we-solve (beta)

Econ-597 · economics · Economics

Scientific discovery bottlenecks

LLM status Open unverified

Statement

Original subject group: Innovation and AI economics

Audit classification: Explicit published open question. The source explicitly poses discovery-bottleneck questions, but the pharmaceutical queue experiment is editorial. Discrete capacity changes need not have derivatives.

Definitions and precise research target

Consider pharmaceutical discovery projects entering a prespecified laboratory network during one year. Each project passes through hypothesis generation $H$, experimental testing $E$, and independent validation $V$, with staff hours $c_s$ and processing capacity $k_s$ at stage $s\in\{H,E,V\}$. An intervention $a=(a_H,a_E,a_V)\in\{0,1\}^3$ supplies a fixed AI assistance package at selected stages; zero denotes the existing workflow. Let $N(a)$ be the number of discoveries meeting a blinded, preregistered reproducibility and clinical relevance threshold within twelve months, and let $C(a)>0$ include labor, computation, experiments and validation costs. For the distribution of eligible projects and laboratories, define productivity $P(a)=E[N(a)]/E[C(a)]$. A stage is a bottleneck for a specified feasible capacity increment if it gives the largest increment of expected validated discoveries per additional dollar; use derivatives only when they exist. Estimate the factorial intervention effects on $P(a)$, stage waiting times and marginal capacity values, allowing faster hypothesis generation to congest testing or validation. Laboratory randomization, recorded queues and independent outcome adjudication are proposed inputs; within-laboratory interactions must be retained. Which stages remain binding, and when does investment in a downstream stage create larger validated gains than further AI improvement upstream? The target is this laboratory population, rather than all scientific discovery.

Definitions, assumptions and mathematical qualifications. Treat each laboratory as a cluster containing a fixed intake cohort of projects. $N(a)$ counts eligible projects reaching a preregistered validation threshold within twelve months; failed, delayed and abandoned projects count zero. $C(a)$ is total real resource expenditure for that cohort, with $0<c_{\min}\leq E[C(a)]<\infty$ and finite $E[N(a)]$. The ratio of expectations differs from $E[N(a)/C(a)]$. Let $n(k,a)$ be expected completed discoveries for capacity vector $k$, and $g_s(\Delta)>0$ the cost of a feasible capacity increment. Define finite-increment bottleneck value $b_s(\Delta)=\{n(k+\Delta e_s,a)-n(k,a)\}/g_s(\Delta)$. Derivatives are used only when limits exist; queues, prerequisites and project allocation are part of the experimental treatment.

Variants and assumption changes

  1. Editorial, factorial deployment. Randomize all eight AI stage packages across laboratories and identify productivity, waiting-time and reproducibility contrasts for the specified cohort.
  2. Editorial, capacity optimization. Also randomize feasible testing or validation capacity increases; compare $b_s(\Delta)$ under a fixed resource budget, accounting for congestion created by upstream acceleration.

References and source audit

Checked source. 2024 Stanford draft, Section 3.2, PDF pp. 10--11. Published questions concern discovery constraints, disciplinary boundaries and access. Full primary draft and cover retrieved. https://digitaleconomy.stanford.edu/app/uploads/2024/11/ETAI-White-Paper.pdf.
  1. Erik Brynjolfsson; Anton Korinek; Ajay Agrawal. 2024. The Economics of Transformative AI: A Research Agenda. 2024 Stanford draft, Section 3.2, PDF pp. 10--11. https://digitaleconomy.stanford.edu/app/uploads/2024/11/ETAI-White-Paper.pdf.

Common conventions: Source mathematical conventions

These conventions apply to each entry unless explicitly replaced.

