Open the AI Black Box

Introduction

30 Second Summary

A polished answer can sound trustworthy before you know where it came from. Without a clear mental model, it is hard to tell what a system learned or when its output needs checking.

In this project, you will build a browser-based lab report that traces artificial intelligence from training data to predictions to generated text. Two interactive experiments give you evidence that you turn into a clear explanation inside ChatGPT.

What You'll Build

Picture showing someone your browser-based report as a failed classifier becomes a working neural network before the same learning ideas lead all the way to ChatGPT.

By the end of this project, you'll have:

• A five-part AI mental map you can use to distinguish machine learning, neural networks, generative AI, and ChatGPT at a glance.
• A visual classifier comparison that puts a poor nonlinear baseline beside a successful neural network. You can point to prediction regions and loss to explain what improved.
• A plain-language generation guide connecting learned parameters, inference, tokens, and next-token prediction. A source-backed verification checklist keeps confident output from becoming an unchecked claim.
• Secret Mission: Change one data or model setting to diagnose the effect on generalization or overfitting risk.

Are there any prerequisites?

All you need is a Windows device with internet access. No prior AI experience is needed because the project covers Microsoft Edge and ChatGPT Free setup.

Before We Start

This opening checkpoint captures the explanation you want to give by the end of the project. Your one-sentence goal will keep the coming experiments focused on evidence.

Set Up the Browser Lab

Your personal AI goal gives this project direction. A dependable browser lab removes setup friction before you study how models learn.

This step prepares Microsoft Edge as the home for your lab.

The lab uses TensorFlow Playground for interactive experiments. Google Machine Learning Crash Course provides guided exercises. ChatGPT Free stores your report.

In this step, get ready to:

• Open Microsoft Edge on Windows.
• Verify both Google machine-learning pages.
• Start a saved ChatGPT lab conversation.

Open Microsoft Edge

A single browser keeps your experiment pages together. You can switch between the model controls and your saved notes without losing your place.

• Press the Windows key to open Start search.
• Type Microsoft Edge into the search field.

✔️ Microsoft Edge appears

• Press Enter to open the selected result.

ⓧ Microsoft Edge is unavailable

• Switch back to the browser showing this project.
• Visit the official Microsoft Edge download page.
• Download Microsoft Edge for Windows.
• Run the downloaded installer.
• Approve any Windows installation permission request.
• Press the Windows key after the installation finishes.
• Type Microsoft Edge into Start search.
• Press Enter to open Microsoft Edge.

Good start. Microsoft Edge is now the home for every browser experiment in this project.

Microsoft Edge Won't Open?

• Restart Windows after the installer finishes.
• Repeat the Start search steps after Windows restarts.

Get help with opening Microsoft Edge on Windows.

Verify the machine-learning pages

TensorFlow Playground turns model behavior into a visible browser experiment. The Crash Course page keeps a guided exercise ready for later.

• Open https://playground.tensorflow.org/ in a new Microsoft Edge tab.
• Scan the Playground interface for its main labeled areas.

You should see the Data, Features, and Output areas.

• Open the Google neural-network interactive exercises in another Microsoft Edge tab.
• Review the page for its interactive exercise content.

You should see interactive exercises for a neural network.

Why Browser-Based Tools?

A coding notebook would add setup tasks before you can see a model learn. Browser widgets keep this project focused on visible model behavior.

Both learning pages are now ready. You have the experiment controls and the later neural-network exercise in separate tabs.

Pages Not Loading?

• Refresh the affected tab once.
• Use the official link above again if the address differs.

Get help with loading the machine-learning pages.

Start the saved ChatGPT lab conversation

A signed-in ChatGPT Free session keeps your lab conversation in chat history. Each later experiment can add evidence to the same report.

Account sign-in can feel sensitive. Your credentials stay on the ChatGPT sign-in page.

• Open https://chatgpt.com/ in a new Microsoft Edge tab.

✔️ I have a ChatGPT account

• Sign in with your existing ChatGPT account.
• Complete any account verification requested during sign-in.

ⓧ I need a free account

• Select the account creation option on the ChatGPT page.
• Create a free ChatGPT account.
• Complete any account verification requested during registration.

