AI Tools Field GuideProbabilistic, deterministic, and the other AIs
Some tools give the same answer every time. Models draw one. Asking a model again is not a check, because you get a second draw from the same odds. The spam filter, the route planner, the chess engine and reinforcement learning sit on that same line, below.
Two kinds of tool
A spreadsheet formula, a script, a database query: same input, same answer, every time. That is what deterministic means.
A language model is probabilistic: each word is drawn from weighted shares, so the same request twice gives two texts that say the same thing in different words. How the model decides covers the draw itself.
| Deterministic tool | Language model | |
|---|---|---|
| The same input twice | The same answer | The same sense, different wording |
| When it is wrong | Wrong the same way every time | Wrong in a different way each time |
| Finding the fault | Reproduce it, read the rule, fix the rule | Compare the output against the source |
| What running it again proves | That the fault is real and repeatable | On its own, nothing |
The last row is the practical difference. A deterministic tool's error sits still while you find it. A model's error is one draw out of a distribution, so the next run may hide it instead of confirming it.
Asking again is not verifying
Ask the same question three times and three agreeing answers feel like confirmation. They are three draws from the same loaded die. If the odds were peaked on something wrong, every draw lands there and the agreement tells you only that the model was sure.
Disagreement is a useful signal. Three answers that differ mean the model had little to go on, which is a reason to stop and supply the material. To verify an answer you compare it with something outside the model: the document, a tool, or the person who knows.
The family map
"AI" covers methods that have almost nothing in common. What they share is that a computer produces an answer nobody wrote down line by line. The useful way to tell them apart is where the randomness sits.
| Kind | What it computes | Where the randomness is | Where you meet it |
|---|---|---|---|
| Chat and writing models | The next word, over and over | In training, and in the draw at the end of every word | ChatGPT, Gemini, Grok, Claude, and every other chat box |
| Image and video generators | A picture, by removing noise step by step | The noise it starts from, and a draw at each step | Image and slide tools |
| The spam filter (a classifier) | A score for each category, then the highest one | In training only; the answer itself is the top score | Mail, security tooling, tagging |
| The route planner (a search) | The cheapest path through a map of choices | None. Same map, same route | Maps, logistics, schedulers |
| A chess engine | A search over moves, guided by a learned scorer | A little, on purpose, so it does not always play the same game | Games, wargaming aids |
| A solver or optimizer | The best assignment that satisfies your constraints | None, or a random restart to escape a dead end | Scheduling, manpower, load planning |
| Reinforcement learning | Which action earns the most reward over time | In exploration while it learns; usually fixed once deployed | Robotics, control systems, and the "thinking" training behind chat models |
| Evolutionary search | Keep the best candidates, mutate them, repeat | Mutation and selection are random by design | Design, layout and tuning tools |
The third column splits the list in two. Most of these are random only while they are being built, and the answer you receive is computed.
Chat models and image generators are the ones that draw the answer itself, which is why they are the ones that need checking every time.
Temperature zero and seeds
Two words turn up on any developer page. Temperature zero (also called greedy) stops flattening the shares and always takes the top-scoring word, so the same input gives the same output. A seed is a number that fixes the random draws, so a run with the same seed and the same input repeats.
Both buy repeatability. Neither buys correctness, and a repeatable wrong answer is still wrong. Either one holds only while the service and the model version stay the same.
Entropy, in one paragraph
Entropy is one number for how spread a draw is. Low entropy means one option holds nearly all of it, as with "the capital of France is", and the answer is already decided. High entropy means many options hold a slice each, and the model is choosing among them.
It is the same quantity information theory uses for how much a message tells you, and the same formula physics uses for how spread out a system is. The equation is shared; nobody is arguing the model is a gas.
At your desk it comes down to one habit. The flatter the draw, the more of the answer came from the model rather than from your material.
Why it makes things up
Training grades every word against the one that came next, and there is no cheap way to answer "I don't know". A blank earns nothing, and a confident guess that lands earns full marks.
So guessing is the habit that gets learned, and a flat draw is where it shows.
What to do with this
Pick the one tool in your workflow whose answers you have never re-run, and re-run it twice on the same input. What moves tells you which kind of tool it is, and how much checking it deserves.
Two longer answers have the arithmetic worked by hand. The Q&A book asks whether all AIs are predictive number generators, and what any of this has to do with entropy.
Why it makes things up, in the LLM book, takes the grading argument apart.