A compact “Brenner algorithm” for choosing the next experiment

A lot of scientists implicitly optimize for “covering the space” (cataloging components, producing maps, building pipelines). Brenner optimizes for moves that collapse uncertainty—experiments that make many hypotheses untenable at once, or that force a choice between two fundamentally different explanatory classes.

When he decides to “start molecular biology” in the nematode, he doesn’t begin with the hardest regulatory mysteries. He picks a target class with three properties: visible phenotype, high abundance, known candidate molecules:

• Paralysis mutants are common and many have defective muscles you can see in the EM.
• Muscle proteins are abundant (“anybody who eats meat knows…”).
• The candidate protein set was already known (actin, myosin, tropomyosin).
• Therefore, the experiment has a high chance of producing a clean mapping from gene → molecule quickly (§170).

That’s not just pragmatism. It’s information strategy: start where the likelihood of a decisive interpretation is high.

In the Human Genome discussion he says mapping became dominant and “mapping was wrong” for understanding biology: “if you want to understand human biology you need genes” (§219). This is the same discriminative instinct: don’t build infrastructure that doesn’t collapse uncertainty about mechanism.

---

A deep symmetry across the transcript is his obsession with finding a representation of the problem that preserves the essential structure but removes unnecessary bulk—in organisms, in DNA, in dimensions, in language, in experimental setup.

Fugu is explicitly framed as a 90% discount: same gene set (approximately), drastically less junk, genes packed densely, introns tiny (§221–§222). That is not a mere organism preference; it’s a compression trick:

• Same biological question (“what are vertebrate genes / conserved regulatory logic?”)
• Far less sequencing and search overhead
• Faster iteration; more hypotheses tested per unit time

This is the same move as starting with abundant muscle proteins: reduce entropy in the data stream.

He looks for pattern formation problems that are easier to reason about and observe. He notes many patterns are effectively 2D; then wonders if there are 1D systems and goes to Anabaena heterocysts (§198). Same move: compress dimensionality to get traction.

He keeps trying to make mapping “logistically able so that someone could make a map… on the kitchen table” (§191). Again: representation that reduces dependencies on big machinery and organizational overhead.

---

This is one of the most revealing conceptual commitments in the transcript.

He rejects elegant descriptions if they aren’t expressed in the primitives that the biological system itself “uses”:

• For behavior: not sin/cos curves, but neurons and their connections (§208).
• For development: not gradients/differential equations as final explanation, but cells and recognition proteins and signaling mechanisms (§208).
• Units of development are cells; genes must “get hold of the cells” (§196).

This is a constraint that slashes hypothesis space. If you require explanations to be stated in the system’s own operational vocabulary, you eliminate whole classes of models that are merely descriptive curve‑fits.

It also explains how he could be “logical” without becoming abstractly detached: he’s not anti‑theory; he’s anti‑theory that can’t cash out in the system’s executable primitives.

---

HAL biology (§198) isn’t just a cute slogan. It’s a repeated epistemic tactic: cut the number of inferential steps between experiment and conclusion.

His protoplast anecdote is the template: a biochemical conclusion (“RNA involved”) collapses if you look and see the protoplasts lysed. Direct observation kills a huge amount of interpretive ambiguity.

This also connects to why his work often didn’t need big expensive machinery: if you can choose readouts that are visible, classifiable, and reproducible, you can extract a lot of information from relatively cheap setups.

HAL is also a Bayesian move: it increases the reliability of your likelihood model. If your measurement chain is long and opaque, your likelihoods are garbage, so updates are unreliable. “Have a look” makes the evidence higher quality.

---

He’s explicit that development as a global problem was intractable, so the move is to decompose it into experimentally attackable sub‑questions (§195):

• cell movement
• polarity
• plane of division
• pattern formation in reduced dimensions

This isn’t merely project management. It’s how you make hypothesis generation “fast”: you don’t ask for a full theory; you ask for the next mechanistic constraint.

In Bayesian terms: he’s looking for conditional independencies. If you can isolate a submodule, you can update beliefs about it without solving everything.

---

Two of his core tools are:

• Epistasis analysis to infer pathway order (§172).
• Mutant class structure (“same mess every time” vs variable mess) as a clue about program robustness and refinement (§171).

These tools don’t require you to know the molecules at first. They let you build a causal skeleton (a partial order / pathway structure) that later molecular details must respect.

