Let’s Get Chemical:
Revisiting the Geoteleomic Origin-of-Life Story with Alkaline Seeps
Path-ensemble guidance to find the “shortest path from non-life to proto-life” still works in a Lane-style seep model — and the local-memory story gets more interesting.
Not very long ago I posted about a framework I’ve been calling geoteleomic origin-of-life modeling. The idea was: instead of asking “what initial conditions produce life?”, take a cue from the Anthropic Principle and ask “given that life did emerge, what path ensembles are least surprising conditional on reaching a life-like endpoint?” Mathematically this is a Schrödinger Bridge (SB) problem on the space of stochastic chemical trajectories, with a Doob h-transform doing the work of identifying the minimum-information deformation of the native chemistry that lands you in the target set.
That earlier post described the framework on minimal toy models: schematic RAF-front simulations, serial transfer, schematic compartmental division. The simulations supported the basic claim — small, coherent, low-information SB-style nudges can convert rare ignition into typical ignition — but the chemical realism was, frankly, near zero.
A new paper I’ve recently uploaded takes the next step, as roughly suggested by my old friend Rafal Smigrodzki in a comment on that previous post. We ask: does the SB approach still help when the native dynamics is shaped by something that at least gestures toward Nick Lane’s alkaline hydrothermal seep picture, with H2/CO2 feeds, pH and redox gates, mineral surfaces, porous retention, and chemical clutter?
Spoiler: yes. And along the way the local-memory story got more interesting.
What’s different this time
The new model is still a fairly coarse preliminary study — a path-ensemble diagnostic, not an actual calibrated geochemical reactor — but it adds several features the earlier RAF model lacked:
H2-rich vent-side feed and CO2-rich ocean-side feed, with gradients across a mixing zone.
pH and redox gates concentrated in that mixing zone, modulating reaction propensities.
FeS-like mineral surface factors and pore retention.
A coarse carbon/energy ladder: H2 + CO2 → C1, C1 + H2 → C2, C2 + proton-motive gradient → E.
An explicit membrane-production channel (C2 + E + productive RAF → L → M).
Decoy autocatalytic loops that compete for the same resources but contribute little or negatively to membrane output.
Side and waste channels (TAR, X, Y, Z, W) representing chemical clutter.
Two productive RAF families: an incumbent T0 and a later-arriving improved variant T1.
Nonstationary evolutionary perturbations (a takeover protocol and a shock protocol).
This is closer to a chemically suggestive environment than the earlier RAF-front toys. It is still abstract — we are not modeling actual Wood–Ljungdahl intermediates, real pH-dependent speciation, or explicit amphiphile geometry. But we now have gradients, gates, surfaces, retention, clutter, decoys, evolving variants, and a competition between membrane conversion and catalytic overcycling.
Result 1: SB guidance survives the move to a richer setting
The native dynamics in this regime is weak. Without bridge guidance, neither protocol reaches viable membrane-producing autocatalytic organization; success rates are essentially zero across 20 replicates.
Add a functional bridge — one trained to favor any productive RAF plus membrane mass — and success goes to 1.00 with substantial membrane output. Switch to a conversion-oriented bridge, which specifically favors membrane-conversion efficiency, and the system produces significantly more membrane mass and a much larger post-event membrane AUC.
Numbers from the shock protocol (means over 20 replicates):
Native: success 0.00, membrane mass 5.33, post-shock M AUC 51.2.
Functional bridge: success 1.00, membrane 35.9, AUC 490.4.
Conversion bridge: success 1.00, membrane 64.1, AUC 878.2.
So the headline finding is that SB / Doob-transform path-space guidance is not a trick that worked only because the earlier RAF model was minimal. Even with chemical clutter, decoys, and an evolving productive family, a low-dimensional learned backward potential robustly improves the path ensemble leading from seep-driven carbon and energy flow to energized autocatalytic compartmental organization.
That alone would have been interesting. But the even funkier story turned out to be about local memory. This turned out to teach me some lessons of much broader import, both philosophically and potentially for AI approaches like algorithmic chemistry.
Result 2: broad biphasic memory was too crude
In the earlier work I had played with biphasic memory rules of the form a_eff = a_base · exp(η · h − ξ · Φ(h)), where h is a recent-use trace, η rewards repetition, and ξΦ kicks in once repetition exceeds a threshold. The intuition was: repeated use first stabilizes a pathway, then risks over-conditioning it. This is adjacent to Sheldrake’s much wilder “morphic resonance” framing, but here it just means a path-dependent bias on propensities, with no mystical commitments.
