An AI agent replicated a landmark economics paper live for the Federal Reserve for $11. Then the economist behind the demo asked whether the same tool was quietly de-skilling him. My conversation with him:
www.forbes.com/sites/johndr...
We tested 30 models for forecasting flu hospitalizations across 48 states. One auto-tuned model with a single covariate beat ensembles of 64. The key: forecast growth rates, not counts, and use national trends to predict state-level dynamics.
An AI system ran a full scientific study end-to-end, and one of its papers beat the median human submission at a major ML workshop. The economics are the real story — a draft paper for a few dollars in compute, landing on institutions that were not built for it.
www.forbes.com/sites/johndr...
Remote islands protect plants from disease. But isolation also favors large-leaved species — and big leaves are easy targets. Two forces, working against each other. New piece in Forbes.
www.forbes.com/sites/johndr...
It was a fun conversation! Thanks @carlbergstrom.com
(Potentially) unpopular opinion. Too long for BlueSky so please read on Substack.
jdrakephd.substack.com/p/in-praise-...
Hi folks - I've mostly migrated to substack. Would love to see my friends here follow me over there, too. It's the main place I'll be posting about ecology, environment, health, epidemiology, academia, and related topics.
jdrakephd.substack.com
Software: we released S4DM (R package) on CRAN implementing plug-and-play + density-ratio workflows:
cran.r-project.org/web/packages...
We advocate ensembles that span assumptions (tight vs broad): they can improve predictions and explicitly map disagreement/uncertainty when data are scarce.
Also: for small samples, training AUC / CV AUC is a weak proxy for independent presence–absence performance; specificity & accuracy carry better.
For ≤20 occurrences, lots of methods are ~indistinguishable from Maxnet on AUC—but they spread across a big sensitivity–specificity gradient, producing very different binary maps.
Main result: no single algorithm wins everywhere.
Maxnet is best on average, but some alternative method beats it for 72% of species.
We compare 3 “data-deficient” families:
• plug-and-play (estimate presence & background densities, take the ratio)
• density-ratio (estimate ratio directly; includes MaxEnt/Maxnet)
• environmental-range (estimate niche limits; e.g., range-bagging)
Problem: SDMs often fail for rare / poorly sampled species—which is most species, and many of the ones we most care about for conservation.
New paper out today 🧵
Flexible methods for species distribution modeling with small samples (Ecography, OA). nsojournals.onlinelibrary.wiley.com/doi/10.1002/...
This #openaccess work was a collaborative effort with Robbie Richards, Ben Carlson, @jdrakephd.bsky.social, and Cory Merow. All code and data are freely available on Github (github.com/bmaitner/sma...) or Dryad (doi.org/10.5061/drya...).
After experimenting for a few months, I've decided mostly to migrate over to Substack. I hope those who follow me here and previously followed me on Twitter will forgive the disruption and click below to follow me there.
Our new paper in #HealthSecurity argues that epidemic intelligence needs a #systems-of-systems framework—integrating #epidemiology, behavior, supply chains, policy, and ecology—rather than siloed models that talk past one another.
Includes an #H5N1 case study.
www.liebertpub.com/doi/10.1177/...
Applied to COVID-19 in California, the approach yields more accurate and more stable short-term forecasts than RNNs, LSTMs, GRUs, Transformers, and naïve baselines.
🔗 royalsocietypublishing.org/doi/10.1098/...
A challenge in epidemic forecasting is that ML models overfit while mechanistic models miss changing transmission conditions. Our new JR Soc Interface paper tests whether physics-informed neural networks—which embed an epidemiological ODE system inside a neural net—can address this.
Biological modeling is organized inquiry, but how should we think about the process?
My new paper in #EcologyLetters argues that we should model like experimentalists: define treatments, measure responses, validate, perturb, repeat.
👉 doi.org/10.1111/ele....
Do you agree?
New essay out in #Science: why is the human fatality rate from the current #H5N1 outbreak so much lower than in past outbreaks?
I explore three possible explanations—and what they mean for pandemic risk.
www.science.org/doi/10.1126/...
Newly expanded version of my guide to scientific writing -- known as the “15 steps” -- published in PLOS Computational Biology. Special thanks to Éric Marty for creating a fantastic visualization.
Check it out: journals.plos.org/ploscompbiol...
#ScientificWriting #PLOSComputationalBiology
Excited to share that I’m joining @bigbiology.bsky.social this season as a recurring guest host. Marty Martin and I teamed up on my first episode with @jaapderoode.bsky.social, and it's just dropped. Hope you enjoy it.
#diseaseecology
“Globalisation, global change and emerging infectious diseases” with @jdrakephd.bsky.social
How do globalisation and climate change influence the rise of new pandemics? Join in person or online – open to all.
Register: www.oxfordmartin.ox.ac.uk/events/globa...
Stream: youtube.com/live/80kXZlQ...
Biological modeling is organized inquiry, but how should we think about the process?
My new paper in #EcologyLetters argues that we should model like experimentalists: define treatments, measure responses, validate, perturb, repeat.
👉 doi.org/10.1111/ele....
Do you agree?
Why are flu vaccination rates stuck?
We studied how “medical mindsets” (naturalist/technologist, minimalist/maximalist, doubter/believer) affect vaccine hesitancy.
These attitudes matter and could help tailor communication to boost uptake.
Read about it in #Vaccine 👉 doi.org/10.1016/j.va...
What if we take a systems approach to thinking about transboundary animal diseases? Our framework, just out in #TrendsInParasitology, suggests common vulnerabilities and opportunities for intervention.
#TADs
authors.elsevier.com/a/1lshW5Eb1x...