🔍 Read the full analysis: Inside The Calculation: AI Recognizes A Handwritten 7 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
ThorstenMeyerAI.com published an explainer on November 8, 2026, tracing how a neural network can classify a handwritten seven through layered arithmetic. It uses a nine-number XOR network to show how one bias can change a function, then describes a 109,386-parameter digit reader and reports a separate, unreviewed MIT CSAIL claim about AI editing model weights.
ThorstenMeyerAI.com published an explainer on November 8 showing how a neural network can turn pixel values from a handwritten seven into a digit prediction, using a small arithmetic example to make the process visible. The piece explains that the network’s answer comes from calculations across layers, not from a handwritten rule defining what a seven looks like, and says a sample digit reader contains 109,386 parameters.
The walkthrough describes a multilayer perceptron, or MLP, in which an image is represented as a list of numbers. A 28-by-28 image has 784 pixel values when flattened into a vector. The input layer holds and passes on those values; hidden layers transform them; and an output layer can produce ten scores, one for each digit. The highest score is the network’s prediction.
Each neuron in a dense layer multiplies inputs by weights, adds a bias, and applies an activation function. The article summarizes that operation as z = W x + b, followed by an activation that produces the next layer’s output. In the described digit-reading model, the author counts 109,386 parameters, with about 92% in the first layer because each pixel connects to each of its 128 units.
To demonstrate how parameters affect behavior, the explainer builds a small network for XOR, a function that returns one when two input bits differ. Its three neurons use nine numbers in total. Changing the output bias from −2 to −1 makes the network return one for all four possible inputs, rather than only the two cases where the bits differ. This is an illustrative example chosen and calculated by the author, not a result from a trained digit-recognition system.
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
From Pixels to Digit Scores
The walkthrough makes a technical process easier to inspect: pixel values pass through successive numerical transformations, and the final output represents competing digit scores. That distinction matters because a model does not need a human-written checklist of the visual features of a seven to produce a classification. It can learn parameter values during training that make its outputs useful on examples.
The XOR demonstration also shows both the power and the limits of focusing on a single number. In the deliberately small network, one bias change flips the function in a predictable way. In a trained model, however, a behavior may depend on patterns distributed across many weights and layers. The example clarifies what a parameter edit is, but does not establish that changing one parameter in a large model will have a similarly simple or isolated effect.
The source also points to reported work in which an AI model edited another model’s weights. If that report is borne out, direct weight editing could offer a way to study or change model behavior. The supplied account, however, describes the work as a blog post with public logs rather than a peer-reviewed paper, so its claims should be treated as reported results rather than settled evidence.
Products Worth Considering
Samsung Odyssey G93SC 49-Inch QD-OLED Curved Gaming Monitor
Amazon Kindle Paperwhite (12th Generation, 2024), 16 GB, 7-inch Display, Black
Samsung Odyssey G9 49-Inch Curved Dual QHD Gaming Monitor
As an affiliate, we earn on qualifying purchases.
A Small Network as a Demonstration
The article begins with a common classification task: provide a network with an image of a handwritten digit and read its prediction. It then reduces the problem to a two-bit example so readers can follow every operation. XOR is useful here because a single straight-line decision rule cannot separate its outputs; the explainer uses two hidden neurons and one output neuron to represent the function.
The author distinguishes the network’s structure from the values learned or assigned within it. Width and depth describe how many units and layers an architecture has, while weights and biases are parameters. The article says architecture choices are made before training and that training normally learns parameter values, though they can also be assigned directly in a hand-built example.
As additional context, the source says researchers have visualized early image-model layers responding to simple patterns such as edges and later layers responding to object parts. It presents this as background on how layers can combine information, not as a specific finding from the digit model described in the walkthrough.
“A neural network is arithmetic in layers.”
— Thorsten Meyer, ThorstenMeyerAI.com
Limits of the Reported Results
The supplied material does not identify the handwritten-digit dataset, the model’s training procedure, its accuracy, or the conditions under which the 109,386 parameters were counted. The example explains how a network can process pixel values, but it does not report a new experiment or independently establish how well a particular system recognizes handwritten sevens.
Details about the cited MIT CSAIL report are also limited. The source says an edit changed 0.016% of a language model’s weights and a math score rose from 7.9% to 48.1%, but it does not name the model, benchmark, evaluation method, or comparison conditions in the supplied account. It characterizes the material as a blog post with public logs, not peer-reviewed research. The results and their reproducibility therefore cannot be assessed from this material alone.
The XOR network’s bias change is fully specified for its small, hand-designed example. It should not be taken as evidence that an individual parameter edit in a larger trained network will produce an equally predictable change. The source itself says that large-model behavior can be spread across many parameters.
What Further Evidence Would Show
The published item is an educational walkthrough, and the supplied source gives no announced follow-up date or next research milestone. Readers can use its arithmetic example to understand the mechanics, but assessing the reported weight-editing result would require the full logs and enough information to identify the model, benchmark, evaluation conditions, and method used to measure the change.
For the digit-recognition example, useful next details would include the dataset, the model architecture and training setup, and measured performance on images not used for training. Those details would show how the illustrated calculations relate to a tested classifier rather than a conceptual demonstration. Until such information is supplied, the confirmed development is the publication of the explainer and its worked examples, while the broader claims about model editing remain attributed to the source.
Key Questions
How does the network turn a handwritten seven into a prediction?
The image can be represented as 784 pixel values for a 28-by-28 image. The network transforms those numbers through hidden layers and produces output scores for possible digits; the highest score is the prediction.
What are weights and biases?
Weights determine how much input values contribute to a neuron’s total. A bias shifts that total before an activation function determines the neuron’s output.
What does the XOR example show?
It shows that in a deliberately small network, changing one output bias from −2 to −1 changes the function from XOR to returning one for every input. It is a hand-built demonstration, not a reported edit to a digit-recognition model.
Is the reported AI weight-editing result peer-reviewed?
The source describes it as a blog post with public logs, not a peer-reviewed paper. The supplied material does not give enough benchmark or evaluation details to independently assess the reported score change.
Source: ThorstenMeyerAI.com