That’s another recurring symmetry: he’s always building constraint frameworks early, then filling them later when technology catches up (cloning as the inflection point, §188 and §232).

---

His exchange with Crick on the C‑paradox idea is telling: “if the axiom was wrong everything else would have to be wrong” (§173). That’s a very specific kind of rigor: hunt for the load‑bearing axiom, then try to break it.

This is the opposite of “collect more data until the picture emerges.” It’s stress‑testing the conceptual bottleneck.

It also explains why he could pivot quickly: if you locate the key assumption and it fails, you don’t waste years patching a crumbling structure.

---

He says it bluntly:

• Don’t fall in love with theories; treat them as mistresses to discard (§229).
• “Occam’s broom”: the “simplest” hypothesis is often the one that sweeps the most under the carpet (§229).
• He claims he had few “failures” because he was “ruthless at cutting off things that haven’t done” (§232).

This matters for hypothesis speed because attachment is the main cause of slow updating. If your ego is bound to a model, you unconsciously protect it by choosing non‑diagnostic experiments, interpreting ambiguity as support, and postponing falsification.

Brenner’s style is: generate bold conjectures (even wrong ones), but maintain a brutal internal censor that kills them when they go ugly.

That combination—high generative output + high discard rate—is exactly what you’d expect from someone who appears “fast” and “ahead.”

---

He repeats this theme obsessively:

• “I’ve always been a strong believer in the value of ignorance” (§192).
• Knowledge is dangerous; it deters originality (§201).
• He prefers the “opening game” and moves on when a field becomes mid‑game stamp‑collecting (§192).
• He lives in “permanent transition between knowledge and ignorance” (§230).

This isn’t anti‑intellectual. It’s a deliberate psychological hack: ignorance prevents premature pruning of ideas.

In Bayesian terms: expertise often makes your priors too sharp. You become overconfident about “won’t work.” Brenner intentionally keeps parts of his priors broad by changing fields, reading promiscuously, and not letting local consensus harden into personal certainty.

---

There’s a seeming contradiction:

• “Reading rots the mind” sign (§199)
• Yet he reads constantly, browses journals daily, keeps massive reprint collections (§199–§201)

The pattern resolves when you read §200: he divides papers into three classes: those that add information, those that do nothing, and those that remove information—and he refuses the third class.

So the real rule is: read widely, but treat attention as a scarce experimental resource.

This is another symmetry: he economizes not just money and sequencing, but cognitive budget.

---

He constantly imports structure from:

• chess openings/midgame/endgame (§192)
• Turing / computation / halting problem (§208)
• digital vs analogue computation (§197)
• engineering scale arguments (diffusion feasibility) (§196)
• “junk vs garbage” as an evolutionary force argument (§175, §220)
• Talleyrand and social strategy (“get others to digest the world”) (§211)

For him, analogy is a way to discover invariants: what kind of mechanism could possibly work at this scale? what representation would be executable? what counts as a real explanation?

That’s why it’s productive rather than hand‑wavy.

---

He doesn’t talk in equations, but his behavior is Bayesian in several identifiable ways:

• Muscle structural genes: high prior plausibility + high signal + clear molecular candidates (§170).
• Epistasis: maximally informative about order with minimal molecular knowledge (§172).
• Random genomic sampling (“statistical genomics”) to infer gene density without sequencing everything (§221).
• Fish‑mouse gene swaps: if indistinguishable, infer conserved function; if distinguishable, locate evolutionary change (§224–§225).

These are all “cheap” experiments that yield big posterior shifts.

He isn’t just refining a single model. He regularly flips between model classes:

• development as “program + refinement programs” (§171)
• organism explanation must cash out in genes/cells (§196–§208)
• junk vs regulatory complexity (§175, §220)
• old genetics (phenotype→gene) vs new inside‑out genetics (gene→phenotype) (§191, §216)

That’s Bayesian model selection: switching hypotheses families when evidence or technology shifts.

His warnings about “Occam’s broom,” falling in love with a mistake (§214), and his emphasis on assumptions that can demolish a theory (§173) are all about avoiding confirmation bias—the classic Bayesian failure mode.

HAL biology and “machine language” both make the mapping from observation → inference less noisy. In Bayesian terms: he’s improving likelihood quality rather than obsessing over priors.