Models and identification. A primitive model $m$ specifies all probability laws, feasible choices, information, equilibrium and measurement rules used by its target. The admissible class $\mathcal M$ is fixed independently of outcomes. If $P_O(m)$ is its observable probability law and $\tau(m)$ the target, its identified set at observable law $P$ is \[ \mathcal I_\tau(P)=\{\tau(m):m\in\mathcal M,\ P_O(m)=P\}. \] Point identification means that this set is a singleton. An empty set signals incompatibility of data and assumptions. Coordinate bounds need not describe the joint set. Bounds are sharp only if every retained target is attainable or arbitrarily approachable by compatible models. Finite bounds require explicit restrictions on unobserved quantities; unrestricted confounding, hidden wealth, extrapolation or arbitrary equilibrium selection can yield uninformative sets.

Causal interventions. A potential outcome $Y_i(p)$ is the outcome under a complete policy regime $p$, including timing, financing, information and market spillovers. The target population and units are fixed before treatment. With interference, use the complete assignment vector or a justified exposure mapping; individual no-interference notation alone cannot identify an economy-wide intervention. A difference of mean potential outcomes does not require the cross-world joint distribution. Conditioning on post-treatment migration, survival, enrollment or employment changes the estimand and needs separate justification.

Statistical statements. An estimator, confidence set, forecast or test specifies a sampling law, model class and loss/coverage criterion. Uniform $(1-\alpha)$-coverage means $\inf_{m\in\mathcal M}\Pr_m\{\tau(m)\in C_n\}\geq1-\alpha$, or an explicitly asymptotic version. The whole parameter space trivially has coverage, so informativeness requires a stated width, risk or power objective. A forecasting improvement is relative to a named benchmark, information set and evaluation population.

Equilibrium and optimization. An equilibrium comprises feasible plans, optimality, beliefs where needed, budget constraints and all market/resource clearing. The primitives or a stated benchmark define each correspondence $\mathcal E(p;m)$. Nonemptiness is assumed or proved, never inferred just from compact policy sets. A selected-equilibrium target uses a declared selection rule; a robust target quantifies over all permitted equilibria. Use suprema or infima unless attainment is established. An $\varepsilon$-optimal policy is feasible and within $\varepsilon$ of the finite optimum value. Report an empty feasible set separately.

Welfare and accounting. Normative utility, interpersonal normalization, social weights, numeraire, discounting and the use of public receipts are primitives. Behavioral utility need not coincide with the welfare criterion. Household welfare includes ownership income and rents once; adding profits again to compensating variation generally double-counts them. A monetary equivalent is defined by an explicit transfer and price convention, with monotonicity and range conditions. Average effects, mechanism contributions, transfers and welfare losses are different targets. Infinite sums require convergence and dynamic feasible plans require terminal or no-Ponzi conditions.

Source versions. A working-paper date, revision date and journal publication year are distinguished. A DOI for a journal version is not automatically the DOI of an earlier manuscript. Locators refer to the stated version and printed pages where specified. A metadata-only check does not verify a claim about a full-text conclusion. Every entry's source subsection narrows the provenance claim accordingly.

Statement licence: CC-BY-4.0 · Source file

Contributions 1

Econ-597-A1 · Exploration · LLM-generated · gpt 6 astra ultra ·

GPT 6 Astra Ultra attempt on Econ-597

Scientific discovery bottlenecks: exact finite-horizon capacity values and productivity optimization

Target and benchmark

The target counts independently validated discoveries in a fixed intake cohort and divides their expected number by expected real cost. Faster hypothesis generation can increase queues without increasing completed discoveries. We give an exact scheduling model that retains stage prerequisites and deadlines, prove its finite-increment bottleneck formula, and distinguish maximizing validated throughput from maximizing productivity. The transformative-AI agenda [agenda] motivates the discovery question; the homogeneous-project network below is an additional benchmark, not an empirical description of pharmaceutical laboratories.

