Discussion Guide — Chapter 35

Six prompts, in order. Each is followed by what to listen for — including the answers that sound good and are not.

Prompt 1 — "A lizard solved a problem that a decade of medicinal chemistry was also solving. What, if anything, does that tell us about how research money should be allocated?"

Open with this. It is deliberately a trap and students enjoy walking into it.

Listen for: the move from a single vivid case to a general allocation principle. Almost everyone does it in the first two minutes, and the productive intervention is not to correct them but to ask what the rest of the exendin-4 story says. Reward the student who brings up that exenatide was displaced within a decade by engineered molecules — that is the whole conversation arriving on its own.

Watch for the base-rate question, which is the sophisticated version: how many venoms were screened for every exendin-4? Nobody in the room will know, and that is the right place to end up — the story is memorable precisely because it worked, and we do not hear about the searches that did not. If a student raises survivorship bias unprompted, stop and make the room notice it.

Bad answer that sounds good: "It shows we should have a balanced portfolio." True, unfalsifiable, and it ends thought. Push back by asking what evidence would change the balance.

Prompt 2 — "Venom is selected to incapacitate. Drugs are meant to help. Why is venom nonetheless a productive source of drug leads — and where does the analogy break down?"

Listen for: the overlap being located precisely. The shared specification is fast, potent, specific, acting on vertebrate nervous / cardiovascular / muscular systems. The break is intent — nothing in a venom's evolutionary history selected for a therapeutic window, a convenient route of administration, or a duration matched to human dosing habits.

Reward the student who notices that the stability bonus (§35.2) is real but partial: venom hands you protease resistance and nothing else on Chapter 4's list.

Bad answer that sounds good: "Because venoms are the products of millions of years of optimization." True and empty until they say optimization for what. Ask.

Prompt 3 — "Display technologies do not design anything. Defend or attack that claim."

This one reliably produces genuine disagreement, which is why it is here.

Listen for: the attack, which is legitimate — the researcher designs the library, chooses the target, sets the wash stringency, and may build in constrained scaffolds or macrocycles. That is real chemical judgment.

Then listen for the defense holding the line anyway: designing the search space is not designing the hit. Nothing in the procedure models the target or reasons about which side chain points where. The output is a sequence that binds, with no explanation attached.

The best outcome is a room that concludes the distinction is real but the boundary is fuzzy, and that the fuzziness is itself worth knowing about — because "AI-designed" in a press release often means something much closer to selection than the phrase suggests. That observation feeds directly into case study 35.2 and into the ⚠️ rating's population-definition problem.

Prompt 4 — "State the strongest possible case FOR the claim that AI has revolutionized drug discovery. Then state what would have to be true for that claim to be settled."

Order matters. Make them build the case before they take it apart.

Listen for: the strong case being genuinely strong — CASP14 as a blind assessment, the scale of public structure release, the collapse of a fifty-year problem, broad adoption across a skeptical field, more shots on goal, marginal targets becoming viable. If the strong case is weak, stop and rebuild it; a room that cannot state the best version of a claim cannot evaluate it.

Then listen for the settlement condition: a prospectively defined cohort of AI-derived compounds, followed to approval, compared against a matched historical base rate, over ten to fifteen years. Reward the student who notices that the duration of the settling study is itself the chapter's argument — you cannot evaluate a claim about approvals faster than drugs get approved.

Bad answer that sounds good: "We'll know when the drugs arrive." True, and it skips the whole design question — matched to what? counted how? defined by whom?

Prompt 5 — "Apply the bottleneck rule to something outside medicine. Find one case where improving a stage DID transform the system, and one where it did not."

The most portable thing in the chapter, and the prompt that most reliably reveals whether it landed.

Listen for: specificity. Vague answers ("computers made everything faster") mean the idea has not landed. Good answers name the stage and the constraint separately.

Insist on the positive case. Students who can only produce the failure mode have converted a diagnostic tool into a general skepticism about technology, which is precisely the failure the chapter's "be fair to the technology" section exists to prevent. Do not let the room leave without a case where the improved stage was the constraint.

Then bring it back: which of their two cases does drug discovery most resemble, and — the harder question — is the answer the same for every disease, or does it depend on whether the bottleneck is a long outcomes trial, a lack of validated targets, or the economics of the indication?

Prompt 6 — "Origin carries no evidentiary weight. Then why does this chapter spend most of its length on origins?"

Close with this. It is the chapter's own tension and students should be made to sit in it.

Listen for: the distinction between interesting and evidentiary. Origins explain how the field got here, why certain molecules exist, and where to look next. None of that is evidence about whether a given compound works in a given population.

Reward the student who connects this to Chapter 31's veterinary argument — a fact about a compound's history offered in place of a fact about its effects — and to Chapter 1's naturalness argument. Three chapters, one error, three costumes.

Push further if the room is strong: if origin is uninformative, is the chapter's ✅ for venom-derived peptides as a productive source of leads inconsistent? It is not, and articulating why is the sharpest thing anyone will say all session. Venom is a good place to look is a claim about search strategy, supported by an approval record. This compound came from venom, therefore it works is a claim about a specific molecule, supported by nothing. The rating is scoped to the first and explicitly excludes the second. A student who can state that distinction cleanly has understood both this chapter and the rating system it belongs to.