You should now see a signed-in ChatGPT chat screen. Your conversations can remain available in chat history.

• Start a new conversation from the ChatGPT home screen.
• Enter the following test prompt in the message box:

Reply with: AI browser lab ready.

What Does This Test Prove?

• The prompt requests a fixed response that is easy to recognize.
• A reply confirms that the chat can accept messages before the lab report begins.

• Send the prompt.
• Check the reply from ChatGPT.

You should see a response containing AI browser lab ready..

• Open your chat history.

You should see the new conversation listed in your history. That saved conversation becomes your lab report in the next step.

ChatGPT Not Responding?

• Refresh ChatGPT once if no reply appears.
• Sign in again if your chat history is missing.

Get help with the ChatGPT lab test.

Before the final check, do you expect every lab tab to respond without another sign-in or reload?

• Return to the TensorFlow Playground tab.
• Inspect its labeled interface areas.
• Switch to the Google neural-network exercise tab.
• Inspect its interactive exercise content.
• Switch to the ChatGPT tab.
• Inspect the latest prompt and response.

You should see Data, Features, and Output in Playground. The neural-network exercise page should still be available.

Your ChatGPT conversation should still show the test exchange in chat history.

Your browser lab is ready. Next up, you'll turn overlapping AI terms into a five-part hierarchy inside your saved report.

Build Your AI Mental Map

Your browser lab is ready. Your saved ChatGPT conversation can now become the report that supports your personal goal.

AI terms often overlap in everyday conversation. A five-part hierarchy gives each term a distinct role.

In this step, get ready to:

• Create a five-row hierarchy for the main AI terms.
• Correct each row with a clear hierarchy anchor.
• Add a glossary grounded in a classifier experiment.

Create the five-row hierarchy

A hierarchy arranges related ideas by scope. Your first draft gives the report a visible structure that you can inspect.

• Switch back to the saved ChatGPT lab conversation from your browser test.
• Start the AI mental-model lab report by copying the prompt below:

Create an AI mental-model lab report. Start with a five-row table covering AI, machine learning, neural networks, generative AI, and ChatGPT.

Use these columns: scope, input, output, and example.

What does this prompt create?

• The AI row gives the table its broadest scope.
• The machine learning row describes a method that learns from data.
• The neural networks row identifies one model family.
• The generative AI row covers systems that create new content.
• The ChatGPT row identifies the product you are using for the report.

• Submit the prompt.

You'll see a table with five rows. Its columns cover scope, input, output, and an example.

Missing rows or columns?

• Ask ChatGPT to include every named term if the response has fewer than five rows.
• Ask ChatGPT to preserve all four requested columns if any column is missing.

Help me repair the five-row AI hierarchy in my ChatGPT lab report.

Correct the hierarchy anchors

Generated explanations can vary between responses. Clear anchors keep the field, learning method, model family, content category, and product in consistent positions.

• Revise the table by copying the prompt below:

Revise the five-row table using these hierarchy anchors:

- AI is the broad category.
- Machine learning trains models from data.
- Neural networks are one model family.
- Generative AI creates new content.
- ChatGPT is a product that provides access to language models.

Keep the scope, input, output, and example columns.

Why do these anchors matter?

• The anchors place each term at a specific level of the hierarchy.
• The scope column shows how broad each level is.
• The input and output columns show what flows through each level.
• The example column turns each definition into something concrete.

• Submit the revision prompt.
• Read the revised table from the AI row through the ChatGPT row.

You'll see five distinct roles in a single hierarchy. The table now moves from the broad field to the product you are using.

Did the table lose its structure?

• Ask ChatGPT to revise the existing table if it created a separate explanation.
• Ask ChatGPT to restore the four requested columns if the response switched to paragraphs.

Help me restore the corrected hierarchy as one five-row table.

Add the classifier glossary

The hierarchy explains where each term sits. A compact glossary now explains how data becomes a prediction.

• Extend the lab report by copying the prompt below:

Add a compact glossary beneath the table for these terms:

- feature
- label
- parameter
- training
- loss
- evaluation
- inference

For every definition, use the same classifier experiment that predicts whether a point belongs to the orange class or the blue class. Explain the role of each term in one sentence.

