AI, explained for people
who'd rather use it than take a six-month course.

What a model actually does

Think of an AI model as a large pattern machine. It has seen a great deal of material, compressed patterns from that material into numbers, and now uses those numbers to make a best guess about what should come next.

Those internal numbers are called weights or parameters. A single weight usually does not represent one complete idea. What looks like comprehension comes from the patterns spread across many weights working together.

A prompt is a brief, not a magic spell

Good prompts give the model a job, the relevant context, and a clear definition of “done.” If important details are missing, add them before asking for the final answer.

You are helping with [job].
Context: [what matters].
Produce: [specific output].
Check your work for: [known risks].

Context is the part people skip. If the model needs to sound like your business, use your actual examples. If it needs to make a recommendation, explain the constraints. “Make it good” leaves too much for the model to guess.

Why AI gets things wrong

AI systems are built to produce a plausible continuation, not to verify every premise. They can be wrong because the source material was missing, the instruction was vague, the task exceeded the model’s context, or the model made a poor prediction.

How a model learns from examples

Training is a repeated feedback loop. The model makes a prediction, a loss score measures how far it was from the target, and the training process adjusts some of the model’s numbers. Then it does the same thing again, often many thousands of times.

1. Show an example. The model receives an input and, when training requires it, the expected result.
2. Make a prediction. This is the forward pass: the model processes the example and produces an output.
3. Measure the difference. The loss function gives the mistake a number.
4. Work backward. Gradients estimate which adjustable numbers contributed to the mistake.
5. Make a small update. The optimizer changes those numbers, then the next example arrives.

What AI cannot know on its own

A model can produce a useful prediction without knowing whether the request is wise, whether a client will accept it, or whether an unstated business rule matters. It only has the information available in its training, its current context, and any tools the surrounding application gives it.

• It does not know whether a source is trustworthy unless that is checked.
• It does not know what happened outside the information it received.
• It does not know which missing constraint will change your decision.
• It does not know whether an answer is appropriate just because it can produce one.

A useful workflow is smaller than you think

Start with one repeated task where the inputs are fairly predictable and the output can be checked. Do not begin by handing over the entire company and asking for “automation.” Start with a small task and learn where the review points belong.

1. Name the job. What should be different when this is finished?
2. Give it the source material. Examples beat adjectives.
3. Ask for a draft. Keep judgment with a person until the pattern is reliable.
4. Review the misses. The mistakes tell you what the workflow still needs.

A prompt recipe for turning notes into something usable

This is the shape I reach for when a pile of notes needs to become a clear first draft.

Turn the notes below into a one-page brief.
Audience: [who will read it]
Goal: [what they need to decide]
Keep: specific facts and useful caveats.
Avoid: invented details and confident filler.
If something is missing, mark it [NEEDS INPUT].
End with three questions a human should answer.

NOTES:
[paste notes here]

The important line is the one about missing information. It gives the model a clear way to show uncertainty instead of filling in details that were never provided.

Before answering, list the assumptions you are making.
If an assumption could change the recommendation, flag it.

Review the draft for unsupported claims, missing constraints,
contradictions, unclear wording, and details a human must confirm.

The review step is part of the workflow

AI can make a fast first draft. It does not know which details matter most to your client, so keep a human in the review loop. Check names, numbers, dates, claims, tone, and anything that could create a real-world consequence.

Before and after: from contact form to personalized reply

When someone fills out the contact form on this site, a serverless function runs the whole response flow. Here is what that replaced.

The work that moved was the lookup and the first draft. The judgment work, deciding whether the reply is right and whether to follow up, stays with a person.

Local AI versus hosted AI

Hosted tools are easier to start with: someone else runs the hardware and maintains the service. Local models give you more control over data, versions, and behavior, but you also manage the hardware and software yourself.

The right choice depends on the work and the data. “Local” is not automatically safer, and a hosted tool is not automatically careless. Read the terms, identify the sensitive parts of the work, and decide what matters most.

Customizing a model, without rebuilding the whole thing

Methods such as LoRA train a small add-on instead of changing every number in the original model. Think of it as fitting a specialist attachment to a general tool. The base model stays intact; the adapter teaches it a particular style, subject, or behavior.

This is useful when the task is specific and repeated. It is not a shortcut around poor examples. The examples still need to be consistent, relevant, and easy to evaluate.

Agents are workflows, not tiny employees

An agent is a model placed inside a task loop with instructions, context, and sometimes tools. Orchestration is the part that coordinates several steps or specialized agents toward one result.

The useful work is in defining what each step receives, what it is allowed to do, and how the result is checked. A name does not make a workflow reliable. Clear handoffs help.

What didn’t work: trusting a good-looking answer

A clean paragraph with the right tone can still contain invented details. That makes it more dangerous than an obviously broken answer because nobody feels the need to check it.

The examples matter more than the impressive setting

Working through model-training notes has made one thing clear: the configuration is visible, but the examples do most of the teaching. A technically valid training run can still produce a poor result if the examples are inconsistent, the target is unclear, or there is no sensible way to evaluate the output.

Privacy and human review

Before putting information into a tool, know what the tool does with it. Client names, private documents, credentials, financial information, and anything covered by a confidentiality agreement deserve more care than “I’m sure it’s probably fine.”

• Remove identifying details when the task does not need them.
• Use approved tools and understand their data-retention settings.
• Keep secrets, passwords, and access tokens out of prompts entirely.
• Have a person review anything that affects a client, customer, or decision.

Sometimes the best AI workflow contains no AI

If a task is faster to do directly, the rules are too sensitive to share, or the cost of checking the output is higher than doing the work, use the normal tool. There is no prize for adding a model to a task that was already fine.

Comparing tools

The useful question is which tool fits the job in front of you, with the constraints you actually have. The two most people meet first are ChatGPT and Claude, and they overlap a good deal.

Either one handles the majority of everyday work, and neither one removes the review step. Image generation, transcription, and local runners each have their own tools, and the same rule applies there too. Name the job, then pick the tool that fits it.

What to watch

Not every announcement deserves attention. The shifts worth watching are the ones that change what you can actually do.

• Longer context windows. Models can hold more material at once. Practically, that means whole documents, codebases, and conversations stay in view instead of getting summarized away. The work is checking whether the model still used the right part.
• Tool use and agents. Models increasingly reach out to tools, data, and other models on their own. The useful part is not the autonomy but the review gates that keep a wrong step from cascading.
• Local and open models. The gap between hosted and local is narrowing. For sensitive work or predictable cost, running a smaller model yourself is becoming a real option.
• Image, video, and audio quality. Media generation keeps improving, and so does the obligation to label what is real. The technical bar falls; the judgment bar rises.

None of these require a pivot. They change the shape of what is possible, and the ones worth acting on are the ones that touch a task you already do.