Staying Up-to-Date with AI — Level 3: Advanced

📅 As of October 2026 · this area changes almost weekly

Everything you have done so far has one thing in common: you start it. You type a question, the AI answers, and when you close the tab, it stops. An always-on assistant is different. You hand it a job, close the laptop, and it keeps working, even while you sleep.

⚙️ How it works, in 3 steps

1. You give it a job. Sometimes you describe it in plain words. Sometimes you show it once.
2. It works on a computer in the cloud. A computer in a data center (a "virtual machine") with a web browser and your logins. It is never your actual PC, which is why it can keep going while yours is switched off.
3. It works, then checks in. The work happens in the background. When it needs a decision (approve this, enter this code, confirm this payment) it comes back to you.

🌙 Picture a night shift

• 10:00 PMYou type: "Every weeknight, go through my inbox and have a summary waiting for me by 7 AM."
• 11:30 PMYour assistant opens email on its own computer and starts reading.
• 2:00 AMIt finds a supplier invoice. It prepares a note, but will not pay it without you.
• 6:45 AMThe summary and a few draft replies are ready.
• 7:00 AMYou wake up to a briefing, review the drafts and approve them with a tap.

An illustration of the idea, not a screenshot of any one product.

🧭 The four names to know

🎯 Which one fits you?

• 📧 Live in Gmail, Google Docs and Calendar? Start with Gemini Spark.
• 💼 Your team works in Slack or Teams? Look at ChatGPT Dots.
• 🎬 Easier to show a job than to explain it? Try Grok Bot.
• 🛍️ Errands, trips and shopping from your phone? Peek at Muse.

🔍 Click to Expand

Grok Bot comes from xAI, the company behind the Grok chatbot. Your bots run on a cloud computer (one shared by all the bots on your account), sign in to the tools you allow, and work around the clock.

✨ Its signature trick: learning by watching

Most assistants need you to describe a job in words. Grok Bot can learn from a screen recording of you doing the job once. It watches, writes down the steps it thinks you took, and shows you those instructions to review.

🪜 How to teach a bot

1. Pick a repetitive chore. Something you do the same way every time, such as downloading a weekly report, renaming it and emailing it to your team.
2. Record yourself doing it once. Recordings can be up to 10 minutes and have no sound, so move clearly and do not skip steps. For now this works for jobs done in a web browser.
3. Read the instructions it writes. Fix anything it got wrong. This is the most important step. A bot that learned a wrong step will repeat it every single time.
4. Switch it on, then watch the first few runs. Trust grows from seeing it get things right.

Humans stay in the loop for the sensitive moments. The bot asks you for 2FA codes, CAPTCHAs and payments.

Spark lives inside the Gemini app. Switch the app over to Spark and your requests go to an agent running on a dedicated computer in Google's cloud. It is built on Gemini 3.5 and uses the same "agent engine" as Google Antigravity, which you will meet later in this level.

📬 How you hand over work

It connects to Gmail, Docs, Sheets, Slides, Calendar and Drive (and Google Keep and Tasks since June), plus a growing list of other apps such as Canva, Instacart, OpenTable, Dropbox and Zillow, so it can read your information and act on it.

Look at my calendar for next week and write me a short briefing: what is happening each day, and anything that looks double-booked. Do not change anything. Just report back.

Dots are OpenAI's always-on assistants. Each dot is one dedicated helper with its own cloud computer and browser, and it can connect to more than 4,000 apps. You might have one dot for your inbox, another for your research and a third for your weekly reports.

You reach your dots through ChatGPT (web, desktop and mobile), inside Slack and Microsoft Teams, and in a limited texting beta.

🪜 Its best idea: you choose how much freedom it gets

Dots come with Custom Rules, four levels of independence. It is a wonderful way to think about any AI agent:

1. 🏃
   Act independently. Just do it and tell me afterwards.
2. ✅
   Act if pre-approved. Only the things I said yes to ahead of time.
3. 🙋
   Ask first. Check with me before each step that matters.
4. 🤝
   Hand off to you. Get me to the finish line, then I do the last step myself.

