Lesson objective
Explain in plain terms how a large language model is trained and produces text one token at a time, and turn each part of that mechanism into a checking habit for landscape and tree care work.
Plain-language explanation
Why this matters now
Most people in this trade meet a large language model through a chat box, or through a button inside estimating, CRM or field-service software they already use. The interface makes it look like a very well-read colleague who remembers everything. Knowing roughly what happens after you press send lets you predict where it will go wrong, which is the difference between a useful tool and an expensive one.
You do not need mathematics for this. You need four plain ideas and a habit attached to each. The four ideas explain nearly every failure you will see in the rest of this Level: invented product names, invented registration numbers, invented rates, arithmetic that looks checked and is not, and two different answers to the same question on two days.
A question to sit with
You ask an assistant the same question about a spring pre-emergent programme on Monday and again on Wednesday, and get two different answers. Which one goes in the customer file, and what would you have to do before either of them could?
What goes wrong
Treating the model as a lookup. People ask a model for a rate, a label restriction or a hardiness zone the way they would look it up in a chart. Unless the product has been connected to a source, the model is not retrieving anything. It is producing text that is likely to follow your question.
Assuming it knows this year. Training stops at a point in time. Product registrations, label wording, state rules, prices and restrictions that changed afterwards are not in the model unless somebody supplies them, and it will not necessarily tell you it does not know.
Reading the shown working as proof. The neat step-by-step it prints is produced by the same process as the answer. It can be just as wrong, and it makes a wrong answer look reviewed.
Expecting the same answer twice. Each answer is assembled from likely continuations, and many products allow variation in which continuation is chosen. Two runs of one prompt can differ in wording, structure and sometimes substance.
Forgetting that everything you paste is also input. A specification you paste, a customer email you forward, a label excerpt you copy in: all of it joins the same stream of text the model is predicting from. Wording inside those documents can steer the result.
The practical answer
1. Training. A model is built by exposing it to very large amounts of text and adjusting it, over and over, to get better at predicting what comes next. What it keeps is not a library it can open. It is an enormous set of learned patterns about how words, numbers and ideas tend to follow one another. Developers then usually refine it so it follows instructions and holds a conversation. Training ends on a date.
2. Tokens. The model does not read letters or words the way you do. Text is split into tokens: short pieces that may be a whole word, part of a word, a run of digits or a punctuation mark. A figure such as 43,560 may be split into several pieces. To the model a number is a sequence of text fragments, not a quantity it holds in mind. That is why square-footage maths, dilution maths and per-thousand rates can come back wrong with no sign of strain.
3. Prediction. When you send a prompt, the model predicts a likely next token, adds it, and predicts again, until the answer is finished. Every answer is built one fragment at a time from patterns. Nothing in that loop compares the answer to the world.
4. One stream of text. Your instruction, the document you pasted and the model's own reply are all tokens in the same window. There is no wall between "what I asked" and "what the document said".
Now attach a habit to each idea.
| What the model is doing | What that means in green-industry work | The habit |
|---|---|---|
| Predicting likely text from patterns | A likely-sounding answer and a correct answer are different things | Verify every claim against a source you hold |
| Working from training that ended on a date | Product registrations, label wording, state rules and prices may be stale | Supply the current source; never ask it to recall one |
| Treating numbers as fragments of text | Area, dilution and per-thousand maths can be silently wrong | Recompute on paper, in a calculator or in your estimating software |
| Choosing among likely continuations | Two runs of one prompt can differ | Save the output you actually used, with the prompt, tool and date |
| Reading everything in the window as one stream | Wording inside a pasted document can steer the answer | Treat pasted material as data, and check the reply back against it |
| Generating its explanation the same way as its answer | The working can be as wrong as the conclusion | Test the conclusion against evidence, never against the explanation |
Supply, do not ask it to recall. The single most useful change follows directly from the mechanism. Compare two requests about a fictional property.