How does the classifier glossary connect?

• A feature is an input value used to make a prediction.
• A label is the answer that the classifier learns to predict.
• A parameter is a learned number that shapes the classifier's predictions.
• Training is the process that updates the model from examples.
• Loss measures the gap between a prediction and the correct label.
• Evaluation checks predictions against known labels.
• Inference applies the trained model to an example without a known label.

Before you submit the final prompt, do you expect the glossary to read like separate facts or one connected learning cycle?

• Submit the glossary prompt.
• Read the finished hierarchy from AI through ChatGPT.
• Confirm each row identifies what the term is.
• Confirm each row describes what the term does.
• Confirm each row includes one example.
• Confirm each glossary definition uses the same classifier experiment.

You'll see five distinct hierarchy rows above seven connected glossary definitions. Each definition now points back to the same orange-or-blue prediction task.

Strong work. Your saved conversation now contains the first complete section of your AI mental-model lab report.

Do the definitions feel disconnected?

• Ask ChatGPT to rewrite all seven definitions around one shared orange-or-blue prediction task.
• Ask ChatGPT to keep each definition to one sentence if the glossary becomes too long.

Help me connect my classifier glossary into one learning cycle.

Your mental map now gives every term a clear place. Next, you'll use that vocabulary to watch a simple classifier struggle with a nonlinear pattern.

Watch a Simple Model Fail

Your lab report now has a five-part AI hierarchy plus a classifier glossary. Next, you will turn those definitions into visible evidence.

Some class patterns curve around each other. A simple classifier using raw features has limited ways to represent that nonlinear pattern.

In this step, get ready to:

• Run the simple baseline for at least 300 epochs.
• Capture its prediction mismatch plus its Test loss.
• Add the baseline evidence to your ChatGPT lab report.

Open the nonlinear baseline

The Google Machine Learning Crash Course includes a deliberately simple baseline in Exercise 2. Its orange and blue classes form a roughly spherical pattern.

• Create a new tab from the Microsoft Edge tab bar.
• Load the feature-cross exercises by entering https://developers.google.com/machine-learning/crash-course/categorical-data/feature-cross-exercises in the address bar.
• Scroll to Exercise 2.
• Confirm that the dataset shows a roughly spherical orange-and-blue pattern.

Why start with the baseline?

The simple baseline gives you a comparison point before you use a more flexible model. Seeing its limit makes the later improvement measurable.

Cannot find Exercise 2?

• Search the page for Exercise 2.
• Confirm that the page address ends with feature-cross-exercises.

Ask for help with the exercise page.

Train the simple baseline

An epoch tracks how long the model has been training. The exercise displays an epoch count while the model updates its predictions.

Before you start training, do you think the prediction background will closely follow the curved point pattern?

• Select the Run/Pause button.
• Wait until the counter reaches at least 300 epochs.
• Select the Run/Pause button again to pause the model.

You have now caught the simple model at its limit. You should see many orange points inside blue prediction regions plus a high Test loss.

What do the colors reveal?

• Orange points generally represent negative values. Blue points generally represent positive values.
• The colored background shows the model's predictions.
• Stronger background color shows greater prediction confidence.
• A high Test loss shows a large mismatch between predictions and the true labels.

Epoch count not moving?

• Select Run/Pause once if the epoch counter stays fixed.
• Confirm that the widget belongs to Exercise 2 on the feature-cross exercises page.

Help me diagnose a frozen exercise.

• Capture the paused epoch count here: your paused epoch count.
• Capture the approximate Test loss here: your approximate Test loss.

Save the evidence and reset

A useful lab observation combines the model setting with the visible result. It also explains what the result means.

What should the observation include?

• Paused epoch count: your paused epoch count.
• Approximate Test loss: your approximate Test loss.
• Visual interpretation: Many orange points appeared in blue predicted regions because the simple model could not represent the nonlinear pattern.

• Switch back to the saved ChatGPT conversation from earlier.
• Compose one request in the conversation input using the three details above.
• Send your request.

You should see ChatGPT add a baseline nonlinear-classifier observation to the lab report. The entry should preserve your measured result plus your interpretation.

Missing part of the observation?

