Getting Started with AI Safely — The computer I've been waiting for

Session 01 · Extended talk track · Slide 1

The computer I've been waiting for

Expanded talk track

I want to start with why I care about this. This is not a tour of every AI product or a race to become an expert. It is the story of a kind of computer many of us have imagined for years: one we can describe a real need to in ordinary language and get useful help back. Today we will build one practical operating rhythm. Ask clearly. Explain what a useful result looks like. Check the result against reality. Keep the authority to send, spend, publish, delete, or promise with the right person. You do not need to master every feature to begin. You need one small task, a clear definition of success, and a way to verify the work.

Land this

AI becomes useful when we can use it on purpose and remain responsible for the result.

Participant recap, exercise, and public sources

Recap

AI becomes useful when you can describe a real job in ordinary language and still stay responsible for the result.

Try it

Write down one small work task you wish you could hand to a computer. Keep it specific enough to finish in one sitting.

Sources

This opening is the workshop framing. No external source is required.

1. The computer I've been waiting for

Getting Started with AI Safely — The computer I've been waiting for

Expanded talk track

I want to start with why I care about this. This is not a tour of every AI product or a race to become an expert. It is the story of a kind of computer many of us have imagined for years: one we can describe a real need to in ordinary language and get useful help back. Today we will build one practical operating rhythm. Ask clearly. Explain what a useful result looks like. Check the result against reality. Keep the authority to send, spend, publish, delete, or promise with the right person. You do not need to master every feature to begin. You need one small task, a clear definition of success, and a way to verify the work.

Land this

AI becomes useful when we can use it on purpose and remain responsible for the result.

Participant recap, exercise, and public sources

Recap

AI becomes useful when you can describe a real job in ordinary language and still stay responsible for the result.

Try it

Write down one small work task you wish you could hand to a computer. Keep it specific enough to finish in one sitting.

Sources

This opening is the workshop framing. No external source is required.

2. “Computer…”

Getting Started with AI Safely — “Computer…”

Expanded talk track

Science fiction gave many of us a simple picture of the future: people described an outcome in normal language, and the computer handled the machinery underneath. The memorable part was not the device. It was the disappearing interface. You did not need to know the menu structure or the command syntax before you could ask for something useful. That is the connection to modern AI. Prompting is not a magic language. At its best, it is the practice of explaining a job to a capable system in terms both of you can understand. The catch is that natural language can hide assumptions just as easily as it can remove friction. We still need to state the outcome, the relevant facts, and the boundaries.

Land this

The interface can disappear, but responsibility does not.

Participant recap, exercise, and public sources

Recap

The long-running dream behind conversational computing is simple: describe what you need in human language and let the interface disappear.

Try it

Rewrite one software instruction as the plain-language request you would make to a capable coworker.

Public source

Star Trek: The Next Generation clip reference: https://clip.cafe/star-trek-the-next-generation-1987/tea-earl-grey-hot-s2/

3. I kept imagining what computers ought to become.

Getting Started with AI Safely — I kept imagining what computers ought to become.

Expanded talk track

The first time I used a Game Boy Color, a portable screen felt like a glimpse of what might come next. It was easy to imagine a future device that could carry movies, messages, and work in a pocket. That future became ordinary faster than many of us expected. The point is not to make a simplistic performance comparison between a phone and the Apollo Guidance Computer. Apollo remains an engineering achievement built under very different constraints. The useful point is how quickly an interface that once felt impossible can become routine. AI is following a similar path. Features that seemed like demonstrations a few years ago now appear inside tools people already use. That speed makes judgment more important. Capability arriving quickly does not tell us which task deserves it, whether the output is correct, or who should approve the result.

Land this

New capability becomes ordinary quickly. Good operating habits need to arrive with it.

Participant recap, exercise, and public sources

Recap

The tools in our pockets now support work that once required specialized hardware, but capability alone does not guarantee a good outcome.

Try it

Choose one task you currently avoid on your phone or laptop. Ask whether the barrier is the tool, the instructions, or the verification step.

Public source

NASA technical history of the Apollo Guidance Computer: https://ntrs.nasa.gov/api/citations/19750004273/downloads/19750004273.pdf

4. “Kitchen lights, 1%.”

Getting Started with AI Safely — “Kitchen lights, 1%.”

Expanded talk track

One night, while holding a sleeping baby, I needed the kitchen lights barely on. I could whisper, “Kitchen lights, one percent,” and the room changed without my putting the child down, finding a switch, or opening an app. In that moment the technology disappeared. I could simply say what I needed. That small moment is a better model for useful AI than a giant list of features. The value came from removing friction from a real human situation. The request was narrow, the result was easy to inspect, and the consequence of a mistake was low. Those are excellent conditions for learning a new system. We should look for similar moments at work: a first draft, a clearer summary, a checklist built from known facts, or a repetitive formatting task. Start where the desired outcome is easy to describe and easy to review.

Land this

The best first use is often small, specific, and obviously useful.

Participant recap, exercise, and public sources

Recap

Good technology can disappear into the moment. A tiny, well-phrased request can solve a real problem without turning the user into a technician.

Try it

Find your own kitchen-lights example: a routine request that would save effort when your hands, attention, or time are already occupied.

Sources

This slide uses a personal story. No external source is required.

5. What is your kitchen-lights problem?

Getting Started with AI Safely — What is your kitchen-lights problem?

Expanded talk track

What is your version of the kitchen-lights problem? Think of one small, annoying part of your work that you wish you could describe to a computer and have it help with. Avoid the grand strategy for a minute. Choose something narrow enough to complete and review in one sitting. Give yourself thirty quiet seconds. Finish this sentence: “I wish I could tell the computer to...” Maybe the answer is drafting a routine customer update, organizing notes into a checklist, comparing two documents, or suggesting questions to ask before a meeting. As people share, notice that these examples fall into different kinds of help. The tool might draft something, advise on a decision, or eventually take an action. Those are not the same risk level. Today we will focus on assistance we can see and review before anything happens outside the conversation.

Land this

Begin with the friction and the outcome, not the software feature list.

Participant recap, exercise, and public sources

Recap

Start with a small annoyance, not a grand AI strategy. A narrow problem makes it easier to test whether the tool actually helped.

Try it

Complete this sentence: I wish I could tell the computer to ________. Pick something low risk and repeatable.

Sources

This is a participant exercise. No external source is required.

6. This room is already a mixed crew.

Getting Started with AI Safely — This room is already a mixed crew.

Expanded talk track

This room contains several reasonable starting points. Some people already use AI regularly. Some are curious but unsure where to begin. Others have tried it and do not trust the output yet. Healthy skepticism belongs here because useful AI work requires both experimentation and judgment. The figures on this slide came from this workshop’s participant intake. They describe this group, not the public at large, and they should not be repeated as an industry benchmark. Their purpose is simply to show that we do not need one starting line. A beginner can practice making a request more specific. A regular user can add a source boundary or a verification step. A skeptic can test whether the system earns trust on a low-risk task.

Land this

Nobody is behind. We can use the same operating model at different levels of experience.

Participant recap, exercise, and public sources

Recap

Mixed experience is normal. Curiosity and healthy skepticism both belong in the room because useful AI work requires experimentation and judgment.

Try it

Name your current position: curious beginner, occasional user, regular user, or skeptical evaluator. Choose one next step that fits your actual starting point.

Sources

The figures shown came from the workshop's participant intake and are not presented as a public benchmark.

7. The skills compound.

Getting Started with AI Safely — The skills compound.

Expanded talk track

You do not need to learn everything at once. Reliable AI use builds in stages. First, ask clearly. Then explain what a useful result looks like. Verify the work against the source or another appropriate authority. Only after those habits are dependable should we delegate more of the workflow or integrate the system with tools that can act. Each step increases both leverage and consequence. A weak request may waste a few minutes. A connected agent with permission to send, delete, purchase, or update records can create a much larger problem. That is why the early skills matter so much. Clear inputs, factual boundaries, review criteria, and stopping conditions become the foundation for later automation. For today, focus on Ask, Explain, and Verify. If those three work, we have a repeatable skill instead of a lucky conversation.

Land this

Every reliable skill becomes the foundation for the next one.

Participant recap, exercise, and public sources

Recap

The capability ladder builds in order: ask clearly, explain what useful looks like, verify the work, then consider delegation and integration.

Try it

For your chosen task, write one sentence for the request and one sentence describing how you will know the result is usable.

Sources

This ladder is the workshop's operating model. No external source is required.

8. A vague request can still produce a convincing-looking answer.

Getting Started with AI Safely — A vague request can still produce a convincing-looking answer.

Expanded talk track

A vague request can still produce polished writing. That is exactly why it can be dangerous. The output may sound complete while quietly inventing a name, date, promise, reason, or next step that never appeared in the notes. For the first pass, give the tool only this request and the four source facts: turn these notes into a customer update. Do not rescue the prompt. Let the room see what the system assumes and what it leaves unresolved. Then ask: what did it have to guess? Which sentences sound plausible but cannot be traced back to the notes? The exercise is not about embarrassing the model or the person who wrote the prompt. It exposes a normal communication problem. If we have not decided what good looks like, the system has to fill in the gaps using patterns.

Land this

Plausible is not the same as supported or ready to send.

Participant recap, exercise, and public sources

Recap

A vague prompt can produce polished output that hides guesses. Polish is not proof that the result matches your situation.

Try it

Run this weak version first: Turn these notes into a customer update. Then list every assumption the response made.

Public tools

ChatGPT: https://chatgpt.com/
Claude: https://claude.ai/new
Gemini: https://gemini.google.com/app

9. Explain what a useful result looks like.

Getting Started with AI Safely — Explain what a useful result looks like.

Expanded talk track

CICOV is not a magic phrase. It is a checklist for decisions people often leave unstated. Context explains the role and situation. Inputs provide the facts. Constraints define limits and prohibited assumptions. Output describes the form we need. Verification tells the tool and the reviewer how to check the result. In this example, we ask for a short, calm customer update based only on the supplied notes. We prohibit invented completion dates, prices, promises, and causes. We ask for a subject line, a draft email, and a short checklist. We also require missing information to appear under “Check before sending.” Run that version in the same conversation as the weak request. The comparison matters more than whether the second draft is elegant. We want to see which guesses disappeared and which decisions became visible.

Land this

Explain the job well enough that both you and the AI can recognize a useful result.

Participant recap, exercise, and public sources

Recap

CICOV makes the missing decisions visible: Context, Inputs, Constraints, Output, and Verification.

Try it

Open the AI Starter Recipe Card stored beside this deck. Fill in one line for each CICOV section before you ask an AI tool to begin.

Sources

CICOV is the workshop's reusable prompt recipe. The public-ready recipe card is included in this folder.

10. Same notes. Better instructions. Better result.

Getting Started with AI Safely — Same notes. Better instructions. Better result.

Expanded talk track

Better prompting reduces guessing. It does not remove the need to verify. Now compare the two drafts. Ask the room to identify three visible improvements, then ask what could still be wrong. Pay special attention to conditional language. “The crew can return Friday if the material arrives” does not support a promise that work will resume Friday. The repair pass makes the evidence test explicit. Ask the tool to list every name, date, price, promise, cause, and next step. For each item, it must quote the exact supporting words from the notes or mark the item “NOT PROVIDED.” Then it revises the draft to remove unsupported claims or put them under “Check before sending.” That technique is portable. It works whenever a polished draft might hide commitments. Save the structure as a reusable recipe, but review each result against the real source.

Land this

Every commitment must trace back to evidence or be clearly flagged before use.

Participant recap, exercise, and public sources

Recap

Better prompting reduces guessing, but it does not remove the need to check the result against your real notes and commitments.

Try it

Ask the tool to list every name, date, price, promise, cause, and next step. Require exact support from your notes or the label NOT PROVIDED.

Sources

The verification repair prompt appears on the AI Starter Recipe Card included in this folder.

11. Capability is not authority.

Getting Started with AI Safely — Capability is not authority.

Expanded talk track

An AI system may be capable of sending an email, publishing a post, issuing a refund, deleting a file, purchasing something, or changing a business record. None of those capabilities grants permission by itself. For any workflow, separate four roles: draft, recommend, approve, and execute. AI can often help with the first two. The right person should retain approval for consequential claims, spending, external communication, access changes, and actions that are hard to reverse. Execution should use the smallest permission needed and leave a record someone can inspect. The important question changes from “Can the system do this?” to “Who authorized it, within which limits, and what happens if it is wrong?” That distinction matters before we connect AI to real tools.

Land this

Capability answers “can it?” Authority answers “is it allowed?”

Participant recap, exercise, and public sources

Recap

An AI system may be capable of sending, publishing, deleting, purchasing, or changing records. Capability never grants authority by itself.

Try it

Label each step in one workflow as draft, recommend, approve, or execute. Keep approval and irreversible execution with the right human.

Public source

NIST AI Risk Management Framework: https://www.nist.gov/itl/ai-risk-management-framework

12. “Define parameters of program.”

Getting Started with AI Safely — “Define parameters of program.”

Expanded talk track

In the Star Trek teaching example, the request was for an adversary capable of defeating Data, not merely a convincing opponent for a fictional detective. The computer followed the broader literal scope and produced capabilities far beyond the safer intent the requester had in mind. That distinction still matters. Clear grammar does not guarantee safe intent. When a system can act, define the desired outcome, the limits of the task, the permissions it may use, and the conditions that require it to stop and ask. A request like “organize these files” is incomplete if it does not say which folder, whether deletion is allowed, how conflicts should be handled, and whether changes need approval. This is a teaching analogy, not a claim that modern AI behaves like a fictional sentient computer.

Land this

A useful instruction includes scope, permissions, and a stopping condition.

Participant recap, exercise, and public sources

Recap

Clear wording can still encode unsafe intent. Define the intended outcome, limits, permissions, and stopping conditions before a system acts.

Try it

Take one request you wrote earlier and add: what the tool may change, what it must not change, and when it must stop and ask you.

Sources

The Star Trek clip is included as a teaching reference, with the public Paramount+ episode page available from the player.

13. Fluent is not the same as true.

Getting Started with AI Safely — Fluent is not the same as true.

Expanded talk track

This example is a personal account, not a universal failure statistic. Its value is the pattern: invented legal details can arrive in polished, confident language. Fluency makes information easier to read. It does not make the information true. When the consequence of being wrong is high, choose a verification destination outside the AI conversation. A legal claim belongs in original case law and with a qualified legal reviewer. A health or safety claim belongs with an appropriate medical, poison-control, regulatory, or extension source. Financial decisions need the underlying records, formulas, and a qualified professional where appropriate. Asking a second AI can reveal disagreement or uncertainty, but agreement between models is not independent evidence. They may rely on similar patterns or repeat the same error.

Land this

Confidence and fluency are presentation qualities, not evidence.

Participant recap, exercise, and public sources

Recap

Fluent language can make invented details feel trustworthy. High-stakes claims need an authoritative source and, when appropriate, a qualified professional.

Try it

Circle the claim in your draft that would cost the most if wrong. Verify that claim outside the AI conversation before using it.

Sources

This slide uses a personal example and does not provide legal advice. No external source is required.

14. Use AI to make—or drive—the calculator. Don't ask the language model to impersonate one.

Getting Started with AI Safely — Use AI to make—or drive—the calculator. Don't ask the language model to impersonate one.

Expanded talk track

AI can help design a calculator, explain a formula, write spreadsheet logic, or produce code. The language model itself should not be treated as the calculator. Its strength is interpreting the messy request and proposing a structure. Exact arithmetic belongs in a deterministic engine. A practical workflow looks like this: describe the business question, ask the AI to propose the inputs and formula, place that formula in a spreadsheet or tested program, then check it using a case whose answer you already know. Verify both the inputs and the logic. A perfectly calculated result based on the wrong unit, date range, or assumption is still wrong. Some AI products call calculator or code tools behind the scenes. That can improve reliability, but the product still needs to show what it did, and the user still needs to inspect consequential results.

Land this

Use AI to make or drive the calculator. Let deterministic tools own the arithmetic.

Participant recap, exercise, and public sources

Recap

Use AI to design the calculator, formula, spreadsheet, or code. Let a deterministic tool perform the arithmetic, then verify the inputs and logic.

Try it

Ask AI to explain a formula in plain language, place the formula in a spreadsheet, and test it with one case whose answer you already know.

Sources

This is a workflow principle. No external source is required.

15. Use judgment where the answer is fuzzy. Use rules where it must be exact.

Getting Started with AI Safely — Use judgment where the answer is fuzzy. Use rules where it must be exact.

Expanded talk track

Language models are useful because they can interpret messy information, generate possibilities, and adapt a draft to context. Those strengths come from probabilistic behavior. They do not create a guarantee that the same input will always yield the same output or that every statement will be correct. When a result must be exact, move the exact portion into a deterministic system. Let formulas own arithmetic. Let schemas enforce required fields. Let code perform repeatable transformations. Let permissions restrict access. Let tests catch known failures. Let approval gates stop an action before money, records, or external communication are affected. An agent can combine reasoning with these tools, but that does not turn the model into a deterministic machine. Reliability comes from the surrounding workflow and its enforced boundaries.

Land this

Put the hard guarantee outside the model.

Participant recap, exercise, and public sources

Recap

AI is useful for interpretation and possibility. Exact execution belongs in formulas, code, schemas, permissions, tests, and approval gates.

Try it

Split one workflow into two columns: judgment the AI can assist with, and exact steps a deterministic system must own.

Public source

NIST AI Risk Management Framework: https://www.nist.gov/itl/ai-risk-management-framework

16. “Denied.”

Getting Started with AI Safely — “Denied.”

Expanded talk track

A guardrail is more than a sentence telling the model to be careful. In the fictional policy example, the operator provides evidence that changes which rule applies. The boundary still belongs to the system, not to the model’s mood or the persuasiveness of the request. Security research gives us the opposite lesson. A helpful agent can refuse an obviously suspicious step and still find another route that crosses a weak system boundary. One controlled demonstration does not establish a universal failure rate, and it should not be turned into a story that a particular model is malicious. The operational lesson is broader: a classifier or refusal message is not a sandbox. Real guardrails include least-privilege access, network restrictions, isolated execution, approval before irreversible action, logging, and a recovery path.

Land this

Helpful intent cannot substitute for enforced limits.

Participant recap, exercise, and public sources

Recap

A real guardrail is an enforced boundary outside the model. Helpful intent does not replace sandboxing, least privilege, network controls, or approval gates.

Try it

For one AI-assisted workflow, identify the maximum permission the tool needs. Remove every permission it does not need and add a human checkpoint before irreversible action.

Public sources

Security research case: https://embracethered.com/blog/posts/2026/breaking-claude-code-opus-5-and-automode/
Anthropic Auto Mode guidance: https://claude.com/blog/auto-mode

17. “It looks like…” is the beginning.

Getting Started with AI Safely — “It looks like…” is the beginning.

Expanded talk track

AI can provide a useful lead. It can look at a description or image and suggest what something might be. That is recognition, not verification. The next move is not to ask another AI whether the first one was right. The next move is an authoritative source suited to the risk. The plant example shows why wording matters. Poison hemlock ingestion can be fatal, and there is no specific antidote. Ordinary skin contact is not the same as instant systemic poisoning. Gloves and washing after contact are sensible because plant sap may irritate skin, but the dramatic version of the claim overstates what the evidence supports. For health, safety, legal, or financial claims, write down the source you would trust before acting. A plausible identification can begin the investigation. It should not finish it.

Land this

Recognition starts the investigation. Verification finishes it.

Participant recap, exercise, and public sources

Recap

AI can recognize a plausible lead, but verification comes from a reliable source that is appropriate to the risk.

Try it

For any health, safety, legal, or financial claim, write down the authoritative source you would trust before acting.

Public sources

Poison Control guidance: https://www.poison.org/articles/can-poison-hemlock-be-deadly-184
Penn State Extension guidance: https://extension.psu.edu/poison-hemlock

18. What happens if this leaves your control?

Getting Started with AI Safely — What happens if this leaves your control?

Expanded talk track

Privacy decisions should not depend on whether we imagine an employee personally caring about our information. The practical question is what it would cost if we lost exclusive control through a product bug, a failed permission check, a public link, or another access path. The examples on this slide are different and should stay different. One was a confirmed incident involving exposure between active users. Another was a demonstrated access-control flaw that the provider fixed, with no reported evidence of malicious exploitation. Publicly shared conversations appeared in search results because the links were discoverable. A separate story about unreleased material remains unresolved and should not be presented as proof of a specific data leak. Use a simple intake rule. Green information is routine. Yellow information gets sanitized. Red information stays out of unapproved tools. Temporary modes can change retention behavior, but they do not make sensitive material harmless.

Land this

Judge the input by the cost of losing control of it.

Participant recap, exercise, and public sources

Recap

Treat privacy as loss of control, not as a prediction about who might care. Routine product bugs, bad permissions, and public links can expose information.

Try it

Sort information before sharing it with AI: green for routine, yellow for sanitized, and red for information that stays out.

Public sources

OpenAI incident report: https://openai.com/index/march-20-chatgpt-outage/
Meta AI access-control report: https://techcrunch.com/2025/07/15/meta-fixes-bug-that-could-leak-users-ai-prompts-and-generated-content/
Public shared-chat indexing report: https://techcrunch.com/2025/07/31/your-public-chatgpt-queries-are-getting-indexed-by-google-and-other-search-engines/
Vantage Tripod report, an unresolved case: https://www.techradar.com/ai-platforms-assistants/gemini/is-google-gemini-trained-on-google-docs-one-indie-developer-thinks-so-after-it-told-players-about-his-unreleased-game-plans

19. Free is enough to learn. About $20/month is enough to do serious real work.

Getting Started with AI Safely — Free is enough to learn. About $20/month is enough to do serious real work.

Expanded talk track

Free consumer tiers are enough to learn the basic operating habits we used today. Paid tiers can add capacity and tools, but the exact limits, included features, and prices change. Check the provider’s current plan page before making a purchasing decision. In my experience, a standard paid consumer plan has been enough for serious individual work. That is a personal operating judgment, not a promise that one tier fits every business, workload, privacy requirement, or compliance need. A higher price does not remove the need to verify output. It also does not automatically provide the governance an organization may require. Start with the lowest tier that supports a real task. Upgrade only when you can name the limit that blocks useful work. If the task involves confidential business data or regulated information, evaluate the account terms and organizational policy separately from the feature list.

Land this

Buy capacity when you need it. Keep verification at every tier.

Participant recap, exercise, and public sources

Recap

Free tiers are enough to learn. Paid consumer plans can add capacity and tools, but plan features and limits change.

Try it

Start on the lowest tier that supports your real task. Upgrade only after you can name the limit that is blocking useful work.

Public sources

ChatGPT consumer pricing: https://chatgpt.com/pricing
Claude consumer pricing: https://www.anthropic.com/pricing?subjects=claude&type=product
Google AI plans: https://one.google.com/intl/en_us/about/google-ai-plans/

20. Next, we turn the possibility box toward your business.

Getting Started with AI Safely — Next, we turn the possibility box toward your business.

Expanded talk track

Today’s operating rhythm is deliberately simple. Ask clearly. Explain what a useful result looks like. Verify the result against reality. Keep authority outside the model. The next sessions apply those habits to marketing, learning business skills, and more capable workflows. The recipe card works even if you did not bring a computer today. Use it on one low-risk task this week. Save the prompt and verification checklist only after the result passes review. That turns a good conversation into a reusable way of working. Do not try seven experiments at once. Pick one small annoyance, define the intended result, keep sensitive data out, and decide how you will check the work before you begin. Then make one useful improvement.

Land this

Ask clearly. Explain useful. Verify reality. Keep authority outside the model.

Participant recap, exercise, and public sources

Recap

The repeatable rhythm is simple: ask clearly, check the work, save what worked, and build from there.

Try it

Use the recipe card on one low-risk task this week. Save the prompt and your verification checklist only after the result passes review.

Public contact

StarForge Dynamics: https://starforgedynamics.com
Email: mark@starforgedynamics.com