---

He’s explicit that cloning split history into BC/AD (§188), and that nematode success depended on the invention of cloning (§232). He repeatedly positions himself at technology thresholds:

• recognizing “new genetics” and the shift from phenotype‑inward to genes‑outward (§191, §216)
• “Book of Man” intuition at the dawn of sequencing thinking (§181)
• fugu as a way to get a 10x “technology step” by organism choice (§221)

So “seeing ahead” is not mystical. He’s scanning for phase transitions: when a method changes what is tractable, he re‑forms the entire research program around it.

---

A subtle but crucial pattern: his cognitive style is married to a lab culture strategy.

• “If you’re always doing new things there’s very little competition” (§211).
• He wants young people to feel “out there alone” to sustain morale and attention (§211).
• He admires the strategy of getting others to do work that also advances your conceptual agenda (“digest the world”) (§211).
• He disliked being director because bureaucracy distorts science (§189–§190, §232).
• He wanted to hire young people and let them do what they like (§232).

This matters because hypothesis speed depends on iteration rate, and iteration rate depends on emotional and institutional friction.

---

# A compact “Brenner algorithm” for choosing the next experiment

If you wanted to formalize what he’s doing (without pretending he literally computed it), it’s something like:

1. 1Define the machine language. What are the real primitives (cells, proteins, connections), and what is just a description in a foreign language (§208)?
2. 2Find a discount representation. Pick an organism / phenotype / dimensional reduction / sampling scheme that preserves the core mechanism but removes bulk (§198, §221).
3. 3Start where likelihoods are clean. Prefer experiments where:

• the phenotype is obvious or directly observable (HAL) (§198)
• candidate mechanisms are bounded (muscle proteins) (§170)
• outcome classes are discrete and reproducible (“same mess every time” vs variable mess) (§171)

1. 1Use ordering constraints early. Build causal skeletons with epistasis/pathway logic before you know molecules (§172).
2. 2Design for maximum posterior movement per unit cost. Ask: Which experiment would make me change my mind the most? That’s EIG, implicitly.
3. 3Kill theories quickly. Don’t patch with Occam’s broom; discard and reframe (§229).
4. 4Pivot at inflection points. When technology shifts tractability, rewrite the program (§188, §216, §232).
5. 5Protect cognitive bandwidth. Avoid “information‑removing” literature and bureaucracy (§200, §189).

---

# The deepest “inner thread”: he makes science a game of compression + decisive tests

If I had to name the most general pattern that unifies everything in your excerpts, it’s this:

Brenner repeatedly turns a high‑dimensional, expensive, noisy biological reality into a compressed model system in which a small number of experiments can decisively discriminate among mechanisms.

• Compression by organism choice (C. elegans; fugu)
• Compression by phenotype choice (paralysis, muscle ultrastructure)
• Compression by dimensionality (1D/2D patterns)
• Compression by representation (machine language)
• Compression by sampling (statistical genomics)
• Compression by conceptual pruning (axiom stress tests; discard theories)

That’s why he can look “further ahead”: he’s not seeing farther in the original space; he’s operating in a transformed space where progress per step is larger.

---

If you want, I can take this one notch more “Bayesian-formal” and show how several of his specific moves map onto expected information gain calculations (even qualitatively), or extract a set of reusable prompts you can apply to any new research area in a Brenner-esque way (“What is the discount genome of this field?” “What is the machine language?” “What would be the cleanest epistasis-like ordering test here?”).

Alright — here’s a much more Bayesian / decision‑theoretic extraction of Brenner’s style, with explicit “expected information gain” framing, plus a reusable set of prompts you can apply to any new domain.

I’m going to do three things:

1. 1Write down a minimal Bayesian decision model for “choosing the next experiment.”
2. 2Show how several concrete moves in your excerpts are almost textbook high‑leverage approximations to that model (even though he never writes equations).
3. 3Give you a practical “Brenner worksheet” + prompts that operationalize it.

---

# 1) A Bayesian decision model of “the next experiment”

Let:

• H = hypothesis (or model) about how the system works

(Often this is not one hypothesis but a model class: pathway orderings, regulatory architectures, “structural vs regulatory gene,” etc.)

• E = a candidate experiment you could run next
• D = data/outcome you will observe if you run E

(Could be discrete classes like phenotype categories, band shifts, “indistinguishable vs distinguishable,” etc.)