• Ask ChatGPT to revise the observation with the missing epoch count.
• Ask ChatGPT to revise the observation with the missing Test loss.
• Ask ChatGPT to add the explanation about the nonlinear pattern.

Help me improve my baseline observation.

Before you reset the widget, which parts of the experiment do you expect to return to their starting state?

• Switch back to the feature-cross Exercise 2 tab.
• Select the curvy reset arrow.

You should see the exercise return to its starting state. The saved observation remains in your ChatGPT conversation.

Before you check the report, do you expect resetting the browser widget to change the evidence saved in ChatGPT?

• Return to the saved ChatGPT conversation.
• Confirm that the baseline observation contains all three recorded details.

That baseline is now preserved in your lab report. You have direct evidence that the simple model could not represent the curved pattern.

Your failed baseline now gives you something concrete to improve. Next, you will use a hidden layer plus nonlinear activation to help a neural network fit the same kind of pattern.

Train a Nonlinear Neural Network

Your baseline exposed the gap. A simple model left many orange points inside blue prediction regions.

Its high test loss showed that the simple model could not fit the nonlinear pattern.

A neural network uses a hidden layer to build new combinations of the inputs. ReLU applies a nonlinear activation so those combinations can shape a useful decision boundary.

In this step, get ready to:

• Inspect the input layer, hidden layer, and output layer.
• Configure one hidden layer with three neurons using ReLU.
• Document a run where both losses fall below 0.2.

Inspect the three layers

• Switch back to the Google Machine Learning Crash Course neural-network interactive-exercises tab from earlier.
• Scroll to Exercise 2.
• Identify the input layer on the left side of the diagram.
• Identify the hidden layer in the middle of the diagram.
• Identify the output layer on the right side of the diagram.

You should see the input layer feed values into the middle layer. The output layer displays the model's prediction.

How do the layers divide the work?

The input layer receives the selected features. The hidden layer combines those values through neurons.

ReLU transforms each hidden neuron's output before it reaches the next layer. The output layer uses those transformed signals to make a prediction.

Train the verified configuration

• Set the diagram to one hidden layer using the layer controls beside the network.
• Set the hidden layer to exactly three neurons using the controls above that layer.

You should see one middle column containing three neuron circles between the input layer and output layer.

• Select 0.01 from the Learning rate control.
• Select ReLU from the Activation control.

The controls should now show a learning rate of 0.01 with ReLU selected.

Before you start training, do you expect the prediction background to remain patchy or form a clearer boundary around the colored points?

• Select Play.
• Watch the prediction background as the network trains.
• Keep the model training until both Training loss and Test loss are below 0.2.

✔️ Both losses are below 0.2

You got it. The network now produces a prediction background that follows the nonlinear pattern more closely.

ⓧ One loss remains at or above 0.2

A stubborn run can happen because the model's parameters start from random values.

• Select the reset control above the Exercise 2 widget.

You should see the network return to its starting state.

• Return the diagram to one hidden layer.
• Return the hidden layer to exactly three neurons.

You should see one middle column containing three neuron circles again.

• Select 0.01 from the Learning rate control.
• Select ReLU from the Activation control.
• Select Play.
• Keep training until both loss readings are below 0.2.

The exact values can differ between runs. The verified configuration should eventually bring both readings below the target.

That's the difficult part complete. Your successful run shows that hidden neurons plus nonlinear activation can fit the pattern that defeated the baseline.

• Record the displayed Training loss here: your Training loss below 0.2.
• Record the displayed Test loss here: your Test loss below 0.2.

Add the result to your lab report

• Switch back to your saved ChatGPT lab conversation from earlier.
• Add the successful run to your report by sending this prompt:

Add a section titled "Successful neural-network experiment" to my AI mental-model lab report.
Record this configuration: one hidden layer with three neurons, learning rate 0.01, and ReLU activation.
Record my Training loss as [[TRAINING_LOSS="your Training loss below 0.2"]].
Record my Test loss as [[TEST_LOSS="your Test loss below 0.2"]].
Explain how hidden neurons combine features.
Explain how ReLU adds nonlinearity so the network can learn a useful nonlinear decision boundary.
Explain that lower loss means the model's predictions are closer to the correct labels.
Keep the explanation concise and in plain language.

