AI Through Glass

Learn AI

Start with a question

What makes a machine intelligent? How can the same leaf get two answers? Try a small experiment, then decide what you would trust.

Could your team use it that day? Try a dated release.

Four questions to start with

What is AI, really?

When a phone groups photos of the same face, software is recognising a pattern. When a chat app drafts a caption, a model is generating words. Both are examples of artificial intelligence.

How many ways could a machine sort something? Choose a shape and watch what happens to the square.

Try the shape sorter

How does AI learn?

During training, a model adjusts internal numbers so its outputs better match examples or earn a higher score. A change to the examples can change what the model learns.

In the leaf game, keep the specimen still and add an example. Can you make the same leaf switch groups?

Change the leaf examples · See how labs train a language model

What makes something an agent?

You ask for a train route after lunch. An assistant could suggest a route from your message. An agent with timetable access could choose where to look, read the result and try another route if the first train is full.

Who should approve buying the ticket? Think about the difference between looking up a seat and spending someone’s money.

Explore agentic AI · See how a tool connects

When can I trust an answer?

A fluent answer can leave out the one condition that matters to you. Pick the claim you would act on and find the sentence in the original source that settles it.

Try it on a reply someone is about to send. Would you approve it, change it or ask for more evidence?

Play “Would you send this?”

From early AI to today’s releases

Lesson 1 · What does a rollout mean?

Read a release note and work out which people could use the feature on the date it was announced.

Read the stage and the access terms

Announced
The lab has told readers about the change. Look for a separate date or access statement.
Beta
The lab describes the product as an early release. Look for eligibility and restrictions.
Rolling out
The lab is extending access in stages. Some eligible accounts may receive the change before others.
Generally available
The release is broadly open on the product or platform named by the lab. Plan, region and admin rules can still apply.

A feature can be in beta and rolling out at the same time. Check the plan, region, product surface and any settings or other requirements named in the announcement.

Worked example

ChatGPT Wallet in dots

OpenAI’s 13 Nov 2026 release notes described a gradual rollout in dots. Eligible users were ChatGPT Pro subscribers aged 18 or older in the US with an eligible US-issued card and a US billing address.

That wording tells you both the group that can receive access and the fact that access arrives in stages. The release note serves as evidence for that date. For current terms, consult the provider.

Read OpenAI’s release notes

Practice

Check your understanding

In this hypothetical example, a feature is rolling out in beta to selected Pro accounts in Canada. Who should expect access first?

Next lesson · Where did the model launch?

Lesson 2 · Reading a launch

Where did the model launch?

By the end, you can tell a colleague which part of a launch they can use and which claim still needs a source.

Model
The system that produces an output. A release can name a particular version.
Product or app
The interface someone actually uses, such as a chat app. Its own notes describe the features and accounts it supports.
API
An interface that software uses to make requests. A developer can use a model through an API to build a product.

Read the noun after “available”

Google’s 12 Nov release named Gemini 4 Hydrogen and the Gemini API. That supports a claim about developers using the model through that API. Google’s separate app announcement was published on 12 Nov and updated on 16 Nov. The updated text says free users could reach Hydrogen through automatic model selection when appropriate. The developer and app claims each have a source that names their audience.

A colleague reads the 12 Nov API note and asks if every Gemini app user had the model that day. What can you establish from that note?

Transfer it to a different lab

Read the Grok Bot announcement in the record. Which product did the publisher name, and where could teams use it? Find the words that answer before opening the explanation.

Compare your answer

The publisher named Grok Bot 1.0 and described team use in Slack, Microsoft Teams and video calls. Those details support a product-workflow claim. A claim about a new API model would require a separate developer release.

Practice guides

Tips from the labs

Any chat app · Everyday practice

Find a detail the notice leaves open

You can try this with the chat app you already use. Paste this invented notice and question together.

Notice: The library’s repair café starts on Saturday at 10 am. Bring one item. Question: What time does it finish, and how many items can I bring? Use only this notice.
Check your reply

One item is supported. The finish time is missing. If the answer supplies a time, ask it to quote the words that establish the time. If an exact time matters, check with the organiser.

Try the same check on a different notice

OpenAI · ChatGPT · Writing

Give the prompt a goal and context

State the outcome, provide the relevant information and specify the format you want.

Try it: Rewrite a public event notice for a reader who is new to the topic. Use three bullets and keep the original dates.

Check: Compare every date with the notice.

Read OpenAI’s guidance

Anthropic · Claude · Document work

Show a few examples of the output

Anthropic recommends using several examples that reflect the task and vary enough to show the intended structure.

Try it: Give Claude a public event notice and a few sample summaries with the same two headings. Ask it to organise the new notice under those headings.

Check: Make sure the headings match and each fact comes from the notice.

Read Anthropic’s guidance

Google · Developer access · Gemini API

Choose the table before asking for it

Google’s developer guide recommends specifying the response format. A fixed table can make omissions easier to spot.

Try it. Give Gemini a public event notice. Ask for a table with columns for date, activity and venue, using “unspecified” for missing details.

Check. Match each row to the notice and look for an invented venue.

Read Google’s format guidance

xAI · Developer access · Grok API

Limit a search to the original publisher

The Grok API web-search tool accepts an allowed list of domains. This gives you control over where it looks for evidence.

Try it. With developer access, allow only the publisher’s domain in the web-search request and ask for the date of one public release.

Check. Open the citation. Does the page name that release and give the claimed date? A relevant domain alone cannot establish either.

Read the domain-filter guide

Can a spreadsheet paint with light?

Compare a launch film with a small grid you can relight yourself.