🔍 Read the full analysis: A Step-by-Step Look At AI Reading A Handwritten 7 on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
A ThorstenMeyerAI.com explainer published Nov. 8, 2026, traces how a dense neural network can turn the pixel values in a handwritten digit image into a prediction. It uses a nine-number XOR network to show how one bias change can flip a function, then discusses a separate, preliminary report about AI editing other models’ weights.
ThorstenMeyerAI.com published an explainer Nov. 8 tracing how a neural network can classify an image of a handwritten seven, showing how pixel values move through layers of arithmetic before the model returns a digit prediction. The piece uses a tiny, hand-constructed network to demonstrate how changing one number can alter behavior, then compares that example with a larger digit reader described as having 109,386 parameters.
The example begins with a 28-by-28 image, represented as a list of 784 pixel values. In the multilayer perceptron described in the article, the input layer holds and passes those values onward; it does not itself calculate a prediction. Hidden layers multiply inputs by learned weights, add biases, and apply an activation function. Their resulting values become inputs to the next layer.
At the end, an output layer produces 10 scores, one for each possible digit. The network’s prediction is the digit with the highest score. This describes the flow of computation, not a complete account of how the particular image was acquired, how the model was trained, or how accurate it is on a broader set of handwriting.
To make the arithmetic easier to inspect, the explainer constructs a network for XOR, a function that returns one when two bits differ. Its two hidden neurons identify whether at least one bit is on and whether at most one bit is on; the output combines those conditions. The network uses nine numbers. Changing only its output bias from −2 to −1 makes it return one for all four inputs, rather than only when the bits differ.
Infographic 01 · One bias, two behaviors
A network you can run by hand
Two hidden neurons check "at least one bit is on" and "at most one bit is on". The output neuron needs both. Change its bias and it needs only one.
Move the output bias from −2 to −1 and the network stops computing XOR: it returns 1 for all four inputs.
The whole network
Hidden 1: step( x₁ + x₂ − 1 ) Hidden 2: step(−x₁ − x₂ + 1 ) Output: step( h₁ + h₂ − 2 )
step: 1 if the total is zero or greater, otherwise 0. Nine numbers in total: six in the hidden layer, three in the output neuron.
All four inputs
Hidden outputs and the final output with the original and the changed bias
| Input | At least one | At most one | Output, bias −2 | Output, bias −1 |
|---|---|---|---|---|
| 0, 0 | 0 | 1 | 0 | 1 |
| 0, 1 | 1 | 1 | 1 | 1 |
| 1, 0 | 1 | 1 | 1 | 1 |
| 1, 1 | 1 | 0 | 0 | 1 |
Sources: own arithmetic, checked with a script · the XOR example as in Goodfellow, Bengio, Courville: Deep Learning, ch. 6.
Powered by Thorsten Meyer AI
Why One Weight Can Change a Prediction
The worked example offers a concrete way to understand what it means for a model to compute through layers: its behavior follows from numerical operations, even when the full network contains too many operations to inspect casually. In the XOR case, a single bias change has a clear, visible consequence. In a trained image classifier, the article says, behavior is usually distributed across many weights and layers, so tracing one adjustment is less straightforward.
The scale figures help explain that gap. The article says a small digit reader contains 109,386 parameters, with about 92% in its first layer, where each pixel connects to each of 128 units. These figures are supplied by the source article; they are not an independent measurement here. The contrast between nine deliberately chosen numbers and a much larger parameter set shows why directly editing trained networks is a different task from designing a toy example.
The source also reports that MIT CSAIL researchers had an AI model read and edit another network’s numerical weights. It says one edit changed 0.016% of a language model’s weights and raised a math score from 7.9% to 48.1%. That is a claim reported by the source, not a result established by the handwritten-seven demonstration. The source characterizes the material as a blog post with public logs, not a peer-reviewed paper, so the figures should be read with that limitation in mind.
Products Worth Considering
Samsung Odyssey G7 40-Inch WUHD Curved Gaming Monitor (G75F)
Bose SoundLink Plus Portable Bluetooth Speaker, Black
Jackery Explorer 300 Portable Power Station with 40W Solar Panel
As an affiliate, we earn on qualifying purchases.
From Pixels to Layer Outputs
A dense layer can be represented as z = W x + b, followed by an activation that produces the layer’s output. Here, x is the incoming list of values, W holds the weights, and b contains the biases. The intermediate totals z are transformed by the activation; those outputs then feed the following layer. “Hidden” refers to these intermediate values, not to a step that cannot be examined.
The distinction between model design and training also matters. The source describes width and depth as architectural choices: how many units are in a layer and how many layers the network has. Weights and biases are parameters that training ordinarily learns, though someone can assign them directly in a simple constructed example. For XOR, the numbers are selected by reasoning about the desired output. A digit classifier’s parameters, by contrast, are numerous and their learned behavior is not presented as a single hand-written rule.
The explainer is accompanied by a video of about seven and a half minutes, described as a hand-drawn account of the path from a handwritten seven to the XOR bias change. The source says its narration uses an AI voice clone of Thorsten Meyer. That description concerns the video’s narration and does not establish anything about the digit model’s design or performance.
“A neural network is arithmetic in layers.”
— ThorstenMeyerAI.com explainer
Limits of the Reported Results
The source does not identify the handwritten-seven image, provide the digit reader’s training details, or report its accuracy, so the example explains a general computation rather than documenting a measured test result. It also does not provide enough information here to independently evaluate how the 109,386-parameter count was obtained or whether the described architecture is representative of other digit classifiers.
The account of the MIT CSAIL work is also limited. The source reports a weight-editing result and links to public logs, while describing the work as a blog post rather than a peer-reviewed paper. The model, evaluation setup, math benchmark, and comparison conditions are not detailed in the supplied material. The reported score change should therefore remain attributed to that account, not treated as independently verified evidence of general performance.
What Readers Can Check Next
The source points readers to a video that walks through the layer calculations and the XOR example. For the separate weight-editing claim, the next useful step would be to examine the public logs and any underlying technical report or evaluation details, if available. The supplied material gives no publication schedule or further milestone for that work, so whether a peer-reviewed study or additional results will follow is not clear.
Key Questions
How does a neural network turn a handwritten seven into a prediction?
In the example, the image is represented by 784 pixel values. Layers transform those values using weights, biases and activation functions. The final layer supplies 10 digit scores, and the highest score determines the prediction.
Does the input layer recognize the digit?
No. The source describes the input layer as holding and passing along the pixel values. The calculations that transform them take place in later layers.
What does the XOR example show?
It shows how a small network can implement a function by combining hidden-neuron outputs. In the source’s nine-number example, changing the output bias from −2 to −1 changes the result for all four possible two-bit inputs.
Is the reported AI weight-editing result peer-reviewed?
The source describes the MIT CSAIL material as a blog post with public logs, not a peer-reviewed paper. Its reported score change is therefore attributed to that account, and the supplied material does not include enough evaluation detail for independent assessment.
Source: ThorstenMeyerAI.com