Start at the bottom of the list. Move up only as the assistant earns your trust. One rule never changes: password changes always stay with you.

Muse is Meta's friendly, phone-first assistant. You find it on iPhone, Android, muse.ai and WhatsApp. Meta has also shown off video conversations with an on-screen avatar and smart-glasses support, but those are still on the way.

🛍️ What makes it different: it can buy things

Where the other assistants focus on work, Muse focuses on life. It can book and buy through Stripe Link and partners such as Expedia, Instacart, Walmart, Best Buy and Sephora.

📅 As of October 2026 · model names and versions change quickly

Every AI you have used so far lives in someone else's data center. When you type a question, it travels across the internet to a company's computers, and the answer travels back. In a sense, you are renting a brain.

A local AI is different. You download the brain once and run it on your own computer. No internet needed. No subscription. No company reading along.

🌟 Why would anyone do this?

This is the strongest answer to the worries we raised in the Privacy Concerns lesson back in Level 1.

🧮 Will my computer handle it?

The size of a model is measured in "B", which means billions of settings. More settings generally means smarter, but it also needs more of your computer's memory (RAM).

💡 How to check your RAM: on Windows press Ctrl + Shift + Esc, then choose Performance and Memory. On a Mac, open the Apple menu and choose About This Mac.

🚪 Two friendly doors in

Both are free. The next two lessons walk through each one.

🔍 Click to Expand

LM Studio looks and feels like a normal chat app. The big difference is the built-in model catalog, which shows you whether each model will actually fit on your computer before you download it. You can even drag in a PDF and ask questions about it.

🪜 Get started

1. Download and install. Get it from lmstudio.ai. It installs like any other program.
2. Find a model. Click the magnifying glass (Discover). Choose one that shows it fits your computer. On a 16 GB machine, try Qwen3 8B.
3. Download it. It is several gigabytes, so this is a good time for a coffee.
4. Load it and chat. Pick the model at the top of the window and type your first question.
5. Chat with a document (optional). Drag a PDF into the chat and ask about it.

Ollama is lightweight. It sits in your system tray with a small icon, and it now includes a simple built-in chat window. For people who like typing commands, it also has a command line. Best of all, many other apps can plug into it and use your local AI as their engine.

🪜 Quick start

1. Download and install. Get it from ollama.com. A small llama icon appears in your tray or menu bar.
2. Open the Ollama window. Pick a model and start chatting. It downloads the model the first time you use it.
3. Or use the command line (optional). Open Terminal on a Mac or PowerShell on Windows and paste this. Type /bye to leave.

ollama run qwen3:8b

💡 Model names change often. Browse the Ollama library and copy the current name of a model that fits your computer.

⚖️ LM Studio versus Ollama

📅 As of October 2026 · features and limits change often

Antigravity is a workshop for building software with AI agents. You describe what you want in plain English. A team of AI helpers plans it, builds it, tests it and shows you proof that it works. (Google offers a few versions: the Antigravity IDE we are exploring here, a desktop app focused on managing many agents, plus a command-line tool and a kit for developers.)

🏗️ The big idea: you become the general contractor

🪟 Two views

🧾 Proof, not promises

Instead of just saying "done!", the agents leave evidence called Artifacts: a plan you can read, a task checklist, screenshots, and even a screen recording of your finished app working. You review those, no code-reading required, and you can leave comments on them, much like comments in a Google Doc.

The agents can even test their own work. Antigravity has a built-in Chrome browser, so an agent can open the website it just built, click through it and take screenshots.

🧠 Pick the brain

You are not locked in to one AI. Antigravity works with Google's Gemini models and, depending on your plan, other models such as Claude and the open GPT-OSS-120B.

🔗 Fun connection: the same agent engine powers Gemini Spark from earlier in this level. And if you want to review how AI builds apps in the first place, revisit Building Apps and Websites in Level 2.

🪜 How to try it

1. Install it. Download Antigravity from Google's official page at antigravity.google.
2. Sign in with your Google account.
3. Open the Agent Manager and describe what you want built.
4. Read the plan the agent proposes. Approve it, or comment to change it.
5. Review the artifacts (screenshots and recordings) and ask for tweaks in plain English.