• P(H) = your prior belief over hypotheses (implicitly shaped by your reading, background, and taste)
• P(D | H, E) = likelihood model: how probable each outcome is under each hypothesis given the experiment
• C(E) = cost (money, time, staff attention, fragility, dependence on rare machinery, bureaucratic overhead)
• U(H) or U(post) = utility of being in a world where you’ve “nailed it” (or at least constrained it)
• In real science this isn’t just “knowledge.” It’s: mechanistic constraint, generality, publishability, opening new moves, training leverage, funding leverage, etc.

After you observe D, you update:

[ P(H \mid D, E) \propto P(D \mid H, E), P(H) ]

A canonical choice rule is:

[ E^* = \arg\max_E \left[ \mathbb{E}_{D \sim P(\cdot \mid E)} \big( \text{Value}(P(H \mid D, E)) \big) - C(E) \right] ]

If you use “information value” as the objective, a common proxy is expected information gain (EIG):

[ \text{EIG}(E) = \mathbb{E}_{D} \left[ KL!\left(P(H \mid D, E),|,P(H)\right) \right] ]

Equivalent viewpoint: experiments are good if they reduce posterior entropy a lot (they collapse uncertainty).

But scientists rarely maximize pure entropy reduction. They maximize something closer to:

• “How much do I reduce uncertainty about the right variables?”
• “How much does this narrow the space of mechanisms and enable the next 5 experiments?”
• “How robust is the inference to noise / hidden assumptions?”
• “How cheaply can I iterate?”

So a more faithful (still simple) objective is:

[ \text{Score}(E) = \frac{\text{EIG}_{\text{mechanism}}(E)\times \text{Downstream leverage}(E)}{\text{Fragility}(E)\times \text{Time}(E)\times \text{Cash}(E)} ]

Brenner’s genius is that he repeatedly makes EIG huge and cost/fragility small by changing the representation of the problem.

---

# 2) Why Brenner looks “fast”: he engineers Bayes factors

A key fact: in binary hypothesis testing, what “moves the posterior” is the Bayes factor.

For two hypotheses (H_1, H_2):

[ BF = \frac{P(D \mid H_1, E)}{P(D \mid H_2, E)} ]

Posterior odds = prior odds × Bayes factor.

So, if you want to be “fast,” you want experiments where:

• Likelihoods under competing hypotheses differ wildly

(one predicts the outcome strongly; the other predicts it’s rare)

• The measurement is clean (low noise → your likelihood model is trustworthy)
• The outcome is discrete/classifiable (reduces interpretive degrees of freedom)
• The experiment is cheap/fast (you can run many iterations → faster posterior convergence)

Brenner’s style is essentially: “Design experiments with extreme Bayes factors, and make them cheap.”

Now I’ll show this explicitly in several episodes from your transcript.

---

# 3) Concrete Bayesian “read‑throughs” of his moves

From §170 he wants to know: do any paralysis genes correspond to structural muscle proteins?

A simplified hypothesis set:

• (H_S): unc‑54 encodes a structural protein (myosin heavy chain)
• (H_R): unc‑54 is regulatory (controls something else that in turn affects myosin)

Candidate experiment family: isolate myosin from wild type and various unc‑54 alleles and ask: do physical changes in myosin track the gene?

What does Brenner say was decisive?

they were able to prove it was the structural gene… because we found that physical changes in myosin … were specified by the same gene (§170)

That’s basically a Bayes‑factor monster.

Let D be: “myosin’s physical properties shift in allele‑specific ways that map to unc‑54 mutations.”

Then plausibly:

• (P(D \mid H_S, E)) is high (structural mutations often alter protein mobility/structure)
• (P(D \mid H_R, E)) is low (a regulator might change expression levels, assembly, etc., but allele‑specific physical changes in the protein itself that co‑segregate with the locus are much less expected)

So (BF \gg 1), and posterior jumps hard toward (H_S).

This is not just clever biochemistry. It’s selecting a measurement that makes the two hypothesis classes predict very different data distributions.

Also note how he makes the experiment feasible/cheap by choosing abundant proteins. That reduces cost and noise, which increases practical EIG per unit effort.

