🔍 Read the full analysis: See How A Neural Network Turns A Handwritten 7 Into A Prediction on ThorstenMeyerAI.com
Get the little things that make your day 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 neural network turns the pixels of a handwritten digit into a prediction. It uses a nine-number XOR network to show how changing one bias can flip behavior, then describes a digit-reading network with 109,386 parameters.
ThorstenMeyerAI.com published an explainer on Nov. 8 showing how a neural network can turn a picture of a handwritten seven into a digit prediction by passing pixel values through layers of arithmetic. The article uses a small network that classifies two input bits to demonstrate how changing a single bias can change a network’s output, before outlining the much larger number of parameters in a digit reader.
For the example, a handwritten image measuring 28 by 28 pixels is represented as a list of 784 values. Those values enter the network; an input layer holds and passes them on but does not itself perform a computation. In the dense, feedforward network described by Meyer, each hidden-layer neuron multiplies inputs by weights, adds a bias and applies an activation function. Its outputs become the next layer’s inputs.
The output layer can produce 10 scores, one for each digit. The prediction is the digit with the highest score. The explainer’s main demonstration is a separate toy example: a network with two input bits, two hidden neurons and one output neuron. It implements XOR, which returns one when the bits differ and zero when they match. Meyer gives the network’s weights and biases and walks through the arithmetic for an input of (1, 0).
In that toy network, changing the output bias from −2 to −1 makes the output one for all four possible bit combinations, according to the explainer. The network has nine numerical parameters in total. Meyer contrasts it with a small digit reader that has 109,386 parameters, about 92% of them in its first layer, where each pixel connects to each of 128 units.
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 Individual Weights Matter
The examples make a technical distinction concrete: neural networks are built from explicit arithmetic operations, but their behavior can depend on many learned weights and biases working together. A one-number change produces an easily traced result in the small XOR demonstration. In a trained digit reader, identifying which parameters matter for a particular prediction is more complicated because behavior may be distributed across layers.
The source also connects this question to weight-space programming—editing a model’s numerical parameters directly rather than changing its input or writing a conventional instruction. Meyer says MIT CSAIL researchers reported that an AI model read and edited other networks’ numbers. The article cites one edit that changed 0.016% of a language model’s weights and raised a math score from 7.9% to 48.1%. Those figures are claims reported in the source article, which describes the material as a blog post with public logs, not a peer-reviewed paper. They are separate from the handwritten-digit demonstration.
Products Worth Considering
Philips Hue White and Color Ambiance A19 Smart Bulb Starter Kit, 2-Pack
eufy Security 5-Piece Home Alarm Kit
Samsung 75-Inch QLED Q7F Series Smart TV (2025)
As an affiliate, we earn on qualifying purchases.
From Pixels to Layered Arithmetic
A multilayer perceptron, or MLP, is a feedforward network made of dense layers. In a dense layer, each output unit connects to every input value. The layer can be written as z = W x + b, followed by an activation: W holds the weights, x the input, and b the biases. The activation turns the intermediate totals into the values passed forward.
Weights and biases are the network’s parameters, which training normally learns from data. Width, the number of units in a layer, and depth, the number of layers, are architectural choices made before training. Meyer notes that neither greater width nor greater depth is automatically better; the choices depend on the task, data and architecture. The explainer’s seven-and-a-half-minute video presents the material as a hand-drawn walkthrough, with narration described as an AI voice clone of Meyer.
“A neural network is arithmetic in layers.”
— Thorsten Meyer, ThorstenMeyerAI.com
Limits of the Demonstration
The XOR network is a deliberately small example with nine parameters; it shows that a numerical change can alter behavior, but it does not explain how to identify the relevant weights in a trained digit reader. The source does not provide a specific test image, the network’s full architecture for digit recognition, its training details or prediction confidence for a handwritten seven.
The article also does not supply the underlying MIT CSAIL post or public logs in enough detail to independently evaluate the reported language-model score change. It identifies that work as a blog post, not a peer-reviewed paper. The exact evaluation setup and how broadly the reported result applies are not established in the provided material.
What the Explainer Shows Next
The published video is presented as a roughly seven-and-a-half-minute walkthrough, moving from a handwritten seven to the XOR network and a single-bias change. The article’s next step is to contrast that hand-built example with a digit reader’s 109,386 learned parameters, illustrating why directly editing a trained network is more difficult.
No follow-up study, additional experiment or publication date is stated in the source. Readers can use the explainer’s equations to verify the toy network’s outputs, but the supplied material does not show a full set of trained weights or a step-by-step calculation for a particular seven image.
Key Questions
How does the network turn a handwritten 7 into a prediction?
The image is represented as 784 pixel values for a 28-by-28 picture. Layers transform those values using weights, biases and activation functions; the output layer returns scores for the possible digits, and the highest score is the prediction.
What does the input layer do?
In Meyer’s example, it holds and passes on the input values. The input layer does not perform the transformations attributed to the hidden and output layers.
Why does changing one bias alter the XOR result?
The output neuron uses its bias as part of the total that determines whether it activates. Changing it from −2 to −1 makes the toy network return one for all four input pairs, according to the explainer.
How many parameters does the digit reader have?
The small digit reader described in the source has 109,386 parameters. About 92% are in its first layer; the source says each pixel connects to each of that layer’s 128 units.
Does the example explain every prediction made by a trained network?
No. The XOR example makes a small network’s behavior easy to calculate by hand. The explainer says a trained network’s behavior is harder to trace because it can depend on patterns distributed across many parameters and layers.
Source: ThorstenMeyerAI.com
