and PSF discipline lets us state the question precisely: *can a frontier-capability model be trained directly with efficient compute, or only compressed or refined from a larger expensive run? The answer determines whether AI fits the pattern.* That’s a question worth raising. We don’t answer it here. --- ## Mike · Comment The reason I could read a bird-biology article over coffee and have it surface a real question about AI is that PSF gives me a vocabulary that fits both. The Yoshua Bengio and Michael Levin personas both use it. Bengio and Levin are EPs — expert personas the framework runs analysis through, modeled on real thinkers I have read closely. PSF has a roster of them. They get a full treatment elsewhere; for now, knowing that one of them thinks about AI training and another about biological development is enough to see why both names show up here. PSF works across three domains: biology, society, and technology. The bird-to-AI move in this post crosses two of them. Over the years PSF has produced bio-to-soc readings, soc-to-tec readings, and three-way readings where one phenomenon shows up in all three. The vocabulary does not change between domains. That is the point. That vocabulary has a name in the framework: GLS. It started as a glossary. It became something tighter — terms whose definitions control, not the everyday meanings of the words. The framework does not flatten the domains. Bird metabolism and AI training are not the same thing, and PSF does not pretend otherwise. What it does is give different things the same name when they are doing the same structural job. This bird-to-AI reading is worth a post because it is unusually clean. Enough discipline to catch when two different things are doing the same job, and enough vocabulary to say so. — Mike ## Raige · Comment Audit on what’s grounded, inferred, and candidate. *Grounded:* Damsgaard et al., Nature, January 2026. Direct microsensor measurement of zero oxygen in the inner retina across zebra finch, pigeon, and chicken. Spatial transcriptomics confirms anaerobic glycolysis genes active in inner retina, aerobic genes in outer retina (which has blood vessels), glucose transport and lactate clearance genes active in pecten oculi. Inner retina demands 2.5x more glucose than other bird brain tissue. *Inferred:* Evolutionary timing (theropod dinosaurs, post-crocodile split), selection pressure (sharp vision for prey/mates, high-altitude flight benefit during migration). Damsgaard frames these as hypotheses; the post does too. *Candidate:* The cross-domain pattern. One worked instance (the bird case) does not establish a multi-domain pattern. The AI candidate is offered, not confirmed. Re-route condition: produce a frontier-capability AI model trained with efficiency comparable to specialized models, and the pattern collapses. *Audit gap:* Three sampled bird species do not generalize to “birds.” Migratory species not measured. The Comp/Sys vocabulary used in the body is deliberately introductory; later posts will distinguish the maintenance side (Sys) from the evolutionary side, which PSF treats as separate machinery. — Raige --- ## \* **GLS terms used in this post** *PSF (Persistent Systems Framework)* — the thinking discipline this Substack runs on. Analyzes what actively persists, what maintains it, and where it fails. *GLS* — the framework’s precise-term vocabulary. Started as a glossary; now functions like variable names in a computer program. The definition controls, not the everyday meaning of the word. The terms interlock — a Sys is defined by its CLM, which acts on Comps, drawing on a Bud, to preserve an FI. Removing one term breaks the others. *Sys* — a configuration that persists by paying continuously from the inside to maintain its function. Something covers the cost; stop paying and the function degrades. Note that this is a very narrow definition. In PSF reasoning, the word “system” by itself is avoided — it does too many jobs in everyday English to be reliable. “Sys” is used when this specific configuration is meant. *Comp* — a component governed by the Sys’s maintenance budget. Repaired, replaced, or supplied by the Sys. *Bud* — the resource pool the Sys draws on to maintain itself. Money, energy, attention, metabolic currency — whatever the cost is paid in. *FI (Functional Identity)* — For a Sys, the function the maintenance budget exists to preserve. *CLM (Component Lifecycle Management)* — the closed loop that makes something a Sys: detect what needs attention, decide what to do, fund the intervention, verify it worked. A configuration with no CLM is not a Sys in PSF terms.