🔍 Read the full analysis: A Handwritten 7 Enters A Neural Network—How Does It Get An Answer? on ThorstenMeyerAI.com
Get smart everyday buys 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 November 8, 2026, explainer from ThorstenMeyerAI.com traces how a dense neural network turns 784 image pixel values into scores for the digits 0 through 9. It also shows, with a nine-number XOR network, how changing one bias can alter a function; the source separately describes an MIT CSAIL report on AI editing model weights, which it says is documented in a blog post rather than a peer-reviewed paper.
ThorstenMeyerAI.com published an explainer on November 8 showing how a neural network can classify a handwritten seven, tracing the process from image pixels through arithmetic layers to a digit prediction. The article also uses a small XOR network to show that changing one bias can change the network’s function, while noting that real trained models often distribute behavior across many parameters.
In the example, a handwritten image is represented as 784 pixel values, corresponding to a 28-by-28 image flattened into a list. The input layer holds and passes along those values; it does not perform the transformations described in the article. Hidden layers process the numbers, and an output layer produces ten scores, one for each digit. The digit with the highest score becomes the prediction.
Each unit in a dense layer multiplies inputs by weights, adds a bias and applies an activation function. In the article’s notation, a layer computes z = W x + b, then applies an activation to produce the values passed to the next layer. Weights control how much inputs contribute, while biases shift when a unit responds. These steps are arithmetic, although a full model may involve many such operations.
To make the effect of a parameter visible, the explainer builds a small network for XOR, a function that returns one when two input bits differ. Its two hidden units check whether at least one bit is on and whether at most one bit is on; an output unit combines those results. The network uses nine numbers. In the given example, changing the output bias from −2 to −1 makes the network return one for all four possible inputs instead of computing XOR.
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 a Network
The example offers a concrete way to understand what it means to edit a model’s parameters: weights and biases are numbers that shape outputs, not rules written in plain language. In the small XOR network, one bias is enough to switch the behavior. That makes the relationship between an internal number and the output easy to follow.
The same lesson also points to a limit of the toy example. The source says behavior in a trained network is often spread across patterns involving many parameters and layers, making the effect of a single edit harder to predict. Understanding that distinction matters when researchers study how models work or try to change their behavior: a simple demonstration does not, by itself, explain a large model.
Products Worth Considering
FLEXISPOT Adjustable Workbench with Hand Crank, 47.2 x 23.6 Rubberwood Top
Logitech MX Keys Combo for Business Keyboard and Mouse – Graphite
LG UltraGear 27GX704A-B 27-inch QHD OLED Gaming Monitor, 240Hz
As an affiliate, we earn on qualifying purchases.
From XOR to Digit Recognition
The explainer focuses on a multilayer perceptron, a feedforward network with dense layers. In a dense layer, each output unit connects to every input value. For a digit image, that means each pixel can contribute to each unit in the next layer. The source says a small digit reader has 109,386 parameters, with about 92% in its first layer because of those connections.
The article distinguishes architectural choices from learned values. Width and depth refer to the number of units and layers, choices made when designing the model. Weights and biases are parameters that training normally learns, though the XOR demonstration assigns them directly. The source also refers to a reported MIT CSAIL example in which an AI model read and edited another network’s numbers. It says one edit changed 0.016% of a language model’s weights and its math score rose from 7.9% to 48.1%; those figures are claims reported by the source, not independent confirmation in the material provided.
“A neural network is arithmetic in layers.”
— ThorstenMeyerAI.com explainer
Limits of the Demonstration
The source material does not provide a dataset, training procedure, accuracy result or test conditions for the handwritten-seven example. It explains a general classification process, but does not establish how a particular model was trained or how reliably it recognizes digits. The 109,386-parameter count applies to the small digit reader described in the source, not to neural networks generally.
The MIT CSAIL account is also presented with a qualification: the source calls it a blog post with public logs, not a peer-reviewed paper. The supplied material does not name the model, link to the logs or give further details about the evaluation behind the reported math-score change. Those details cannot be verified from the source text alone.
Following the Numbers Further
The explainer points readers to an accompanying roughly seven-and-a-half-minute video that follows the handwritten-digit example and the XOR bias change. It describes the narration as an AI voice clone of Thorsten Meyer. The source material does not announce a later study, product release or new research milestone.
The next step for readers seeking to assess the reported model-editing result is to consult the linked public logs and the underlying MIT CSAIL account, then check what model and evaluation were used. For the digit example itself, the article’s arithmetic shows how values move through layers; details such as training data and recognition performance remain outside the material provided.
Key Questions
How does the network turn a handwritten seven into a prediction?
The image is represented as 784 pixel values. Hidden layers transform those values using weights, biases and activation functions, and the output layer returns ten scores. The highest score is the predicted digit.
What does a bias do in a neural network?
A bias shifts the point at which a unit responds. In the XOR example, changing one output bias changes the results for all four possible input pairs.
Does one changed number usually control a large model’s behavior?
Not necessarily. The explainer says behavior in trained networks is often distributed across many parameters and layers, so the effect of an individual edit can be harder to isolate than in the nine-number XOR demonstration.
What is confirmed about the reported MIT CSAIL model edit?
The source reports an edit affecting 0.016% of a language model’s weights and a math score change from 7.9% to 48.1%. It describes the material as a blog post with public logs, not a peer-reviewed paper; the supplied text does not provide enough information to independently assess the result.
Source: ThorstenMeyerAI.com