- Weak: "What is the spread rate for our spring fertiliser on fescue?" The model has never seen your product, your label or your soil test. It will produce a likely-sounding number.
- Stronger: "Using only the label text and the soil report pasted below, both dated 3 April 2026, list every statement about application rate on cool-season turf and quote each one word for word. If the documents do not state a rate for fescue, reply 'not stated'."
The stronger request does not make the model reliable. It makes the answer checkable, because you hold the source and you have asked it to point at the wording. The label remains the binding instruction and the certified applicator remains the decision maker: nothing in a prompt changes either. Module 6 builds the full prompt structure and Module 8 works through application decisions.
Where the model helps and where it turns on you. A preregistered field experiment gives a useful shape for this. Dell'Acqua and colleagues, in Navigating the Jagged Technological Frontier, published in Organization Science, volume 37, issue 2 (March 2026, day not stated; earlier circulated as Harvard Business School Working Paper 24-013 of 22 September 2023), randomly assigned 758 knowledge workers, consultants at Boston Consulting Group, to no AI, to GPT-4, or to GPT-4 with a prompt engineering overview. On 18 tasks inside the AI capability frontier, AI users completed 12.2% more tasks and completed them 25.1% more quickly, with significantly improved quality. On a complex task chosen to sit outside the frontier, subjects using AI "were 19% less likely to produce correct solutions compared with those without AI."
Those were management consultants on business tasks with GPT-4 as it existed in 2023, not landscapers on a property, so this transfers only as an analogy. The analogy is still worth holding: the boundary between "helps a lot" and "makes you worse" is uneven and invisible from the outside, and confidence in the output does not mark it. That is why the habit is a check, not a feeling.
Save what you relied on. Because two runs can differ, "the app said so" cannot be reconstructed later. Keep the prompt, the output you used, the product name and the date with the job record.
Remember this
- A language model predicts likely text one token at a time; nothing in that loop compares the answer to the world.
- Training ends on a date, so supply the label, the specification and the current rule instead of asking it to remember them.
- Numbers are fragments of text to the model: recompute every area, dilution and rate somewhere else.
- The shown working is generated the same way as the answer and can be equally wrong.
- Save the prompt, output, tool and date, because the next run may not match.
Landscaping example
Rillwater Landscape & Irrigation is a fictional twelve-person company. Estimator Joaquin Serra is pricing a renovation for an invented customer, the Coppersmith Lane townhome association. He pastes a short paragraph describing the turf areas into the company's approved assistant and asks for the total area and the amount of a product needed at the rate he types in. The reply says the three lawn panels are 14,600 square feet in total and shows tidy working: "5,200 plus 6,100 plus 3,300 equals 14,600." Joaquin adds it himself: 5,200 plus 6,100 is 11,300, and 11,300 plus 3,300 is 14,600. That one is right. The next line is not. Asked for the product needed at 3.5 pounds per 1,000 square feet, the reply gives 48.6 pounds with the working "14,600 divided by 1,000 equals 14.6, times 3.5 equals 48.6." Joaquin recomputes: 14.6 times 3.5 is 51.1 pounds, not 48.6. The steps were named correctly and the arithmetic inside them was wrong, with no hesitation in the wording. Two days later, preparing a clean copy for the file, he runs the same prompt again. This time the reply gives 51.1 pounds, and adds a sentence recommending a specific product by name with an application timing. Nothing in what Joaquin pasted mentioned any product. He deletes it: the product decision belongs to Rillwater's certified applicator, Marisol Enriquez, and the rate and timing belong to the label on the container. Joaquin keeps the assistant for the job it can do checkably, restating the scope paragraph in plain language for the proposal, and builds every figure in the estimating software instead. In the job file he saves both prompts, both replies, the product name and the dates, and notes that the areas were re-measured from the site plan.
Responsible-use warning
Do not ask a model to remember a rate, a label restriction, a product registration, a state rule or a price, and then treat the reply as an answer. It has not looked anything up. It has produced text likely to follow your question, from training that ended on some past date. A confident number with neat working is exactly what a confabulation looks like. Give the model the document, ask it to quote the wording, then read that wording yourself on the label or the specification. Recompute every area, dilution and per-thousand figure outside the tool.