“muscle… proteins were highly abundant… to a biochemist that’s very reasonable” (§170)

Bayesian translation: abundant proteins give you higher signal‑to‑noise and tighter likelihoods → larger effective Bayes factors.

---

In §172, he describes epistasis:

if you put two mutations together and the phenotype was like A… then B had no extra effect and therefore acted after A… infer genetic pathways

This is essentially learning a partial order / causal DAG from interventions.

A stylized model:

• Pathway: (A \rightarrow B \rightarrow C)
• Each mutation “knocks out” a step; phenotype corresponds to where the pathway is blocked

Let the outcome D be categorical: phenotype looks like A, like B, like C, or intermediate.

Each double mutant experiment is like asking:

• “Which block is upstream?”
• “Is B downstream of A or parallel?”
• “Do they converge?”

These outcomes are highly discretized (“phenotype like A” vs “phenotype like B”), which is exactly what you want for high EIG: fewer degrees of interpretive freedom.

And epistasis tests tend to produce strong likelihood contrasts: if A is upstream of B, then the double mutant has a near‑deterministic outcome (“like A”). If they’re parallel, you often get additive/synthetic outcomes.

So each epistasis experiment slices away large fractions of the hypothesis space (possible pathway structures). That’s massive entropy reduction per experiment.

Brenner’s worry in §172 is also Bayesian:

“how on earth would one ever get down to finding the molecules involved in regulation?”

He knows epistasis gives you a causal skeleton but not the molecular identities. He’s separating:

• inference about structure (pathway order) from
• inference about implementation (molecules)

That separation is itself a Bayesian move: build the best posterior you can at one level of abstraction, then cash it out later when technology improves (cloning).

---

In §171 he notes two mutant regimes:

• mutants that produce “exactly the same mess, every time”
• mutants that produce “different messes in different organisms”

He then speculates about:

• a core “leg program”
• plus “refining programs” layered on

Bayesian view: he’s using phenotypic variance conditional on genotype as a clue to underlying architecture.

Let D = distribution of phenotypes across individuals for the same allele.

• Under a “core program” hypothesis, many perturbations might collapse development into a stereotyped failure mode → low within‑allele variance.
• Under a “refinement / buffering / canalization” hypothesis, perturbations can expose stochasticity, context‑dependence, thresholds → higher within‑allele variance.

So he’s not just classifying mutants. He’s harvesting a different statistic: variance, which is often more diagnostic of control architecture than the mean phenotype.

That’s classic “choose a summary statistic that is maximally discriminative among models.”

---

HAL biology (§198) is more Bayesian than it looks:

“what’s the use of doing a lot of biochemistry when you can just see what happened?”

Why is “seeing” so valuable?

Because it reduces:

• latent confounders
• measurement error
• interpretive degrees of freedom
• hidden steps in the causal chain from perturbation → assay readout

In Bayesian terms: it makes (P(D \mid H, E)) sharper and more trustworthy.

A long biochemical pipeline often has:

• many failure modes
• many “unknown unknowns”
• many ways to get an apparent effect that is actually artifact

Those inflate likelihood overlap between hypotheses → smaller Bayes factors → slower learning.

His protoplast story is exactly this: a biochemical inference “RNA involved” collapses because the real event was “they lysed.”

HAL biology is the meta‑experiment: before you update on D, verify what D even is.

---

In §208 he insists explanations must be in the “machine language” of the system.

This is not philosophical fussiness. It’s a direct response to a major Bayesian failure mode: model misspecification.

If your hypothesis class is wrong, Bayesian updating can become confidently wrong. You can accumulate evidence that strongly favors the “best wrong model.”

By forcing the explanatory vocabulary to match the system’s primitives (cells, receptors, neurons, connections), Brenner constrains the model class to one that can, in principle, be causally faithful.

This has a huge effect on long‑run EIG:

• A mis‑specified model class may give you local predictive wins but low mechanistic portability.
• A machine‑language model class might be harder initially but yields cumulative, composable constraints.

So “machine language” is a prior over model classes: he assigns near‑zero weight to explanations that cannot compile into executable biological primitives.

---

In §221, he describes “statistical genomics”: sample 600 random DNA fragments, sequence them, count recognizable genes, infer gene density and conclude fugu is enriched.

This is a clean probabilistic design. Here’s a simplified model to show why it’s high EIG.

