🔍 Read the full analysis: A Neural Network’s View Of A Handwritten 7, From Input To Output on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A November 8 explainer from ThorstenMeyerAI.com follows a handwritten seven through a neural network, showing how pixel values pass through layers of weighted calculations to produce a digit prediction. It also uses a nine-number XOR network to show how changing one bias can change a network’s output; the article then connects that example to a reported, but not peer-reviewed, MIT CSAIL project on AI editing model weights.
ThorstenMeyerAI.com published an explainer on November 8 tracing how a neural network can turn a picture of a handwritten seven into a digit prediction. The article lays out the calculations between image pixels and output scores, then uses a small network to show how changing one number can change what the network returns.
For its digit example, the explainer uses a 28-by-28 image, represented as a list of 784 pixel values. Those values enter the network as input; the input layer holds and passes them on rather than performing the transformations. In the hidden layers, neurons multiply inputs by weights, add biases and apply activation functions. The resulting values feed the next layer.
The output layer can produce ten scores, one for each digit. The explainer describes the prediction as the digit with the highest score. It presents this process as arithmetic across layers, not as a single handwritten rule that tells the network what a seven looks like.
To make the mechanics easier to inspect, the article constructs a small XOR network with two input bits, two hidden neurons and one output neuron. Its nine weights and biases are set so it returns one when the two bits differ and zero when they match. Changing the output bias from −2 to −1 makes the network return one for all four possible inputs, according to the explainer’s calculations.
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
What the Pixel Calculations Show
The walkthrough offers readers a concrete way to understand what happens inside a basic neural network: a prediction comes from successive numerical transformations, with learned weights and biases affecting how strongly neurons respond. The XOR example makes the effect of a parameter change visible because the network is small enough to calculate by hand.
The article also warns against treating the toy example as a direct map of how a trained digit reader behaves. In the XOR network, one bias flips the result across every input. In a trained model, a behavior may depend on patterns spread across many parameters and layers, making the effect of a particular edit harder to predict. That distinction matters when people discuss interpreting or editing AI models: understanding one simple calculation is not the same as explaining a large model’s behavior.
The source says a small digit reader in its example has 109,386 parameters, with about 92% in its first layer because each pixel connects to each of 128 units. Those figures describe the article’s chosen network, not all digit-recognition systems. They illustrate why even a seemingly modest image task can involve many numbers.
Products Worth Considering
FLEXISPOT ErgoX Ergonomic Office Chair with Footrest, Black
Dell UltraSharp U2724DE 27-Inch WQHD IPS Black Monitor
Brother Work Smart 1360 Wireless Color Inkjet All-in-One Printer
As an affiliate, we earn on qualifying purchases.
From XOR to Digit Recognition
The article focuses on a multilayer perceptron, or MLP: a feedforward network made from dense layers. It distinguishes architectural choices from learned values. Width, the number of units in a layer, and depth, the number of layers, are chosen before training; weights and biases are parameters that training ordinarily learns. The explainer notes that the right architecture depends on the task and data.
Its nine-number XOR example is designed to make the arithmetic legible, not to reproduce the internal operation of a modern image model. The source also points to a separate report from MIT CSAIL researchers, who said an AI model read and edited other networks’ weights. According to the source, one edit changed 0.016% of a language model’s weights and raised a math score from 7.9% to 48.1%. The source describes that work as a blog post with public logs, not a peer-reviewed paper, so the figures should be understood as reported results rather than independently established findings here.
“A neural network is arithmetic in layers.”
— ThorstenMeyerAI.com explainer
Limits of the Demonstration
The explainer does not identify a particular trained model or provide a test set, accuracy rate or benchmark for the handwritten-seven example. It describes how a network can process digit pixels in general, rather than reporting a new experiment showing that a specific system recognized a specific image.
The source also does not provide enough detail to independently evaluate the MIT CSAIL weight-editing figures, including the full evaluation setup or whether the reported score change holds across other tests. Its characterization of the material as a blog post with public logs, rather than peer-reviewed research, is a relevant limit. The source does not establish that the same kind of targeted edit would work as simply in a large image-recognition network as it does in the hand-built XOR example.
What Readers Can Follow Next
The explainer points readers to a roughly seven-and-a-half-minute video that follows the handwritten-digit example and the XOR bias change. The source says the narration uses an AI voice clone of Thorsten Meyer. It does not announce a follow-up experiment or a publication date for further research.
For the separate weight-editing claim, the next useful evidence would be the cited public logs and a detailed account of the evaluation: which model and test were used, how the score was calculated, and whether the result can be reproduced. Until those details are assessed, the reported change remains an attributed result in the source material, not a general conclusion about what AI systems can reliably do.
Key Questions
How does the network receive a picture of a handwritten seven?
In the article’s example, a 28-by-28 image is flattened into a list of 784 pixel values. The input layer passes those values to the network’s later layers.
What do weights and biases do?
Weights determine how much inputs contribute to a neuron’s total. A bias shifts that total before the activation function is applied. The resulting value is passed onward through the network.
What does the XOR example demonstrate?
It shows, in a small hand-built network, that changing one number can change the output for every input. Moving the output bias from −2 to −1 changes the described XOR network into one that returns one for all four input pairs.
Does the explainer report a new test of a digit-recognition system?
No specific model test or recognition score is reported. The piece is a mechanics explainer showing how a network can transform pixel values and produce digit scores.
Are the reported language-model editing results peer-reviewed?
The source describes the MIT CSAIL material as a blog post with public logs, not a peer-reviewed paper. It reports a score increase from 7.9% to 48.1% after an edit to 0.016% of a model’s weights, but the supplied material does not include enough evaluation detail to verify how broadly that result applies.
Source: ThorstenMeyerAI.com