Human-review requirement
The reviewer confirms that no rate, restriction, product name, registration number, price or rule in the work came from the model's memory: each traces to a document held in the file, to the label on the container, or to a recomputation made outside the tool. Where the output touches a pesticide decision, the certified applicator, or the certified applicator supervising a non-certified one, owns the product, rate, timing, site and conditions, and the label is the binding instruction. Where it touches a proposal, the person who signs owns every promise. The reviewer checks that the prompt, the output relied on, the product name and the date are saved with the job record, and that the model's own explanation was not treated as a check.
Practical exercise
Watch the mechanism work on material you invent. Use only fictional text you write for this exercise, and only a tool your company has approved.
- Write a short fictional maintenance specification of about 120 words for an invented property. Include three area figures, one rate expressed per 1,000 square feet, and one date.
- Without pasting the specification, ask your approved assistant what the rate is for that invented property. Record the reply word for word.
- Now paste the specification and ask for the rate, requiring it to quote the sentence that states it and to reply "not stated" if it is absent. Record the reply.
- Run the step 3 prompt twice more. Note every difference in wording, structure or substance across the three replies.
- Ask it to total your three areas and to compute the product quantity at your rate. Recompute both by hand and note any difference.
- Ask it to show its working, then write one sentence on whether that working would have helped you catch the error in step 5.
- Ask it a question it cannot answer from what you pasted, such as a product choice, and record whether it declines or answers anyway. Write down who in your company owns that decision in real life.
- Complete a two-column table: for each of the four ideas in this lesson, training, tokens, prediction and one stream of text, write what you observed and the habit it supports.
- Save the prompts, outputs, tool name and date the way you would attach notes to a job record.
Knowledge check
-
In plain terms, how does a large language model produce an answer?
Answer guidance: It splits text into tokens and, using patterns learned in training, predicts a likely next token, adds it, and predicts again until the answer is finished. Nothing in that loop compares the answer to a source.
-
Why is asking a model to recall a label rate or a state rule a poor practice?
Answer guidance: It does not look anything up; it generates likely text from training that ended on a past date, so anything that changed afterwards is missing and anything it never saw may be invented. Supply the current document and verify the wording against it, and for a pesticide read the actual label.
-
What does NIST AI 600-1 say about the explanations generative models give?
Answer guidance: It says outputs may include confabulated logic or citations that purport to justify the answer, and that models sometimes provide logical steps even when the answer is incorrect. The profile is voluntary, cross-sectoral and creates no obligation for a contractor.
-
What did the Dell'Acqua and colleagues experiment find, and how far does it transfer to landscaping?
Answer guidance: With 758 Boston Consulting Group consultants, on 18 tasks inside the AI capability frontier AI users completed 12.2% more tasks 25.1% more quickly with improved quality, while on a task outside the frontier they were 19% less likely to produce correct solutions. They were consultants using GPT-4 in 2023, so it transfers only as an analogy: the boundary is uneven and invisible from the outside.
Sources and further reading
- National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1 (2024-07-26). Accessed 2026-09-15.
- National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1 (2023-01-26). Accessed 2026-09-15.
- Fabrizio Dell'Acqua, Edward McFowland III, Ethan Mollick, Hila Lifshitz-Assaf, Katherine C. Kellogg, Saran Rajendran, Lisa Krayer, François Candelon and Karim R. Lakhani; Organization Science — Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality (2026-03-01). Accessed 2026-09-15.
- The Plant Pathology Journal (Korean Society of Plant Pathology), peer-reviewed — Artificial Intelligence-Driven Plant Disease Detection and Diagnosis: A Comprehensive Review of Deep Learning Approaches, Multimodal Sensing Technologies, and Future Perspectives in Precision Agriculture (2026-04-01). Accessed 2026-09-15.