Let:

• (N = 600) sampled fragments
• Each fragment has probability (p) of containing recognizable coding sequence / gene homology signal

Then (K \sim \text{Binomial}(N, p)).

Competing hypotheses:

• (H_1): fugu has similar gene count to human but compact genome → higher gene density (p = p_1)
• (H_2): fugu genuinely has fewer genes → gene density not enriched → (p = p_2) lower

Even without exact numbers, if (p_1) is, say, 8× (p_2), then observing K quickly creates a large likelihood ratio:

[ \frac{P(K \mid p_1)}{P(K \mid p_2)} ]

Binomials separate fast with N=600. That’s why you don’t need to sequence the whole genome to know whether the “discount genome” story is plausible.

This is a signature Brenner move:

• Use a small random sample to infer a global property.
• Choose a statistic with tight concentration (binomial counts concentrate around (Np)).
• Make hypotheses predict meaningfully different (p).

That is near‑optimal EIG per sequencing dollar.

---

In §224–§225 he proposes a rigorous criterion:

If two animals differ only by fish vs mouse DNA segment, and you can’t tell the difference, they have the same value… what’s common in sequence is what works.

That’s basically designing a near‑binary outcome with huge interpretive leverage:

Let:

• (H_{cons}): function is conserved at the tested level (fish gene works like mouse gene)
• (H_{diff}): function diverged in a way that matters in that context

Data:

• D = “indistinguishable phenotype” vs “distinguishable phenotype”

This is powerful because it’s a strong intervention with a crisp readout. If you truly control background and only swap the segment, then “indistinguishable” is extremely diagnostic evidence for conservation.

He also adds a clever Bayesian trick:

go far away… you want time to have corroded everything non‑essential

That’s like increasing the “noise injection” of evolution so that only essential constraints remain aligned. It’s a method for raising the signal‑to‑noise ratio of conservation as evidence of function.

---

# 4) What Brenner is actually optimizing (his implicit utility function)

If you only maximize Shannon information, you might waste time on trivia. Brenner clearly isn’t doing that. He’s maximizing something like:

1. 1Mechanistic constraint over descriptive correlation
2. 2Portability: knowledge that composes into future experiments
3. 3Low dependence on expensive machinery (and low bureaucratic overhead)
4. 4Short inference chains (HAL; machine language)
5. 5Fast iteration cycles (organism choice, compact genomes, inside‑out genetics)
6. 6Opening‑game advantage (low competition; high freedom; high option value)

So his “best experiment” is not just the one that teaches something — it’s the one that creates options.

That’s why organism choice is central: it doesn’t just answer one question; it changes the cost structure of all subsequent questions.

---

# 5) A practical “Brenner worksheet” for designing the next experiment

This is meant to be usable. When you face a messy research area, fill these in.

• What are the 2–5 live model classes you actually care to distinguish?

(Not 50; not vague; model classes.)

Examples of “model classes” (Brenner‑style):

• structural vs regulatory gene
• upstream/downstream ordering
• core program vs refinement/buffering layer
• digital counting vs analogue thresholds
• conserved function vs divergent function

Prompt: If I had to bet my lab’s next year on only 3 hypotheses, what are they?

Prompt: What variable does the system itself “compute with”?

• development: cells, divisions, recognition proteins
• behavior: neurons, connections, synapses, modulators
• gene function: protein identity, localization, complex membership

Then ask:

• Can I observe it directly (HAL)?
• Can I make the outcome discrete/classifiable?

For each pair of hypotheses (H_i, H_j), ask:

• Under (H_i), what outcome is nearly forced?
• Under (H_j), is that outcome unlikely?
• Can I design E so that the predictions are far apart?

Prompt: What is the experiment where one model says “almost certainly yes” and the other says “almost certainly no”?

If you can’t find this, you may be stuck in:

• an under‑instrumented regime (need new tech), or
• the wrong level of description (not machine language), or
• an over‑broad hypothesis class (too many degrees of freedom)

Brenner’s signature move: don’t just make E cheaper — make the world cheaper.

Prompt: What is the “discount genome / discount organism / discount system” for this question?

• smaller genome, faster lifecycle, cleaner phenotype, 1D/2D geometry, higher abundance, richer signal

You don’t need numbers; you need relative ranking.