Build a one-page website for our church's fall bake sale on Saturday, October 24, from 9 AM to 1 PM in the fellowship hall. Include a friendly headline, the date and time, a list of featured baked goods with emojis, and a simple volunteer sign-up form (name and email only). Use warm autumn colors and make it look good on a phone. When it is finished, open it in the browser and show me a screenshot.

🔍 Click to Expand

📅 As of October 2026 · limits and fees can change

OpenRouter is a service that gives you access to more than 500 AI models from over 80 providers (OpenAI, Google, Anthropic, Meta, xAI and many open models) through a single account.

🌟 Five reasons people love it

• ⚖️ Compare side by side. At openrouter.ai/chat you can ask the same question to several models at once and see which answer you like best. A perfect companion to the AI Leaderboard lesson in Level 2.
• 💳 Pay only for what you use. Instead of a monthly subscription for each AI, you add credits and spend them as you go. It pairs nicely with the Audit Your AI Subscriptions lesson.
• 🆓 Free models. Model names that end in :free cost nothing to use, within limits.
• 🛟 A backup plan. If one provider goes down, OpenRouter can automatically switch to another.
• 🔑 One key for many apps. Lots of AI tools let you plug in your own key, and one OpenRouter key opens every model.

🪜 Get started

1. Create a free account at openrouter.ai.
2. Browse the models. You can filter for free ones.
3. Open the Chat page and add two or three models side by side.
4. Later, if you like it, add a few dollars of credit and create an API key to use in other apps.

Explain the "K-Curve" of AI adoption to a farmer in four sentences, then give one practical tip.

🔍 Click to Expand

On any big AI project, your AI will need to report progress, explain a problem or ask you questions. Left alone, it uses the chat window: long blocks of text that are tiring to read and soon scrolled out of sight. Chat scrolls away. A page stays.

So ask for a page that is beautiful, color-coded and easy to follow. Reports, problems and questions stop being a chore and become a pleasure.

🌟 Five marks of an excellent report

• 🔝 Big on top. The most important thing comes first, and largest. Lesser things sit lower.
• 🗂️ Folded away. Minor details collapse until you click. Nothing is lost, and nothing is in your way.
• 🖱️ Click to answer. Buttons and boxes replace typing, and a Copy My Answers button carries your choices back to the AI.
• ✨ Beautiful and alive. Elegant design and gentle animation make it a pleasure to open.
• 🔄 Always current. The AI updates the same page as the project moves on, so you have one place to look.

Color is the fastest language on a page. A simple code works well:

1. 🔴
   Needs you now. Work is blocked until you act.
2. 🟡
   Your call. A decision only you can make.
3. 🔵
   Good to know. Nothing to do but read.
4. 🟢
   Done. Finished, with nothing left to do.

✍️ The report prompt

Paste this the moment your AI starts to report, explain or ask:

Please give me a clear status report. Design it as a local HTML page that makes the next step obvious.

- Put what matters most at the top, big and bold. Let lesser things sit lower, and fold the least important away.
- Use great color coding: 🟢 done, 🟡 needs my decision, 🔴 needs me now, 🔵 good to know.
- Use plain words, clear structure, indentation and emojis, so I can take it in at a glance.
- Wherever you need me, let me click or type my answer. Add a "Copy My Answers" button so I can paste them back to you.
- Make it beautiful and elegant, with gentle animation.
- Keep it in one file and update that file as we make progress. When it is ready, open it in my browser.

🔁 Change the first line to fit the moment

🪜 How to use it

1. Spot the moment. Your AI starts a long report, explains a problem or asks questions in the chat.
2. Paste the prompt, with the first line that fits.
3. Open the page. The AI saves it as a file and opens it in your browser.
4. Answer with a click. Press Copy My Answers, then paste them into the chat.
5. Keep the same page. As the project moves on, ask the AI to update it.

🔗 Best partner: an AI that can create files on your computer, such as Google Antigravity from earlier in this section.

I'm organizing a potluck dinner for 40 people at our church. Ask me your five most important questions.