Module 3 · suggested week 3
Life is a collection of interacting, evolving systems.
What you will be able to do
Understand genes, regulation, selection, feedback and neural signalling well enough to read biology-inspired AI claims and spot where the analogy to computation holds and where it breaks.
Recognition
DNA is a sequence over a four-letter alphabet (A, C, G, T). It is copied into RNA (with U instead of T), and the ribosome reads RNA three letters at a time. Each three-letter codon maps to one of 20 amino acids or a stop signal. With possible codons and only 21 outputs, the map is many-to-one, which makes it robust to some single-letter mutations.
const CODE = { AUG: 'Met', UUU: 'Phe', UUC: 'Phe', GGA: 'Gly', UAA: 'STOP', /* …64 entries */ };
function translate(rna) {
const protein = [];
for (let i = 0; i + 3 <= rna.length; i += 3) {
const aa = CODE[rna.slice(i, i + 3)];
if (aa === 'STOP') break;
protein.push(aa);
}
return protein;
}The genetic code is nearly universal across life, with a small number of known variants (for example in mitochondria). The lookup-table description is accurate. It is a fact about molecular biology, not a metaphor.
Almost every cell in your body carries the same DNA. A neuron and a skin cell differ because different genes are switched on. Proteins called transcription factors bind to DNA near genes and raise or lower their expression. Transcription factors are themselves made from genes, so genes regulate genes: a gene regulatory network.
A useful idealisation treats each gene as a node and each regulatory interaction as a directed, signed edge (activates or represses). Recurring wiring patterns, called network motifs (e.g. feed-forward loops), appear more often than chance would predict and have characteristic behaviours, such as filtering out brief noisy signals.
Boolean network models (each gene ON or OFF, updated by logical rules) capture some qualitative behaviour of real regulatory systems, such as stable cell types as attractors. Real regulation is graded, noisy and chemical. These models are deliberate simplifications, useful for intuition and hypothesis generation.
Natural selection needs just three ingredients:
Given those, the composition of a population changes over generations toward variants that reproduce better in that environment. Nothing aims. There is no goal, no designer and no gradient computed anywhere. There is only sorting by consequence.
Evolution by natural selection is among the best-supported theories in science, confirmed by genetics, the fossil record, comparative anatomy, direct observation in fast-reproducing organisms, and molecular phylogenetics. Debate among biologists is about how much different processes (selection, drift, constraints, gene flow) contribute, not about whether evolution happens.
Fitness landscape is a metaphor (Sewall Wright, 1932): imagine each possible genotype as a location and its reproductive success as altitude. Selection tends to move populations uphill. The metaphor is useful but has known problems. Real genotype spaces are astronomically high-dimensional, landscapes shift as environments and other species change, and populations are clouds, not points.
One landscape, two processes. The gradient follower computes the local slope and steps along it; it can stall on a local peak. The population never computes a slope: it mutates randomly, and selection keeps what reproduces. Watch how differently they explore.
Biological evolution vs gradient-based training
| Natural selection | Gradient descent | ||
|---|---|---|---|
| What changes | Allele frequencies in a population | One parameter vector θ | |
| Direction | None computed. Random variation, then filtering by reproduction. | Computed exactly: the negative gradient of a loss | |
| Objective | None exists. "Fitness" is a description of outcomes, and it changes with the environment. | An explicit loss function chosen by engineers | |
| Information used | Only success or failure of whole organisms | Per-parameter sensitivity (∂L/∂θ_i) for every parameter | |
| Units of change | Generations, populations, often millions of years | Steps, minutes to months of compute | |
| Closest computational cousin | Evolutionary algorithms / evolution strategies | Itself; its relatives are other gradient methods |
"Training is evolution in fast-forward" is an analogy. Both are iterative optimisation-like processes that accumulate improvements. They differ in what varies, how direction is found, whether an objective exists and what information flows. The analogy can suggest questions; it cannot carry an argument.
Evolutionary algorithms (genetic algorithms, evolution strategies) are a genuine computational analogue: populations of candidate solutions, random mutation, selection by score. They are established optimisation methods. Note the direction of inspiration: biology → algorithm. It does not follow that biology "is" running that algorithm, or that gradient descent is evolution.