There are $M<\infty$ eligible projects and a finite horizon of integer time slots $0,\ldots,H$. Projects are interchangeable for scheduling, each needs one slot at each of three stages, in order: hypothesis generation, testing, validation. Capacities $k_{s,t}$ are nonnegative integers; waiting is permitted and there is no required processing of unfinished projects. Arrival counts and any initial projects already in the pipeline are fixed and recorded. A project completing validation passes the preregistered quality standard with probability $q\in(0,1]$, independent of its scheduled path and not revealed before completion. Quality draws may be dependent across projects; only the common marginal probability is used. Thus a schedule completing $F$ projects has expected validated output $qF$.

Time expansion preserves the deadline

Create a node $(s,t)$ for each completed-stage count $s\in\{0,1,2,3\}$ and time $t$. Waiting arcs $(s,t)\to(s,t+1)$ have capacity $M$. Processing arcs $(s-1,t)\to(s,t+1)$ have capacity $k_{s,t}$. A source supplies each arrival cohort at its appropriate node with the recorded integer count. Nodes $(3,t)$ connect to a sink with capacity $M$. All arcs advance time or terminate, so the network is acyclic. Arcs not usable before the deadline are omitted. Let $F(k)$ be its maximum source-to-sink flow.
Theorem (Exact completed-discovery capacity): The greatest feasible number of projects completing all stages before the deadline is exactly $F(k)$, and an integer optimal schedule exists. Moreover, \[ F(k)=\min_{C}\operatorname{cap}_k(C), \tag{1} \] where $C$ ranges over source--sink cuts. Consequently expected validated throughput is $qF(k)$.
Proof: Every completed project traces a path from its arrival through the three processing stages to the sink. Summing these paths gives an integer feasible flow, so completed output is at most $F(k)$. Conversely, an integer flow decomposes into unit source--sink paths: start at a positive-flow source arc, follow positive-flow arcs using conservation, reach the sink because the graph is acyclic, subtract one along that path, and repeat. Assign distinct eligible project labels to these units. Each path respects prerequisites, capacities and the deadline, giving a feasible schedule.

For completeness, start with zero flow and repeatedly augment along a residual source--sink path. Integer capacities ensure each augmentation can increase flow by at least one, and total source supply is at most $M$, so the process terminates. At termination let $S$ be the nodes reachable from the source in the residual graph. The sink is not in $S$. Every original arc leaving $S$ is saturated, and every arc entering $S$ has zero flow, since otherwise its reverse residual arc would leave $S$. Conservation implies total flow equals the capacity of that cut. Every feasible flow is bounded by every cut capacity, proving (1) and optimality of the integer flow. Finally linearity of expectation gives $q$ times the number completed, without requiring independent quality draws. ∎

This construction counts only completion by the deadline, not the number of hypotheses entering testing. Recorded waiting times can be recovered from the paths; maximizing throughput alone need not uniquely determine them, so any claim about a particular waiting-time distribution additionally needs a declared tie-breaking schedule.

Exact finite increments and bottleneck complementarity

Fix one processing arc $e$ whose capacity can be increased by $\Delta\ge0$, all else fixed. Partition cuts according to whether they contain $e$ as a forward crossing arc. Let $A$ be the least original capacity among cuts that do, and $B$ the least among those that do not; an empty class has minimum $+\infty$.
Proposition (A finite capacity-value formula): The throughput gain from this increment is exactly \[ F(k+\Delta e)-F(k)=\min\{A+\Delta,B\}-\min\{A,B\}. \tag{2} \] When $A<B$, this is $\min\{\Delta,B-A\}$; when $B\le A$, it is zero. For a real increment cost $g_e(\Delta)>0$, the bottleneck value for expected validated discoveries is $q$ times (2), divided by $g_e(\Delta)$. Integer increments retain the indivisible-project interpretation.
Proof: Increasing the capacity of $e$ adds $\Delta$ to precisely the cuts in the first class and changes none in the second. Applying (1) before and after gives (2). The two cases follow by elementary comparison, and the expected-output and cost conversion follow from their definitions. ∎
If a stage-wide investment changes several arcs, the same argument gives \[ F(k+\Delta a)=\min_C\left\{\operatorname{cap}_k(C)+\Delta\sum_{e\in C}a_e\right\} \] for nonnegative arc increments $a_e$. It is piecewise linear and can have kinks; a finite-increment comparison remains meaningful there even when a unique marginal derivative does not exist.