What does this prompt add?

• The configuration records exactly how you produced the successful run.
• The two loss values preserve evidence from the experiment.
• The explanation connects hidden neurons to the improved decision boundary.
• The loss explanation connects the displayed values to prediction error.

Before you review the new section, do you expect both low losses to support the improvement you saw in the prediction background?

• Confirm that the report names one hidden layer with three neurons.
• Confirm that the report records a learning rate of 0.01.
• Confirm that the report names ReLU as the activation.
• Confirm that both recorded losses are below 0.2.
• Confirm that the explanation connects hidden neurons to a useful nonlinear decision boundary.

Your report should now contain the successful configuration, both observed loss values, and the explanation behind the improvement.

Missing part of the report?

• Compare the generated section with the five checks above.
• Send a follow-up request that names any missing result.

Ask for help reviewing the section: Help me check whether my neural-network lab result includes the configuration, both losses, and the explanation.

You've replaced a failing baseline with a neural network that handles the nonlinear pattern. Next, you'll connect learned parameters plus inference to language generation in ChatGPT.

Connect Neural Networks to ChatGPT

Your earlier neural network experiment made learning visible through its changing prediction background. The language model behind ChatGPT applies learned patterns to language.

In this step, you'll connect learned parameters to language generation. First, you'll see how several next pieces of text can all be plausible.

You'll then trace the path from training to inference. Finally, you'll add source-backed verification guidance to your completed lab report.

In this step, get ready to:

• Test how one sentence can have several plausible one-word completions.
• Explain the path from training to next-token prediction.
• Add source-backed verification guidance to the final lab report.

Test plausible sentence completions

A language model predicts a plausible next piece of text from context. Several completions can fit the same sentence, so generated output can vary.

• Switch back to your saved ChatGPT lab conversation in Microsoft Edge.
• Add the sentence-completion activity by pasting this prompt into the conversation:

List five plausible one-word completions for: Instead of turning left, she turned ___. Then explain why more than one completion can be reasonable.

What Does This Prompt Test?

• The fixed sentence keeps the context consistent across every completion.
• The five answers reveal that several continuations can fit the same context.
• The explanation connects those valid alternatives to variation in generated responses.

• Press Enter to send the prompt.
• Count the one-word completions in the response.

You should find five one-word completions. The explanation should show why the same context can support several reasonable continuations.

Missing Five One-Word Answers?

• Edit the request so it explicitly asks for exactly five single-word answers.

Help me tighten this sentence-completion prompt.

You now have a concrete language-generation observation in your lab report. Multiple plausible completions make response variation visible.

Connect training to next-token prediction

Text can be divided into tokens that represent whole words, parts of words, or punctuation. During inference, a language model uses its learned parameters to predict a plausible next token from the context.

• Add the technical connection to your lab report by pasting this prompt into the same conversation:

Explain, in light technical detail, how training updates weights or parameters, how inference uses those learned values, and how text can be divided into tokens for next-token prediction. Add my observed baseline and neural-network loss results to the lab report.

How Does This Prompt Build the Connection?

• The first part links training to learned weights or parameters.
• The second part explains how inference uses the values learned during training.
• The final part connects tokens to next-token prediction.
• Your recorded experiments anchor the explanation in evidence you observed.

• Press Enter to send the prompt.
• Read the section that explains how training updates learned values.
• Compare the included experiment results with your recorded observations.

You should see training described as the process that updates weights or parameters. You should see inference described as the process that uses those learned values.

Your experiment results should anchor the explanation in the failed baseline. The successful neural-network result should show how a more suitable model handled the nonlinear pattern.

Seeing Different Experiment Results?

• Compare the reported values with the observations already stored in your lab conversation.
• Ask ChatGPT to preserve your recorded results if it replaces them with invented values.

Help me keep my recorded experiment results accurate.

Add an official verification checklist

Output that sounds polished can include incorrect information. Verification checks important claims against reliable sources.

• Add the accuracy warning to your report by pasting this prompt into the same conversation:

Search for OpenAI's guidance on whether ChatGPT always tells the truth. Cite the official source. Add a warning about inaccurate output and a verification checklist to my lab report.

Why Require an Official Source?