In the richer Lane-style model this rule did not work as hoped. It sometimes increased mutant participation, but it also damaged membrane conversion. Why?
Because the rule penalizes repetition itself — and repetition is how a chemical individual forms. Without recurrence, there is no self-maintaining organization at all. Suppressing repetition broadly is a bit like trying to fix overfitting by deleting your training data.
The fix turns out to be conceptual rather than parametric. What should decline is not function but excess canalization: the dominance of an implementation beyond its current contribution to viable future organization.
I’m calling the refined rule “protected release,” or value-relative anti-precedence. The structure is straightforward:
For each pathway family g, compare its actual share S_g(t) of the recent ecology to a value-justified share S*_g(t), computed by a softmax over per-family value scores. Value here is approximated by recent membrane-conversion payoff per unit of resource or catalytic cost.
Define excess canalization as O_g = max(S_g − S*_g − θ, 0). Penalize this, not raw use.
Give each reaction a functional-protection coefficient c_r ∈ [0,1] — high for shared membrane-conversion machinery, lower for identity-specific autocatalytic cycling — so anti-precedence falls mainly on identity, not on shared function.
Diffuse the “release pressure” from over-canalized families into neighboring pathway implementations via a motif-neighborhood kernel.
Put more simply: penalize a pathway only when it occupies more of the ecology than its function justifies. Protect the shared machinery that constitutes individuation. Release pressure toward neighboring implementations, not into the void.
Crucially, this is not species-specific. The rule does not say “boost the new RAF” or “kill the old one.” It says: any pathway identity whose realized dominance exceeds its value-justified share is gently relaxed, while shared functional machinery is preserved.
Putting it together: shock, takeover, and the winner
We tested two nonstationary protocols:
Takeover: T0 establishes, then T1 is introduced at low abundance. Does the system shift toward the better variant while staying viable?
Shock: T0 establishes, T1 appears, and then incumbent capability is degraded. Does the system recover by adopting T1 while preserving membrane production?
The full condition list runs from a no-bridge native baseline through bridge-only conditions (functional, conversion, plastic), broad biphasic memory, an overuse-only anti-precedence variant, and combinations of each bridge with protected release.
In the shock protocol, conversion bridge plus protected release wins by essentially every metric we tracked, relative to a conversion bridge alone:
Mutant takeover rate up by ~0.35 (95% CI: 0.05 to 0.60).
Mutant share up by 0.072 (CI: 0.029 to 0.118).
Incumbent lock AUC down by 2.46 (CI: −4.05 to −0.97).
Membrane / product conversion efficiency up by 0.011 (CI: 0.001 to 0.020).
Adaptive score up by 0.034 (CI: 0.004 to 0.061).
Multi-objective evolutionary regret: lowest of any tested condition.
The takeover protocol gives a weaker but consistent advantage. With no explicit incumbent degradation forcing change, plasticity-oriented conditions don’t get to shine as brightly. Even so, conversion plus protected release still posts the lowest multi-objective regret and the best adaptive score.
The weight-sensitivity grid is the result I find most reassuring. Across every tested combination of weights on membrane output, mutant adoption, lock-in, and decoy burden, conversion plus protected release wins in all shock cases and all takeover cases. So the headline isn’t an artifact of how we weighted four objectives.
The intuitive picture: conversion plus protected release is not simply more plastic. It preserves or improves conversion while allowing self-transformation. On a Pareto plot of post-shock membrane performance versus adaptive plasticity, it occupies a sweet spot that broad biphasic memory and pure plastic bridges miss. Plastic bridges trade output for flexibility. Conversion bridges alone tend to entrench the incumbent. Protected release lets you keep the throughput while loosening the lock.
Why this matters conceptually
The concept regarding “tendency to take habits” and “habit saturation” is now roughly:
Precedence enables individuation. Anti-precedence, properly defined as value-relative anti-precedence, enables ongoing self-transformation.
A protocell-like system needs both. It must stabilize a self-maintaining organization (precedence). And it must let better-performing implementations replace older ones without losing the macro-organization (protected release).
In information-theoretic terms, the rough form of value you want is something like
V_g ≈ I(g; viable future function) − α · I(g; frozen implementation).
A motif gets credit for predicting future viable function and a penalty for entrenching a particular frozen implementation. The crude simulations I’ve been running use a much simpler proxy — recent membrane-conversion payoff — but the conceptual direction is clear.