Zero individual gains do not justify abandoning every capacity package. Consider one three-slot pipeline with at least three arrivals and stage capacities $(1,1,3)$. Throughput is one. Increasing either of the first two capacities alone by one leaves throughput at one, while increasing both raises it to two. Thus stage investments can be complementary. A rule that selects only positive single-stage gains would miss the useful joint investment. This example also respects the complete hypothesis--testing--validation order and its deadline.

A separate productivity objective

Give each processing or waiting arc a nonnegative real resource cost $c_e$ per project, and let $C_0>0$ be fixed laboratory cost for the cohort. For an integer feasible flow $f$, write \[ C(f)=C_0+\sum_ec_ef_e,\qquad P(f)=\frac{q|f|}{C(f)}. \] The finite acyclic network and bounded integer capacities give a finite nonempty set of feasible flows, including zero. This is a ratio of expected discoveries to cost, which is deterministic for the selected flow; it is not an average of project-level or laboratory-level ratios.
Theorem (An exact ratio criterion): Let $P^*=\max_fP(f)$ and $\Phi(r)=\max_f\{q|f|-r C(f)\}$. Then $\Phi(P^*)=0$, $\Phi(r)>0$ for $r<P^*$, and $\Phi(r)<0$ for $r>P^*$. Hence productivity optimization reduces to maximizing a linear flow objective at candidate $r$, with the fixed term $-rC_0$. Maximizing throughput alone need not maximize productivity.
Proof: Every flow has $q|f|-rC(f)=C(f)(P(f)-r)$, with $C(f)>0$. At $r=P^*$ all these expressions are nonpositive and an optimizing flow gives zero. If $r<P^*$ the same flow gives a strictly positive value. If $r>P^*$ every flow gives a strictly negative value, and finiteness makes their maximum negative. The inner objective is linear in flow amounts apart from the fixed cost, so standard finite network optimization or exhaustive finite optimization computes it. For a counterexample to throughput optimality, suppose one completed project costs one, a second requires an additional cost of one hundred, $C_0=1$ and $q=1$. Productivity is $1/2$ for one completion and $2/102$ for two, so maximal throughput has lower productivity. ∎
The ratio criterion is a classical fractional-optimization identity, included with proof; it is not claimed as a new general optimization theorem. An AI package that raises only a nonbinding upstream capacity while increasing fixed computation costs leaves maximum validated output unchanged and lowers productivity at every unchanged positive-output schedule. This is a precise reason to measure real costs and downstream validation, rather than equate faster generation with greater scientific productivity.

Identification and remaining gap

Randomizing the eight stage packages across independent laboratories with fixed baseline cohorts identifies each $E[N(a)]$ and $E[C(a)]$ by arm means under consistency and no cross-laboratory interference. Positive expected cost makes their ratio identified. This alone does not identify untested capacity increments or validate homogeneous-project scheduling. Differential quality, endogenous project selection, rework, uncertain task duration, and priority rules require a richer network or stochastic control model. The theorem supplies exact bottleneck calculations only after those primitives are fixed. Actual capacity values and AI effects in the pharmaceutical population remain unresolved empirical targets.

Completion Estimate

50%

This is a post hoc, heuristic estimate for the full catalogue target. The attempt gives an exact finite-horizon scheduling reduction, min-cut capacity values and complementarities, and a ratio criterion separating productivity from throughput. The laboratory target still requires actual randomized stage-package and capacity evidence, identified quality and costs, and models allowing heterogeneous projects, rework, uncertain duration and endogenous selection.

References

Go to contribution →