Industries / Food, farming and forestry / Agriculture and farming / Animal feed
Formulation
Deciding the recipe: working out what each kind of animal needs at each stage of life, and which mix of ingredients supplies it at the lowest cost.
4
Milestones achieved, from 1951 to 2016.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1951
The least-cost ration
Linear programming is a method for finding the cheapest plan that meets a set of limits. In 1951 Frederick V. Waugh used it to find the cheapest mix of feeds that met a dairy cow's needs, in the Journal of Farm Economics (now the American Journal of Agricultural Economics). An IBM manual for feed makers from the 1960s, based on Purdue University research, said "many of the major manufacturers use LP to reduce formula costs" and that firms that did not "frequently cannot be cost-competitive", and put the saving at a minimum of $2 per ton of feed. Recipe software still starts there: Bestmix, a formulation software maker, wrote in 2026 that for decades such software "has focused on one essential question: what is the least-cost formula".
Credit
Frederick V. Waugh; Purdue University's Department of Agricultural Economics and IBM
1976
Light reads feed quality
Karl Norris and colleagues recorded the near-infrared light reflected from 87 ground forage samples and matched it to laboratory results. Their equations tracked crude protein with a correlation of 0.99 and fiber with 0.96 to 0.98, and predicted crude protein in samples not used to build them with a standard error of 0.95 percentage points. They concluded it had "potential for use in rapid evaluation of forage quality". In 2023 a handheld scanner using the same principle proved accurate enough for routine checks of oilseed cakes in East African markets.
Credit
Karl H. Norris, Robert F. Barnes, John E. Moore and John S. Shenk (Journal of Animal Science, 1976)
1981
Ideal protein for pigs
An ideal protein supplies each essential amino acid in exactly the proportion the animal needs, usually written as ratios to lysine, so none is wasted. The idea dates to the 1950s or 1960s, depending on the account. The UK's Agricultural Research Council first put it into practice for pigs in its 1981 nutrient requirements, by the account of Jaap van Milgen and Jean-Yves Dourmad (2015), and it has since been a common way to state a pig's needs. Together with factory-made amino acids, it lets nutritionists lower a diet's crude protein and so the surplus nitrogen pigs excrete.
LessonA fixed profile is a simplification; Guoyao Wu and Peng Li argued in 2022 that it ignores amino acids animals can make for themselves but still need in the diet for best growth and health.
Credit
the Agricultural Research Council (UK); later work by groups led by M. F. Fuller and David Baker
2016
A daily blend per pig
Candido Pomar's group built feeders that blend two feeds in a new proportion for each pig every day, to match that pig's estimated needs. In an 84-day trial with 70 pigs, feeding each pig 100% of its estimated lysine need, in place of a three-phase group program, cut lysine intake by 26%, estimated nitrogen excretion by 30% and feed cost by US$7.60 a pig (10%), with no loss of growth. A 2026 review by Devi and colleagues puts precision feeding of pigs, across studies, at about 25% less protein and phosphorus eaten and more than 8% lower production cost.
Credit
Ines Andretta, Candido Pomar and colleagues (Animal, 2016)
Watch for
China cuts soymeal in feed
China imports about 100 million tonnes of soybeans a year, more than any other country. In 2023 its Ministry of Agriculture and Rural Affairs (MARA) set targets for the average share of soybean meal in all feed: under 13% by 2025 and 10% by 2030. Its guidelines call for 0.5 to 1.5 percentage points less crude protein in pig and poultry feed, with synthetic amino acids such as lysine, methionine and threonine filling the gap, and more rapeseed, cottonseed, peanut, sunflower and sesame meal. Official figures put the share at around 17% in 2017 and below 13% from 2023. Estimates from the US Department of Agriculture's Foreign Agricultural Service and the industry data provider Huitong put it at about 16 to 16.5% in 2024, and University of Illinois economists Jilang Qing and Joe Janzen concluded that "official Chinese government data likely overstate the size and smoothness of the reduction".
Achieved whenestimates made independently of MARA, from crush, import and feed output data, show the share below 13%.
Claimed by
MARA, official series as reported in December 2025
Open
01
Test AI formulas against solvers
A public test of whether AI chat models and agents can write a least-cost feed formula. Each task gives an animal and stage (broiler starter, growing pig, laying hen, dairy cow, tilapia), 10 to 30 ingredients with prices and nutrient values, and the nutrient limits to meet. A linear programming solver supplies the true cheapest answer. Each model is scored on whether its formula meets every limit and how far its cost sits above the solver's.
The briefCredit · unclaimed02
Open count of China's soymeal
A public, repeatable estimate of the average share of soybean meal in China's feed, for each year since 2017 and each quarter from now on, built only from public data: soybean imports and crush, soybean meal use and stocks, and industrial feed output. Publish the method, the code and an uncertainty range beside MARA's official figure.
The briefCredit · unclaimed
Open 01 · the brief
Test AI formulas against solvers
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of whether AI chat models and agents can write a least-cost feed formula. Each task gives an animal and stage (broiler starter, growing pig, laying hen, dairy cow, tilapia), 10 to 30 ingredients with prices and nutrient values, and the nutrient limits to meet. A linear programming solver supplies the true cheapest answer. Each model is scored on whether its formula meets every limit and how far its cost sits above the solver's.
Why it is still open
Least-cost formulation has been solved by linear programming since the 1950s, but farmers and small mills can now ask a chat model instead. The one published test of AI models on animal diets found in November 2026 asked four models (Manus, ChatGPT, DeepSeek and Gemini) for dog recipes: 61.3% of nutrients fell below the European pet food federation's minimums, and 81% of the diets supplied too little energy. Searches of Europe PMC, trade press and GitHub found no such test for farm feed formulas, which must meet a price as well as nutrient limits.
Built with what exists
An open linear programming solver such as HiGHS, called through SciPy or PuLP; nutrient values from Feedipedia, the open feed database run by INRAE, CIRAD, AFZ and FAO; the APIs of commercial models and open-weight models such as Mistral Large 4 (Notes: "A trillion open parameters"); Python.
First step
Write 20 broiler and pig tasks from Feedipedia values, solve them with SciPy, and run three models on them to see how often they break a limit.
Done looks like
300 tasks across at least five species, each with the solver's optimum; at least six models scored; the share of answers that meet every limit and the median cost gap above the optimum published, with every prompt and answer, under an open licence.
Open 02 · the brief
Open count of China's soymeal
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public, repeatable estimate of the average share of soybean meal in China's feed, for each year since 2017 and each quarter from now on, built only from public data: soybean imports and crush, soybean meal use and stocks, and industrial feed output. Publish the method, the code and an uncertainty range beside MARA's official figure.
Why it is still open
The official series and outside estimates disagree by about 3 percentage points for 2024 (13% against 16 to 16.5%), in a market that made 322 million tonnes of industrial feed in 2023. The University of Illinois comparison that set them side by side was a single article, and one of its outside sources is a private data provider. Searches in November 2026 found no open series that anyone can rerun.
Built with what exists
The US Department of Agriculture's Production, Supply and Distribution (PSD) online database; China's customs import statistics; official industrial feed output totals; a low-cost AI model to pull figures out of Chinese-language reports (Notes: "The jobs agents repeat"); Python.
First step
Download China's soybean meal supply and use table from USDA PSD and the yearly industrial feed output for 2017–2024, and compute a first share for each year.
Done looks like
A series from 2017 to 2025 with a range for each year; a second analyst reproduces it from the raw files to within 0.3 percentage points; and an estimate for 2026 published before MARA releases its 2026 figure, then compared with it.