The next step is real chemistry
So the work I’ve done so far is just a path-ensemble diagnostic, not a full-fledged chemical simulation. The chemistry is still abstract: H2, CO2, C1, C2, E, named RAF families, generic membrane mass M. To turn this into a real chemical simulation, some significant upgrades are needed:
Replace the C1/C2/E ladder with a real reduced network. Candidate ingredients are well known: CO2/HCO3 chemistry with proper pH-dependent speciation; H2 oxidation and explicit redox coupling; FeS/NiS/greigite-like surface states; formate, CO, methyl, acetyl-like, acetate, thioester, and acetyl-phosphate-like intermediates; and adsorption/desorption dynamics on mineral surfaces. The Wood–Ljungdahl literature, the Lane group’s reactor work, and the iron-sulfide catalysis literature provide most of what’s needed.
Make the spatial / hydrodynamic part less coarse. Real alkaline seeps have explicit pore geometry, flow, retention, surface-area distributions, and pH gradients that are not separable from the chemistry. A reduced reaction-diffusion-flow simulation with realistic pore-scale features would let memory variables emerge from the physics rather than be put in by hand.
Discover motifs instead of naming them. The current model bakes in T0, T1, and the decoy. This is an initial attempt to understand the situation conceptually and – to an extend – quantitatively. What we really need to do is generate random reaction networks and uses exact or approximate RAF detection (Hordijk–Steel and successors) to identify motif families dynamically. Pathway families g would then be discovered, not declared. Anti-precedence is studied as a general rule over discovered motif occupancy, value, and motif-neighborhood structure..
Derive memory variables from chemistry. In the current paper U_g (recent use), V_g (value), c_r (functional protection), and O_g (excess canalization) are all phenomenological. In a chemically grounded version, U_g should correspond to adsorbed pathway-specific intermediates; V_g to measured conversion into membrane / energy-coupled products; c_r to shared energy/membrane machinery; O_g to surface occupation beyond productive turnover. Surface conditioning, adsorption, depletion, inhibition, clogging, washout, pH exhaustion — the microphysics is already path-dependent. The question is whether real surface and flow microphysics naturally produces something like value-relative anti-precedence, or only broad decay (which the current paper shows is too crude).
Experimental analogues. A microfluidic or mineral-reactor experiment looking for adaptive replacement — an incumbent product network persists, a better-performing variant takes over after a perturbation, while membrane / compartmental function remains continuous — would be the cleanest experimental signature of protected release in actual chemistry. The relevant knobs (pH cycling, redox cycling, periodic flow changes, substrate pulses, surface renewal, controlled poisoning/recovery of catalytic surfaces) are not exotic.
Better bridge training. The current paper uses frozen coarse Doob/SB-style bridges trained on near-native rollouts. A more serious version would use online amortized bridge inference, with the backward potential ψ updated as the chemistry runs, and would test whether bridge KL cost itself is reduced in Lane-style settings relative to undirected baselines.
Connect to algorithmic chemistry and AGI. More speculative, but the same formalism applies to rewrite-rule systems and cognitive routines. Precedence lowers the cost of repeatedly useful rule neighborhoods. Anti-precedence penalizes rule families whose dominance exceeds their current contribution to viable future function. Functional protection preserves useful skills while letting implementations change. I think this gives a principled handle on the individuation-versus-self-transcendence balance any open-ended cognitive system has to negotiate. But before leaning hard on that interpretation, I’d want the chemical version to clear its own bar.
Summing up
Two main claims:
1. SB / geoteleomic guidance is not a quirk of the schematic RAF setting. It still robustly improves the path ensemble in a Lane-style cluttered, gradient-driven, nonstationary chemical model.
2. Local memory should not be modeled as broad reinforcement-then-decay of pathway use. Repetition is how individuation works. What you want to penalize is excess canalization — dominance beyond value-justified share — while protecting shared functional machinery. Protected release improves adaptive robustness across every weighting we tested.
Together these point to a refined picture of life-like emergence. Structured energy environments (Lane-style gradients, mineral surfaces, pores, flow) bias chemistry toward viable paths. Autocatalytic recurrence stabilizes local organization. And protected-release-style dynamics prevent any one early implementation from monopolizing the future before a better neighboring implementation has a chance to take over — while the macro-function (membrane / compartmental coherence) stays continuous through the substitution.
This is not yet a chemical claim about specific reactions or specific reactor geometries. The paper is explicit that the next step is real chemistry, real motif discovery, and ideally real reactor experiments designed to look for adaptive replacement signatures.
That’s where the geoteleomic program for understanding the origin of life is now: still abstract, but no longer schematic. Real chemistry simulations would be the next step – but what we have now is some fairly clear guidance about what kind of simulations to run and what kind of results to hopefully expect….
Read the full paper here: