Kurka Labs / Project COHO / Seminar 1
A practical orientation

Introduction
to AI

Who makes it, what chatbots can and cannot do, and why browser, terminal, cloud, and local tools feel so different.

Designed for a conversation of 1-6 people · approximately 60 minutes

Participant handout ↗ · Presenter guide ↗ · Explore 25 AI tools ↗

First question

What have you already tried?

A chatbot?

ChatGPT, Claude, Gemini, Copilot, Meta AI, or something else?

A useful result?

Writing, research, planning, images, coding, or solving a practical problem?

A disappointment?

A wrong answer, generic prose, confusion, refusal, or a task it could not finish?

A concern?

Privacy, work, trust, cost, ownership, bias, or what happens next?

Five different questions

Do not call all of this “the AI.”

VendorWho offers it?
ProductWhat do you open?
ModelWhat generates?
InterfaceHow do you work?
LocationWhere does it run?
One example
OpenAI makes ChatGPT, which offers access to models through a browser, apps, APIs, and specialized tools.

The company is not the chatbot. The chatbot is not one fixed model. The subscription is not the same thing as API access.

A changing market

Several ecosystems, not one ladder.

OpenAI

ChatGPT, Codex, APIs, and GPT-family models.

Anthropic

Claude, Claude Code, APIs, and Claude-family models.

Google

Gemini, Gemini Notebook/NotebookLM, Workspace tools, and APIs.

Microsoft

Copilot experiences across Windows, Microsoft 365, GitHub, and Azure.

Meta & open models

Model families that can be offered by many services or run more directly.

Many specialists

Search, images, video, voice, coding, research, and industry-specific systems.

Open the clickable AI tools directory ↗

A useful first approximation

Generative AI produces a plausible continuation from patterns and context.

It can transform information and create new combinations. It does not automatically know whether the result is true, current, authorized, or wise.

A foundational distinction

Generating an answer is not the same as finding a record.

Generation

Produces a response from learned patterns and the present context. Flexible, adaptive—and capable of confident invention.

Retrieval

Returns or consults an external source. More inspectable, but still dependent on source quality and interpretation.

Useful strengths

Chatbots are unusually broad transformation tools.

The limits matter

Fluency is not proof.

Demonstration 1

The same question, two chatbots.

I run a small professional service. Give me three practical ways AI could help this month, and name one risk for each.

Compare emphasis, assumptions, uncertainty, and what each system asks before answering.

How you work

The browser makes AI feel like a conversation.

Strengths

Accessible, visual, easy to start, good for questions, uploads, voice, images, research, and iterative discussion.

Boundaries

Usually sees only what you provide or explicitly connect. It may not share the state of your files, programs, or authenticated websites.

How you work

The terminal lets AI work inside an environment.

Strengths

Can inspect permitted files, run commands, test results, compare changes, and carry work across many related artifacts.

Boundaries

More capability creates more responsibility. Permissions, directories, version control, tests, backups, and review become essential.

Different reach

It is not merely a different-looking chat box.

DimensionBrowser chatbotTerminal or coding agent
Starting pointConversationFiles and a working environment
Typical contextMessages, uploads, connected servicesRepository, commands, tests, tools
Typical outputAnswer or artifactVerified changes and runnable work
Primary riskBelieving an unsupported answerAllowing an incorrect action
Human roleQuestioner and evaluatorDirector, reviewer, and authority
Demonstration 2

Advice versus action on a real file.

Ask how to improve a sample document
→
Let an agent inspect the actual file
→
Save a separate revision
→
Review the exact difference

The important step is not that AI can edit. It is that the change can be inspected, tested, accepted, or rejected.

Where computation happens

Interface and location are separate choices.

QuestionCloud serviceLocal system
SetupUsually immediateHardware and software required
CapabilityOften the strongest available modelsDepends heavily on your machine and model
Privacy/controlDepends on provider, plan, and settingsPotentially greater, if the whole workflow stays local
CostSubscription, usage, or API chargesHardware, electricity, setup, and maintenance
AvailabilityInternet and provider dependentCan work offline and remain under your control
Most useful systems mix both

The future is probably hybrid, not a winner-take-all contest.

Local context

Private files, browser sessions, devices, and organization-specific knowledge.

Cloud intelligence

Powerful models and services that would be difficult to operate independently.

Human authority

Purpose, judgment, consent, responsibility, and control of consequential actions.

Working effectively

Give the system something to work with.

1

Objective

What outcome do you actually need?

2

Context

What should it understand about you, the audience, and the situation?

3

Constraints

What must it preserve, avoid, verify, or leave for you?

4

Review

How will you decide whether the result is useful and trustworthy?

Match trust to consequence

Not every AI action deserves the same permission.

Ask
reversible
→
Draft
reviewable
→
Prepare
state-changing
→
Act
consequential

Reading a selected page is different from publishing, purchasing, deleting, granting access, or sending a message.

A glimpse of Seminar 2

From one conversation to a supervised publishing workflow.

Idea
→
Draft
→
Editorial review
→
Platform versions
→
Evidence

Open the completed Why I Email Myself campaign in PubDesk. Show status and proof. Do not publish live.

Seminar 2

From chatbot to collaborator

What changes when AI can work with files, browsers, and workflows—but the human still controls the consequential actions?

Takeaways

Five things to keep.

The experiment continues

What became clearer—and what still feels confusing?

Clearer?

What distinction helped?

Unresolved?

What still feels confusing or unbelievable?

Next?

Which demonstration should we explore?

kurkalabs.dev/coho · Participant handout ↗

Official starting points

Explore the products directly.

Product names, capabilities, prices, and availability change. Verify current details before making a purchasing or deployment decision.

Presenter note