For more than 100 million years, mosquitoes have adapted to ice ages, global droughts, and our efforts to eradicate them.

A fossil mosquito preserved in pale stone, photographed under a microscope. Its wings, legs and proboscis are still distinct against the rock.
The fossil of a female Culiseta lemniscata mosquito from the Eocene.

And the more we use insecticides against them, the more likely they are to pass their resistance genes on to another generation. Anopheles stephensi, an invasive mosquito that thrives in cities and breeds in discarded water bottles, is an example of this cycle. It’s spreading across Africa with resistance to the insecticides communities have relied on. “We’ve never seen anything like this, this kind of spread,” said Dr. Ochomo, the head of entomology for the Kenya Medical Research Institute. Scientists are also up against more than the breeding and feeding drives of the animal: conflict, climate-driven displacement, and fragile health systems make this one of the most technically challenging problems of our time.

Machine intelligence needs to face a mosquito

Last month, a frontier model disproved an 80-year-old idea in geometry and solved some of the toughest open problems in mathematics and theoretical computer science. Much of the recent progress in math and coding has come from training models in environments with clear, verifiable answers. What is 458x490? A model will generate many different attempts and receive a clear score for each one without producing any new evidence from the physical world. Here’s what that looks like in practice:

A reinforcement-learning example based on a basic multiplication prompt
How experience becomes learning
Agent The AI model
Environment The prompt and an automated answer checker
State St The model receives: “What is 458 × 490?”
Action At The model answers 224,400
Next state St+1 The response is complete and the attempt ends
Reward Rt+1 −1 in this simplified example, because 224,400 does not match 224,420
Learning update Across many scored attempts, training makes higher-reward solution paths more likely

Acceleration in digital intelligence is happening through verification in domains that can easily access and generate those answer keys. Mosquito biology has no equivalent answer key. We just have to run the experiments to get the data; we can’t buy it from data companies, we can’t scrape it from the internet, and we can’t generate reliable data in silico. AI can read everything humanity has written about mosquitoes and still cannot know how an untested compound will change a living animal. That answer does not exist until the experiment is run and teams of scientists and thousands of labs need to run them.

What if we could change that decision?

It means understanding their behavior is organized around a simple and ancient biological reward: find a viable blood meal while wasting as little energy as possible.

It means treating its behavior, its policy, as a sequence of decisions, rather than as a single outcome such as “landed” or “did not land.”

It’s understanding that verifiable answers do exist in biology, but that they are intrinsically more complex than math and coding, making it harder to isolate the effect of any part of these complex living systems from the other 10,000 things happening. (Did small differences in nutrition change how their olfactory systems developed? Had the colony adapted to laboratory conditions over many generations in a way that changed its behavior?)

How would we go about generating verifiable answers? How do we generate our answer key, and do so at scale? We’re starting by asking this question over and over again: How does compound X impact the behavior of a female Aedes aegypti compared to the control, under controlled temperature, humidity, and airflow in our standardized assay? And because lab results do not always translate to the field, we will use field results as training data too. By comparing what happened in the lab with what happened in the field, models can learn which laboratory signals matter most, helping us design better tests and choose compounds more likely to work at scale. The model strengthens patterns associated with compounds that actually work and weakens patterns associated with compounds that fail (which is the case for most compounds). Across many comparable experiments, it becomes better at recommending the next experiment: which compound, dose, formulation, or controlled assay adjustment is most likely to produce useful evidence and increase the probability of finding another effective compound. The learning signal strengthens, and we take a step closer to finding compounds that will reduce malaria transmission.

Here’s what it looks like:

A simplified example of mosquito behavior becoming a model learning update
How mosquito behavior becomes model learning
Agent A host-seeking female Aedes aegypti
Environment A standardized assay with host cues, compound X, and a matched control
State St The mosquito senses carbon dioxide, odor, heat, airflow, and compound X
Action At Approach the treated area
Next state St+1 Turn away before landing
Experimental reward Rt+1 +1 because, across repeated trials, X reduced landing compared with the control by a predefined amount
Model learning update Add the verified result to the training data, update the model’s predictions and uncertainty, and recommend the next experiment: the compound, dose, formulation, or controlled assay adjustment to test

We need to evolve faster than the mosquito

mZero turns mosquito behavior data into training data. Across thousands of assay results, models can learn which kinds of compounds are most likely to change a mosquito’s decision to land on human skin. Each experiment closes the distance between what a model predicts and what a living mosquito actually does.

Often decades can pass with little progress, then months can produce decades’ worth of progress. The acceleration of machine intelligence, together with the global scientific community’s commitment to generating the data needed to unlock its potential, gives us a rare opportunity to make significant progress in the near-term.

Our goal is that this scientific effort will help scientists get ahead of this highly adaptable animal and create a path toward protecting millions of people from the harm that can follow when a mosquito decides to land on human skin.