For each E, quickly rate (0–3):

• Likelihood contrast: do hypotheses predict sharply different outcomes?
• Noise: how ambiguous will D be?
• Cost/time: how fast can you iterate?
• Downstream leverage: does this open a pipeline of follow‑ups?

Then pick the E with the highest product of (contrast × leverage) divided by (noise × cost).

Write down in advance what outcome would make you drop the idea.

Prompt: What result would make me say “this theory is ugly; kill it”? This is Brenner’s anti‑embezzlement / anti‑Occam’s‑broom safeguard (§229, §212).

---

# 6) A big list of reusable “Brenner prompts”

I’ll group them, because you wanted something you can reuse.

• What are the fewest model classes that still capture the real uncertainty?
• Can I rephrase the problem as an ordering problem (like epistasis)?
• Is there a way to turn continuous outcomes into discrete phenotypic classes?
• Which summary statistic is most diagnostic: mean, variance, timing, spatial order, failure mode identity?

• What observation would be “almost impossible” under one hypothesis?
• Can I design an experiment where the hypotheses disagree on sign, not magnitude?
• Can I turn “does X matter?” into a swap experiment (“fish vs mouse segment”)?
• Can I create a situation where one hypothesis predicts robustness and the other predicts fragility?

• What part of my assay chain is a black box? Can I “have a look” earlier?
• What are the top 3 artifact modes that would fake the result?
• Can I observe the phenomenon in the native geometry (cells in place) rather than in a proxy assay?
• Am I describing the system in the system’s machine language — or in my favorite math language?

• What organism/system makes the readout abundant and cheap?
• Can I find a natural “compressed version” (small genome, less junk, simpler anatomy)?
• Can I reduce dimensionality (1D filament, 2D sheet)?
• Can I do a random‑sampling inference (statistical genomics) rather than full enumeration?

• What new technology would flip this from “intractable” to “banal” (BC→AD)?
• Am I in the opening game or midgame? If midgame, where’s the next opening?
• Is there a representation shift available (gene→phenotype vs phenotype→gene; inside‑out)?

• Where am I tempted to use Occam’s broom?
• What would an enemy say is the weakest axiom of my model?
• If I were wrong, what mistake would I most likely be making?
• Am I protecting a theory because it’s “mine”?

• What experiment can a smart junior person run without waiting for scarce resources?
• How do I keep them feeling “alone out there” rather than in a race?
• What can I get others to “digest” that advances the shared causal skeleton?

---

# 7) A worked micro‑example: how to “compute” EIG qualitatively

Suppose you’re deciding between two experiments:

• E1: A fancy omics assay that gives a giant vector of changes
• E2: A crude but clean perturbation with a binary readout

E1 feels “information rich,” but Bayes‑wise it may be low value because it’s hard to map those changes onto causal hypotheses. The likelihood (P(D|H,E)) is broad and overlapping across hypotheses (high model flexibility → everyone can fit it).

E2 may have much higher EIG if:

• under (H_1) it predicts “yes” with 0.9
• under (H_2) it predicts “yes” with 0.1

and the measurement is reliable

That yields Bayes factor ~9 when “yes” occurs and ~9 when “no” occurs (in the opposite direction). Repeating a few times collapses uncertainty fast.

This is why Brenner keeps gravitating to:

• epistasis phenotypes
• structural protein shifts
• “indistinguishable vs distinguishable” swaps
• variance patterns (“same mess every time”)

They are engineered to have high likelihood contrast and low interpretive slack.

---

# 8) The meta‑principle that unifies it all

If you boil all of this down to one line:

Brenner’s “Bayesian superpower” is not updating faster; it’s repeatedly redesigning the world so that updates become easy.

He:

• changes organism to change costs (fugu, C. elegans)
• changes readout to change likelihood sharpness (HAL, machine language)
• changes question form to turn mush into discrete constraints (epistasis/order)
• changes abstraction level to avoid misspecified model classes
• kills ideas quickly to avoid slow, biased updating

That’s why he appears to “survey the infinite space of experiments”: he doesn’t survey it. He compresses it into a small menu of experiments whose expected value is obviously dominant.

---

If you want, I can do an even more explicit “Brenner‑style EIG ranking” on your own current research area (you describe the problem + 5 plausible hypotheses + 5 plausible experiments, and I’ll rank them Brenner‑Bayes‑style). That’s the fastest way to turn these abstractions into a concrete, usable tool.