Meta Built an AI That Optimizes Concrete. I'm Going to Point It at Malaria.

On March 30th, Meta released BOxCrete, an open-source Bayesian optimization model for designing concrete mixes. It sits on top of their Ax platform, an adaptive experimentation framework that uses Gaussian process surrogates to navigate complex design spaces with noisy data and competing objectives. The concrete application delivered real results: an AI-optimized mix reached full structural strength 43% faster than the original formula while reducing cracking risk, using domestically sourced materials.

I read the engineering blog post and the methodology paper and recognized the problem structure immediately. Not because I know anything about concrete, but because I’ve spent the last few weeks inside a very similar optimization problem. One where the stakes are measured in lives rather than compressive strength.

The Problem I’ve Been Working On

I recently built a multi-agent pipeline that red-teams GiveWell’s cost-effectiveness analyses, the models that inform where hundreds of millions of dollars go for interventions like malaria bed nets, water chlorination, and seasonal malaria chemoprevention. Running the pipeline across three interventions revealed a consistent structural pattern: key parameters are modeled as constants when the evidence says they change over time. Adherence decays. Insecticide efficacy degrades. Drug resistance accumulates. The models treat a dynamic environment as static.

That work addresses the evaluation side: are the models that inform funding decisions as accurate as they could be? But there’s a complementary question it doesn’t touch: are the interventions themselves configured as effectively as they could be?

That’s an optimization problem. And it’s structurally the same one Meta just solved for concrete.

The Parallel

BOxCrete’s problem: a set of input variables (cement types, aggregates, water ratios, admixtures), competing objectives (strength, cure speed, workability, cost, sustainability), physical constraints on ingredients, and noisy lab data that’s expensive to collect. The optimizer learns from each experiment, proposes the most informative next test, and converges on high-performing formulations without exhaustive trial-and-error.

Bed net distribution: input variables (net type, distribution frequency, coverage targets per district, timing relative to rainy season), competing objectives (mortality reduction, cost per DALY averted, coverage equity across income quintiles), constraints (budget, supply chain capacity, community health worker availability), and sparse field data that’s expensive and noisy. Water chlorination dosing: variable source water quality, seasonal turbidity fluctuations, competing demands on treatment resources. SMC scheduling: regional variation in transmission seasonality, accumulating drug resistance, different adherence patterns across socioeconomic strata.

These aren’t loose analogies. Swap the variable names and the optimization structure is identical.

What I’m Building

Ax is open-source under MIT license. The platform is domain-agnostic: you define your parameter space, objectives, and constraints, and it handles the surrogate modeling, acquisition functions, and experiment sequencing. BOxCrete isn’t a new model architecture; it’s Ax adapted to a specific domain through curated data, domain constraints, and validation against real results.

I can do the same adaptation for global health interventions. I’ve already built the spreadsheet integration. My red-teaming pipeline’s Quantifier stage connects to GiveWell’s actual CEA models and runs sensitivity analysis against real parameters. That’s halfway to being the objective function Ax needs.

Here’s what the proof-of-concept looks like:

Phase 1. Take the three interventions I’ve already instrumented (bed nets, water chlorination, SMC) and define their design spaces in Ax. Map the input variables, encode the constraints, and connect the optimizer to the CEA spreadsheets as objective functions. Run optimization and see whether Ax proposes intervention configurations that outperform current defaults on cost-effectiveness.

Phase 2. Incorporate the temporal dynamics the red-teaming pipeline identified. Instead of optimizing against static parameters, feed the optimizer time-varying efficacy estimates: adherence decay curves, resistance accumulation rates, net degradation profiles. This is where Ax’s ability to handle noisy observations matters most, because the field data on these dynamics is sparse and uncertain.

Phase 3. Multi-intervention optimization. GiveWell evaluates interventions independently, but a funder choosing between bed nets, SMC, and chlorination in a specific region faces a portfolio allocation problem with interaction effects. Ax supports multi-objective optimization natively and can produce Pareto frontiers showing the trade-offs between competing interventions under a shared budget constraint.

I can build Phases 1 and 2 with public data and open-source tools. Phase 3 is more ambitious but still tractable as a modeling exercise. All of it will be open-source, same as the red-teaming pipeline.

Where I Hit a Wall

I can build a technically sound prototype. What I can’t do as a solo operator:

Access real-time field data. The proof-of-concept will run against published study data and GiveWell’s public CEA parameters. But a production system needs live connections to health management information systems, DHS surveys, and sentinel surveillance sites. That data exists, but accessing it requires institutional relationships, including data sharing agreements with ministries of health and partnerships with implementing organizations. Code doesn’t solve that.

Validate in the field. An optimizer can propose that shifting bed net distribution timing by six weeks in a specific region would improve cost-effectiveness by 15%. Testing that requires an implementing partner (the Against Malaria Foundation, Malaria Consortium, Evidence Action) willing to run a modified distribution alongside their standard approach. No proof-of-concept, however elegant, earns that trust by itself.

Build production data pipelines. Field data from global health programs is messy, heterogeneous, and inconsistently structured. I documented exactly this problem with GiveWell’s CEA spreadsheets in my first post. Connecting Ax to real-world monitoring systems at scale is the same kind of engineering Meta did when Quadrel integrated BOxCrete into their commercial concrete platform. It’s plumbing, not research, but it’s plumbing that requires a team.

Get it adopted. GiveWell and its grantees are rightly conservative about changing how they design interventions. Millions of lives depend on these programs. A prototype from an independent researcher is interesting. A validated framework backed by an organization with Meta’s engineering resources and a demonstrated track record (they shipped optimized concrete in an actual data center) is a different conversation entirely.

The Ask

Meta’s Adaptive Experimentation team has already proven that Ax can move from software optimization into a physical-world domain with real constraints, noisy data, and material consequences. The jump from concrete to global health is a change of domain, not a change of methodology.

I’m not asking Meta to build this for me. I’m building the proof-of-concept and I’ll publish the results, same as with the red-teaming work. What I am asking is that Meta consider whether their adaptive experimentation framework, already open-source and already validated in a non-software domain, could do more good applied to intervention optimization than to any other single problem their team works on.

The specific contributions that only an organization at Meta’s scale can make: brokering data partnerships with global health implementing organizations, engineering production-grade integrations with field monitoring systems, and lending the institutional credibility that turns a prototype into something the Against Malaria Foundation or Malaria Consortium would actually pilot.

Meta is investing real capital in positioning AI as a public good: open-source Llama, the concrete sustainability work, the domestic manufacturing narrative. I respect that, and the BOxCrete work is genuinely impressive. But the same methodology, applied to how we optimize malaria interventions, would operate on a problem where a 10% improvement in cost-effectiveness doesn’t mean stronger concrete. It means more children alive.

What’s Next

I’m starting Phase 1 now. The proof-of-concept will wire Ax to the GiveWell CEA models I’ve already instrumented and run optimization against the bed net intervention design space. Results and code will be published to the same repo as the red-teaming pipeline.

github.com/tsondo/givewell_redteam

If you work on Meta’s Ax team, at GiveWell, at an implementing organization, or in global health intervention design, I’d like to hear from you, whether that’s to collaborate, to tell me what I’m getting wrong, or to point me at data sources I should know about.

[email protected] · BlueSky @tsondo.com · or via the about page.