🔍 Read the full analysis: Follow One Handwritten 7 From Input To AI Answer 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 November 8 explainer from ThorstenMeyerAI.com traces how a neural network turns pixel values into a handwritten-digit prediction. It uses a nine-number XOR network to show how one bias can change behavior, then describes a 109,386-parameter digit reader and a separate, unreviewed MIT CSAIL report about AI editing model weights.
ThorstenMeyerAI.com published an explainer on November 8 showing how a neural network can turn the pixel values in a handwritten image into a digit prediction. The article follows the arithmetic through a small, constructed network, then gives the scale of a digit-reading example: 109,386 learned numbers, or parameters.
The example begins with a 28-by-28 image, represented as 784 pixel values. An input layer holds those values and passes them along. In each dense hidden layer, neurons multiply inputs by weights, add a bias and apply an activation function. The next layer takes the resulting values as its input. An output layer can produce ten scores, one for each digit; the greatest score is the network’s prediction.
To make the process easier to follow, the explainer constructs a network for XOR, a function that returns one when two input bits differ and zero when they match. The network has two input values, two hidden neurons and one output neuron. With step activations, its weights and biases total nine numbers. The article works through the input (1, 0): both hidden neurons return one, and the output neuron also returns one.
Then it changes just the output bias, from −2 to −1. The same network now returns one for all four possible input pairs, rather than only the pairs with differing bits. This is a deliberately designed teaching example, not a claim that a single parameter usually controls a particular behavior in a trained AI model. The article says behavior in trained systems can depend on patterns distributed across many parameters and layers.
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 walkthrough gives readers a concrete way to understand what it means to say that a model’s behavior is encoded in its parameters. A weight determines how strongly an input contributes to a neuron’s total; a bias shifts the point at which that neuron responds. In the XOR example, changing one bias alters the function the entire network computes. The arithmetic is visible, even when following every operation in a much larger model is difficult.
The explainer’s scale comparison helps explain that difficulty. Its digit-reading network has 109,386 parameters, with about 92% in the first layer because each pixel connects to each of its 128 units. That count is specific to the example, not a universal size for digit recognition. It illustrates why tracing an input through a working model involves far more calculations than inspecting the nine numbers in the toy network.
The article also connects this idea to a separate report about AI systems editing other networks’ parameters. That connection matters because direct parameter edits are one possible way to change model behavior. But the reported result, as described in the source, is a blog-post account with public logs—not a peer-reviewed finding—and should be read with that limitation in mind.
Products Worth Considering
Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Eucalyptus Green
Samsung Odyssey G93SC 49-Inch QD-OLED Curved Gaming Monitor
Amazon Kindle Paperwhite 16GB, 7-Inch Display, Graphite
As an affiliate, we earn on qualifying purchases.
From XOR to Digit Recognition
The explainer focuses on a multilayer perceptron, a feed-forward network made from dense layers. In this architecture, each unit in a layer connects to the inputs from the previous layer. The article distinguishes architectural choices—such as layer width and depth—from parameters. Width and depth are selected when designing the network; training typically learns the weights and biases, though they can also be assigned directly.
The XOR example is useful because it shows why a simple, single straight-line decision rule is not enough for that function. Two hidden neurons can represent the conditions “at least one bit is on” and “at most one bit is on”; both are true exactly when one bit is on. The example’s hand-set weights make that logic inspectable. A trained image model, by contrast, learns its parameters from data, and the source does not provide a full account of the training process or evaluation for its digit-reader count.
The source also describes a MIT CSAIL report in which an AI model reportedly read and edited another network’s numbers. 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 figures are claims reported by the source; it characterizes the material as a blog post with public logs, not a peer-reviewed paper. They are separate from the handwritten-seven walkthrough.
“A neural network is arithmetic in layers.”
— ThorstenMeyerAI.com explainer
Limits of the Reported Examples
The source does not identify the precise dataset, training setup or evaluation details for the 109,386-parameter digit reader, so that figure should be understood as the size of the example described, not a benchmark for all digit-recognition systems. It also does not provide enough information here to independently assess the model’s accuracy on handwritten sevens or other digits.
The MIT CSAIL account is described as a blog post with public logs, rather than a peer-reviewed paper. The source gives a reported change in a math score after an edit affecting 0.016% of weights, but does not specify the benchmark, evaluation conditions or other details needed to interpret the result fully. The XOR demonstration is a hand-designed network; it does not establish that one parameter in a trained language model will have a similarly isolated or predictable effect.
What Readers Can Check Next
The source links to a video of about seven and a half minutes that presents the same progression, from a handwritten seven to the XOR bias change. It says the narration uses an AI voice clone of Thorsten Meyer. Readers seeking to assess the separate weight-editing claim can consult the public logs referenced by the source, while keeping in mind that logs do not replace independent peer review.
No further research milestone or follow-up date is given in the source. For now, the explainer’s key contribution is pedagogical: it makes a small network’s calculations explicit and distinguishes that demonstration from the harder task of interpreting parameter changes in larger, trained models.
Key Questions
How does the network identify a handwritten 7?
It receives the image as 784 pixel values for a 28-by-28 picture. Hidden layers transform those values using weights, biases and activation functions; the output layer produces scores for possible digits, and the highest score is the prediction.
What does the XOR example show?
It shows how a small network can compute XOR and how changing one bias can change its output for every input pair. The network is constructed for explanation; it is not a description of how a real digit model is trained.
What does 109,386 refer to?
It is the number of parameters in the digit-reading example described by ThorstenMeyerAI.com. The source says about 92% are in the first layer; it does not present the count as a universal standard.
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. It reports a score change, but the information provided does not fully specify the benchmark or evaluation conditions.
Source: ThorstenMeyerAI.com
