Moonshots
The big ideas we are not sure about yet: AI helping design strains, a cannabis-trained model, and a test to find out which AI is best at this industry.Where AI in cannabis might go: strain genetics, a domain-tuned model, and a public benchmark. Honest about what is unproven.Genomic models for cultivars, fine-tuning vs. retrieval for domain models, and a public cannabis benchmark. Honest about what is unproven.
Carry on withCarry on withnext:
- 018 min
Could AI Help Design Strains?Could AI Help Design Strains? Genomic Prediction and the Data Growers Would NeedGenomic prediction for cannabis: marker data, the phenotype bottleneck, what is realistic in 2026DoneDonedone
Corn and dairy breeders already predict how offspring will turn out from DNA. Cannabis could too, but nobody has logged enough plants the same way. Here is what is real, what is hype, and the log you could start this week.Genomic prediction works in corn and cattle because millions of plants and animals were genotyped and measured the same way. Cannabis has a fraction of that data. Here is what the models do, what is hype, and a phenotype log a grower could start this week.Marker-assisted selection and genomic prediction: what they do in agriculture, why cannabis training sets are thin and inconsistent, what is realistic (trait prediction from markers with enough labelled plants) versus hype (a cultivar from a prompt), and a minimum phenotype schema.
- 028 min
Your Own AI Model: When It Makes Sense (Rarely)Your Own AI Model: Prompt vs. Retrieval vs. Skill vs. Decision Model vs. Fine-Tuning (Rarely)Fine-tuning vs. RAG vs. skills vs. a trained classifier: data, cost, maintenance, the one plausible caseDoneDonedone
Everyone asks 'should we train our own AI on our data.' Almost always no. Here is the honest version of when yes, and the one small case where it works.Everyone asks 'should we train our own AI on our data.' Almost always no: a good model plus your documents gets you there. Here is the ladder from prompt to retrieval to a skill to a decision model to fine-tuning, and the one case that is plausible.The decision ladder (prompt, skill, retrieval, small decision model, fine-tune), data and cost per rung, what a 'cannabis LLM' would and would not know, and why the plausible case is a classifier on your own labelled tickets or orders.
- 038 min
How Do We Know Which AI Is Best at Cannabis?Which AI Is Best at Cannabis? Building a Small Private BenchmarkA domain benchmark: why public evals do not measure your work, a 50-task private set, rubric grading, calibrationDoneDonedone
Nobody has tested the big AI models on real cannabis tasks: Metrc rules, 280E, wholesale math. The leaderboards do not measure your work. Here is how to build a small test of your own, and the scorer to run it.Public benchmarks score trivia the models have often already seen. A cannabis benchmark is 50 real tasks, expected answers, a rubric, and a rerun on every model change. Here is how to build one, plus the calibration check for decision models and a scorer you can run today.Why public benchmarks (contaminated, saturated) do not predict domain performance; a 50-task private eval with gold answers and a rubric, rerun per model change; calibration check (ECE, reliability table) for decision models; starter repo with scorer.