Some theoretical work shows formal connections. Under particular assumptions (small mutations, large populations), the change in mean trait under selection is proportional to a fitness gradient, which is the content of results in quantitative genetics such as the Lande equation. That is a Level 2 link: a shared mathematical structure under stated assumptions. It is not a general claim that evolution performs gradient descent. Note how the honest version is narrower, conditional and cites its assumptions.
Homeostasis is the maintenance of internal conditions (body temperature, blood glucose, pH) within narrow ranges despite external change. The workhorse is negative feedback: sense a deviation from a set point, respond to reduce it. Your thermostat does this. So does insulin: blood glucose rises, insulin is released, cells take up glucose, levels fall.
Positive feedback amplifies deviations instead: blood clotting cascades, the hormone surge before ovulation, an action potential firing. It is used for decisive, all-or-nothing transitions, and it must be switched off by something else.
A simple negative-feedback controller. Raise the gain and correction becomes faster, until it overshoots. Add delay and even moderate gain starts to oscillate. Delay plus strong correction is the classic recipe for instability.
Delayed negative feedback can produce oscillation. This is a mathematical property of feedback systems (control theory), seen in engineered thermostats, physiological rhythms, predator–prey cycles and supply chains. Here the shared mathematics is the explanation, which makes this a genuine Level 2 connection that often supports Level 3 predictions in specific systems.
A biological neuron receives chemical signals at thousands of synapses, integrates them across a branching dendritic tree, and, if the membrane voltage crosses a threshold, fires an action potential, an all-or-nothing electrical spike that travels down its axon. Information is carried in spike timing and rates, modulated by dozens of neurotransmitters and neuromodulators, glial cells, and synapses that change strength with experience over milliseconds to years.
An artificial "neuron" computes a weighted sum and applies a fixed function:
Multiply each input by a weight, add them up (a Σ loop), add a bias, pass the result through a nonlinearity φ. That is the whole unit.
const neuron = (x, w, b) => relu(x.reduce((s, xi, i) => s + w[i] * xi, b));
const relu = (z) => Math.max(0, z);Does not say: Inspired by a 1940s abstraction of neurons (McCulloch & Pitts). It was never intended as, and is not, a model of how real neurons compute.
Biological brain vs artificial neural network
| Brain | Artificial network | ||
|---|---|---|---|
| Signal | Spikes in time; chemical + electrical | Real numbers passed in one synchronous step | |
| Unit | A neuron is itself a complex computer (dendritic computation) | A dot product plus a nonlinearity | |
| Learning | Local synaptic plasticity, neuromodulation; mechanisms still being worked out | Backpropagation of a global error signal, which brains are not known to implement directly | |
| Energy | Roughly 20 W for the whole brain | Training runs at megawatt scale | |
| Architecture | Evolved; massively recurrent; embodied | Designed; usually feed-forward within a step | |
| Relationship | Historical inspiration | A distinct engineered mathematical object |
There are measured representational similarities. For example, activations in some deep vision networks predict responses in primate visual cortex better than earlier models did. That is an interesting empirical correspondence at the level of representations. It does not show that brains and networks compute the same way, and it is debated how far it extends.
Emergence in biology: properties of the whole not present in the parts. One ant follows simple chemical cues; a colony builds ventilated nests and finds shortest paths to food. One neuron does not remember your grandmother; networks of neurons do. Emergence is a description of the phenomenon. Explaining it means finding the mechanism (the local rules and interactions) that produces it. Module 5 goes deeper.
Metabolism is the network of chemical reactions that converts food into energy and building blocks. Metabolic rate scales with body size in a regular way across species:
Basal metabolic rate B scales roughly as body mass to the 3/4 power. Per gram, bigger animals burn energy more slowly.
Does not say: An empirical regularity. The exponent varies by taxon, temperature and dataset; whether it is "really" 3/4, 2/3 or not a single number is an active debate. B₀ differs between groups.
West, Brown and Enquist (1997) proposed that 3/4 scaling arises from the geometry of space-filling, energy-minimising branching networks (blood vessels, for example). It is an influential theoretical proposal. Its assumptions and its predictions for specific organisms have been challenged, and alternative explanations exist.
Genetic algorithms and evolution strategies borrow variation + selection from biology to optimise without gradients, useful when gradients are unavailable or misleading.