• The search request asks ChatGPT to find current guidance.
• The source restriction keeps the warning grounded in OpenAI's own documentation.
• The checklist turns the warning into a repeatable verification habit.

• Press Enter to send the prompt.
• Open the official OpenAI citation included in the response.
• Read the cited guidance to confirm that it warns about incorrect or misleading output.
• Return to your saved ChatGPT lab conversation.
• Compare the report warning with the official guidance.
• Check that the verification checklist recommends reliable sources for important information.

The official page should support the warning that confident output can be inaccurate. Your checklist should give you a practical way to verify important claims.

Citation Does Not Open an Official Page?

• Check that the cited page belongs to OpenAI before treating it as the official source.
• Ask ChatGPT to repeat the search using only official OpenAI guidance if the citation points to another publisher.

Help me find the correct official OpenAI guidance.

Before you inspect the full report, predict whether every required piece now has direct evidence behind it.

• Confirm that the report still contains the corrected five-row AI hierarchy.
• Confirm that the report contains the baseline observation from at least 300 epochs.
• Confirm that the report contains the successful neural-network observations.
• Confirm that the report contains five plausible one-word sentence completions.
• Confirm that the report explains learned parameters during training.
• Confirm that the report explains tokens during next-token prediction.
• Confirm that the report contains the official OpenAI citation.
• Confirm that the report ends with a verification checklist.

The report should contain the baseline failure from at least 300 epochs. It should also contain the neural-network result with both Training loss and Test loss below 0.2.

The language section should contain five completions. It should connect learned parameters to tokens through inference.

The final section should contain the opened OpenAI citation. It should also contain a verification checklist.

Strong finish. Your lab report now traces a clear path from training data to verified AI output.

Secret mission

Test Generalization Under Pressure

Your network handled nonlinear data with a known-good configuration. Increase data noise alone to discover whether the model still performs well on unseen examples or begins to struggle.

Clean Up Your Resources

Clean Up Your Resources

This browser lab leaves no paid resources running. Decide whether to keep the saved conversation, archive it for later, or delete it.

Resources you used:

• Saved ChatGPT Free conversation containing your completed AI mental-model lab report.

Keep everything running

No action needed. Choose this if you want to keep the lab report in your chat history as a study note.

• Keep the saved ChatGPT Free conversation in your chat history.
• Return to the report whenever you want to revisit the AI hierarchy.
• Use the recorded losses to compare future Playground experiments.

Pause - I'll come back to this later

Archive the saved conversation to remove it from your active chat history while keeping the lab report. The browser experiments have no running process to pause.

• Switch back to your saved lab conversation in ChatGPT.
• Select the ••• more options menu beside the conversation in chat history.
• Select Archive.
• Select Settings when you are ready to retrieve the report.
• Select Data controls.
• Select Archived chats to find your archived lab conversation.

Delete - I don't want to use this again

Remove the saved report when you no longer need it. The TensorFlow Playground experiments require no deletion because they did not create a billed cloud resource.

This is permanent

That finality can feel drastic. The saved hierarchy, experiment observations, and generalization conclusion disappear with the conversation.

Deleted conversations cannot be restored. The browser exercises remain available if you decide to rebuild the report.

• Switch back to your saved lab conversation in ChatGPT.
• Select the ••• more options menu beside the conversation in chat history.
• Select Delete.
• Confirm the deletion.

You should no longer see the lab conversation in your chat history. The deleted report cannot be restored.

Nice Work!

Nice Work!

Outstanding work! Your completed browser lab turns AI from a black box into a set of visible experiments. You can now explain learning, prediction, generation, and failure with evidence.

What you learned:

• Built a five-part AI hierarchy that distinguishes machine learning, neural networks, generative AI, and ChatGPT by their scope and purpose.
• Tested a nonlinear classifier that produced poor predictions. You then trained a neural network with a hidden layer and ReLU to bring both Training loss and Test loss below 0.2.
• Connected learned parameters and inference to tokens and next-token prediction. You also built a verification habit for checking fluent AI output against reliable sources.
• Secret Mission: Changed one Playground variable to test generalization. You compared both losses to diagnose underfitting or increased overfitting risk.

Ready to quiz yourself?