S3 E18 Yaniv Altshuler: AI to Reduce Carbon Emissions from Cow Burps and Farts

In this talk

1. 2:37

   Why cow methane matters

   Altshuler says methane released when cows burp represents lost energy that could support milk production. He estimates the resulting opportunity for dairy farmers at $20 billion in additional annual yield and cites a report that puts cattle emissions at about 18% of global emissions. He says suitable feed additives can also improve milk quality, cow health, and mortality.

2. 9:06

   Using AI to match additives to herds

   Altshuler connects his work on swarm intelligence and data analysis to predicting which additives will work for a given herd. Meta AI samples microbes from about 15 cows, sequences their genetic material, and looks for patterns in raw data that conventional methods may miss. The goal is to estimate an additive’s effect at the time of sampling, since a herd’s microbiome can change.

3. 17:11

   Building evidence and finding value

   Altshuler says the company collected microbiome data and direct methane readings across 20 farms, measuring treated cows weekly for several months and comparing results with control groups. He reports that a widely used additive reduced methane by about 8% overall in published findings, but had little effect on roughly one-third of farms and about twice the effect on another third. Better matching could help farmers avoid ineffective treatments and target subsidies or purchasing decisions where benefits are expected.

4. 35:39

   From research to companies

   Altshuler recounts earlier ventures: a game tool that identified highlights from player behavior, and Endor, which applied data analytics to problems including marketing and financial systems. He says Endor helped Mastercard affiliates triple campaign conversions, but the company was eventually unwound during COVID. His advice from that experience is to choose a specific use case and understand the customer journey end to end.

5. 38:23

   Limits of language models for microbiome data

   Asked about applying next-token prediction to genetic data, Altshuler points to limited sample sizes and the rapid growth in possible sequence features. He also explains that shotgun sequencing breaks samples into short fragments of roughly 100 to 150 units, which limits the context available. Meta AI instead searches for statistically meaningful correlations among small pieces of information.