Why this level: The algorithm implements the same abstract mechanism (variation, heritability, selection) by design. This says nothing about gradient descent, which is a different mechanism.
Artificial neurons were inspired by a simplified 1940s model of biological neurons.
Why this level: Historical inspiration only. The computation, learning rule and signalling all differ. Representational similarities are an empirical question studied separately.
Homeostasis, thermostats, ecosystems and agent loops share the mathematics of feedback, gain and delay.
Why this level: Control theory makes specific predictions (delay + high gain → oscillation) that hold across engineered and biological systems when the model assumptions are met.
Metabolic scaling has been explained using physical constraints on transport networks.
Why this level: A proposed mechanism with physics content; it remains a contested theory rather than an established mechanism.
"Neurons", "evolution", "DNA", "immune system", "organism": biology lends AI writing vividness. Check whether each metaphor names a shared mechanism or just a mood.
Lecture-level depth on DNA, gene regulation and cell signalling for Module 3.
Watch for: Gene regulation lectures: transcription factors as switches whose wiring forms networks.
Verified: HTTP 200 + page content checked, 2026-10-30
The cleanest picture of a network as layers of numbers, weights as knobs, and a layer as a matrix-vector product followed by a nonlinearity.
Watch for: The moment the weights are written as a matrix W times a vector a, plus a bias b. That single line is the equation inside every dense layer you will meet on arXiv.
Verified: YouTube oEmbed title+channel check, 2026-10-30
No prerequisites, and it covers dynamics, chaos, fractals, information, self-organisation, agent-based models and networks, the same toolkit as Module 5.
Watch for: The units on networks and scaling. Notice how often Mitchell separates "this is a model" from "this is how nature works".
Verified: HTTP 200 + page content checked, 2026-10-30
The editorial benchmark for this curriculum: serious science, precise about what is shown versus hoped.
Watch for: Read with the Skeptic’s Toolkit open. Label each claim with the evidence badges used here.
Verified: HTTP 200 + page content checked, 2026-10-30
A General Model for the Origin of Allometric Scaling Laws in Biology
Read as a model-building exemplar: assumptions → network geometry → exponent. Limitation: contested assumptions; empirical exponents vary.
Neural networks and physical systems with emergent collective computational abilities
A landmark physics → neural-computation bridge: memory as energy minimisation. Limitation: an idealised network, not a model of real brain circuits.
There are no right answers to check against. These questions are for thinking on the page, and your answers go to your notebook.
Pick an AI product described with a biology metaphor. Is it Level 1, 2 or 3? What would move it up a level?
Your content pipeline as a regulatory network: which components switch others on or off? Where is the feedback, and is it negative or positive?
What is something "selected for" in your industry without anyone choosing it? What gets copied, and what decides which variants survive?
Write two sentences on why a neural network is not a digital brain, aimed at a smart executive. Then write the strongest objection.
Selection vs gradient descent · same landscape
Gradient descent · one point, uses ∇L
Selection · 40 variants, no gradient
Darker = lower loss (higher fitness). Gradient descent computes the local slope of a differentiable function and steps downhill: θ ← θ − η∇L. From a bad start it slides into whichever basin it is in. Selection computes no gradient. It needs only variation (random mutation), selection (keep the better half) and inheritance, and because a population explores many points at once, mutations sometimes hop into a better basin. Evolution strategies are a related optimisation family in machine learning, but biological evolution has no global loss function, changing environments, and no designer choosing η. Both are search on a landscape, a level-2 structural analogy at most, not a shared mechanism.
Negative feedback · homeostasis with delay
A toy temperature regulator: each step it senses the error from the set point (37 °C) and corrects by gain × error, using a reading that is delay steps old. Low gain returns smoothly. High gain overshoots. Add delay and even moderate gain oscillates, because the system keeps correcting an error that is already gone. Flip to positive feedback and the error runs away. The block diagram (sense → compare → act → sense) recurs in thermostats, glucose–insulin regulation and agentic AI loops (act → observe → correct). That shared structure is a level-2 connection, the mathematics of control theory. The mechanisms (neurons and hormones, relays, LLM calls) differ completely. Numbers are illustrative, not physiological.