🔍 Read the full analysis: The Numbers Behind A Neural Network’s Handwritten 7 Prediction on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A ThorstenMeyerAI.com explainer published Nov. 8, 2026, traces how a dense neural network turns a 28-by-28 handwritten digit image into scores for digits 0 through 9. It also demonstrates, with a small XOR network, how changing one bias can alter a network’s output; the article cites a separate MIT CSAIL report about AI editing model weights, described as a blog post rather than a peer-reviewed paper.
ThorstenMeyerAI.com published an explainer on Nov. 8 showing how a neural network can turn the 784 pixel values in a 28-by-28 handwritten digit image into a prediction such as “7.” The article uses a nine-number XOR network to demonstrate that changing one bias can change the function a network computes, then gives a parameter count for a small digit reader and cites a separate report about AI editing model weights, a topic related to how neural networks work.
The explainer describes a multilayer perceptron, a feedforward network made of dense layers, similar to those discussed in an explainer on neural network numbers. Its input layer holds the image’s 784 pixel values; it does not calculate a prediction. Hidden layers apply weights and biases to their inputs, then pass the results through an activation function. The output layer returns 10 scores, one for each digit, and the highest score determines the network’s prediction in the example.
To make the arithmetic visible, the article constructs a small network for XOR, which returns one when two input bits differ and zero when they match. It uses two hidden neurons and one output neuron, with step activations. The weights and biases are set so the network produces the desired result. For input (1, 0), both hidden neurons return one, and the output neuron returns one.
The example then changes the output neuron’s bias from −2 to −1. The revised network returns one for all four possible two-bit inputs, rather than only the inputs whose bits differ. The article says a small digit reader has 109,386 parameters, about 92% in its first layer, where each pixel connects to each of 128 units. These figures describe the example in the source, not a universal network design.
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 Bias Can Change the Output
The demonstration makes a basic point about neural networks concrete: their behavior is implemented through weights, biases and layer-by-layer arithmetic. In a deliberately small network, a person can trace every calculation and identify exactly how one changed value affects the answer. In a trained model, the same kind of direct interpretation is harder, the explainer says, because a behavior may depend on patterns distributed across many parameters and layers.
That distinction matters as researchers explore whether AI systems can inspect or edit other models. The source cites MIT CSAIL researchers who reportedly had an AI model read and modify another network’s 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%. Those numbers are attributed to the source’s account of that work; the post labels the material a blog post with public logs, not a peer-reviewed paper. The XOR example helps explain the underlying idea, but does not establish how reliable or general such model editing is.
Products Worth Considering
Samsung Odyssey G9 49-Inch Curved Dual QHD Gaming Monitor
Amazon Kindle Paperwhite (12th Generation, 2024), 16 GB, 7-inch Display, Black
LG UltraGear 32GX850A-B 32-inch 4K UHD OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
From Pixel Values to Layer Outputs
The source frames a handwritten digit as a list of numbers: a 28-by-28 image contains 784 pixel values when flattened into a vector. In a dense layer, each output unit combines the input values with its own weights, adds a bias and applies an activation. In compact notation, the article gives the calculation as z = W x + b, followed by an activation that produces the layer’s output. That output becomes the next layer’s input.
The article distinguishes the network’s design from the values learned during training. Width and depth describe how many units and layers the architecture has; these are choices made before training. Weights and biases are parameters that training normally learns, though they can also be assigned directly, as in the hand-built XOR example. The source notes that more width or depth is not automatically better: choices depend on the task, data and architecture.
“A neural network is arithmetic in layers.”
— Thorsten Meyer, ThorstenMeyerAI.com
Limits of the Reported Weight Edit
The source does not identify the specific digit-reading model, its training data, evaluation procedure or performance on a test set. The 109,386-parameter count is given for a small example, but its complete architecture is not specified in the supplied material. The article’s account of the MIT CSAIL work also does not provide enough detail here to independently assess the reported math-score change, including the benchmark, test conditions or how repeatable the result was. The source explicitly describes that account as a blog post with public logs, not a peer-reviewed paper. The XOR demonstration shows what a parameter change can do in a toy network; it does not show that changing one parameter will predictably produce a chosen behavior in a large trained model.
Reading the Model-Editing Evidence
The source provides a video accompanying the explainer, described as a roughly seven-and-a-half-minute hand-drawn walkthrough from a handwritten 7 to the bias change in the XOR example. It does not announce a new experiment or a formal publication date for the cited MIT CSAIL work. Readers assessing the weight-editing claim would need the underlying logs, benchmark details and a peer-reviewed account or independent replication to judge the result beyond the figures reported in the blog post. The explainer’s immediate next step is conceptual: follow the arithmetic from input pixels through intermediate layers to the output scores.
Key Questions
How does the network decide that the image is a 7?
In the example, the image is represented by 784 pixel values. The network processes them through layers and returns 10 output scores, one for each digit. The digit with the highest score is the prediction.
What does the XOR example show?
It shows that a small network can compute XOR using two hidden neurons and one output neuron. Changing the output bias from −2 to −1 makes the network return one for all four two-bit inputs instead of only when the bits differ.
What are weights and biases?
Weights determine how much each input contributes to a neuron’s calculation. A bias shifts the neuron’s response threshold. Together with an activation function, they determine the value passed to the next layer.
Is the cited AI weight-editing result peer-reviewed?
Not according to the supplied source, which describes the account as a blog post with public logs, not a peer-reviewed paper. Its reported score change and editing details should be read with that qualification.
Source: ThorstenMeyerAI.com
