Available today
The film begins with explanation, feedback and guided practice. Here's what that can look like today.

Learning after closing time
Invented example: Maya, 34, runs an independent shop.
Rain catches the amber light outside Maya's shop. Her assistant has the stock notes she's supplied.
She asks for a product description and a reply to someone who wants two matching cups. The description calls the glaze dishwasher-safe. Maya checks the maker's card. Hand washing is recommended. She removes that claim and asks for another version using the care instructions.
The reply offers to reserve a pair. There's one cup left. She explains that another is expected, leaving the date open, and reads both drafts aloud.
Then she tackles markup and margin in her supplier's pricing worksheet.
"Explain it using a cup that costs me $12," she types. "Then let me try."
The assistant walks through a sale at $20. Maya writes that the $8 difference is a 40% markup. It points out that she's divided by the selling price, which gives the gross margin. For markup, she needs to divide by the cost.
"Don't give me the next answer yet."
She tries another price, shows her calculation and asks for feedback. This time she uses the right base. She checks both examples against the worksheet and works a third one on paper before looking at the response. Fees and other costs still need their own lines.
The customer message arrives as she's putting the pen away: one cup will be fine. Maya reaches for a box, then comes back for the care card.
What these tools can do today
I haven't tested these for this page. This is what the companies say their tools do, and what you get depends on your account.
Conversation. Google's Gemini guidance describes asking questions and getting explanations. Ask about the part of a calculation you missed. The answer can still be wrong.
Generation. The same guidance covers first drafts, including emails. Turn stock notes into a product description. A draft may add an unsupported detail, as Maya's care claim did.
Retrieval. Microsoft's Copilot Studio documentation describes finding information in connected sources. An assistant could look up a shop policy. Check the passage: answers can include material beyond that source.
Tool-assisted work. Anthropic's engineering guidance describes connecting language models with retrieval and tools. That could let an assistant check a stock record before drafting a reply. Errors can build across steps, especially when the connection can make changes.
How this looks in my own work
What excites me about AI is how quickly I can take an idea and start doing something with it. I use it daily for everything from reviewing SIEM logs and comparing configurations to creating documents, proposals, SOPs and DIY guides. Sometimes I’ve got a technical problem. Sometimes I’ve just got a voice note and an idea I haven’t quite worked out yet.
One early example was troubleshooting between SonicWall and Fortinet devices. The terminology in the two interfaces wasn’t lining up, and AI helped me work through those differences, identify CLI changes and figure out which logs to review. I could validate the findings faster than I could get answers elsewhere. That got my attention.
Now I’ll talk through a process and use that explanation to build instructions, or take a project into Notion and work through its design, dependencies and missing steps. Being able to ask, “What have I overlooked?” and then explore the answer is a huge part of the appeal. Each revision gives me something more concrete to evaluate, which leads to better questions and a more comprehensive result.
And I enjoy that back-and-forth. I want to understand why a recommendation makes sense, try it against the information I have and learn something along the way. I still check the sources and decide what gets used or published. The exciting part is how much further I can take an idea while staying involved in the thinking.
A research signal
Experimental brain interfaces that restore communication are real research today. Berkeley's 2025 research account describes streaming speech from neural recordings during attempted speech. The online work involved one participant and had limitations; it doesn't establish a consumer learning connection. The brain–computer interfaces guide follows that research more closely.
What we can examine now
What does a specific system do? How is it evaluated? What can it access, and who corrects its mistakes? Those questions matter even when the task looks small. NIST's 2024 generative AI profile describes confident false content and the need to check sources and performance in context. For background, Google's machine learning introduction explains the underlying idea of learning patterns from data.
What to check before you rely on it
- The evidence. Keep the original notes beside the answer. Check names, dates, conditions and promises against them.
- The access. Know which files and tools it can reach. Approve consequential changes before they're made.
- The consequence. Decide who reviews the result and how to fix it if someone relies on a mistake.
Try the exercise in Start Here. Governance and Impact cover responsibility and the people affected. You can carry these questions into Future scenarios and the film.
Sources
- Use Gemini Apps. Google, n.d. Checked October 6, 2026.
- Knowledge sources summary. Microsoft, 2026. Checked October 6, 2026.
- Building effective agents. Anthropic, 2024. Checked October 6, 2026.
- Generative Artificial Intelligence Profile. NIST, 2024. Checked October 6, 2026.
- What is machine learning?. Google, 2026. Checked October 6, 2026.
- A Fast Streaming Brain-to-Voice Neuroprosthesis. UC Berkeley, 2025. Checked October 6, 2026.