🔍 Read the full analysis: What Does A Neural Network Do With A Handwritten 7? on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A neural network classifies a handwritten digit by transforming pixel values through layers of weighted arithmetic and selecting the highest output score. A ThorstenMeyerAI.com explainer uses a nine-number XOR network to show how changing one bias can alter behavior, then describes the much larger parameter count in a simple digit reader.
A ThorstenMeyerAI.com explainer published November 8, 2026, shows how a neural network can turn the pixel values in a handwritten digit image into a prediction such as “7.” The account breaks the process into layer-by-layer arithmetic, using a tiny network to demonstrate how changing one bias can change the function a network computes.
In the example, a 28-by-28 image is flattened into 784 pixel values. Those numbers enter a multilayer perceptron, or MLP, a feedforward network with dense layers. The input layer holds and passes on the values; it does not perform the transformations described in the explainer. Hidden layers transform them, and an output layer can return 10 scores, one for each digit. The prediction is the digit with the highest score.
Each neuron in a dense layer multiplies its inputs by weights, adds a bias, then applies an activation function. The explainer represents a layer’s calculation as z = W x + b, followed by an activation that produces the values passed to the next layer. Weights determine how much inputs count, while a bias shifts when a neuron responds. The arithmetic is visible in principle; the difficulty is the quantity of calculations and parameters.
To make the effect concrete, the article builds an XOR network with two input bits and nine numbers across its weights and biases. XOR returns one when the bits differ and zero when they match. The example’s two hidden neurons check whether at least one bit is on and whether at most one bit is on. Its output returns one when both checks are satisfied. Changing the output bias from −2 to −1 makes the network return one for all four possible inputs, the explainer says.
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 Changes the Output
The small example makes a broader point: a neural network’s answer comes from its learned or assigned parameters, not from a written rule that says what a seven looks like. Altering a parameter can change a network’s behavior. In the XOR demonstration, that effect is easy to trace because the whole model has only nine numbers. In a trained image or language model, a behavior can depend on patterns distributed across many parameters and layers, making the same kind of intervention harder to interpret.
The explainer estimates that its digit-reader example has 109,386 parameters, with about 92% in the first layer. That concentration follows from the design described: each pixel connects to each of the layer’s 128 units. The figures help explain why even a small digit classifier is not comparable to the hand-built XOR example in scale. They also show why researchers’ ability to inspect and edit model weights is a subject of interest, though a particular edit does not by itself explain all of a model’s behavior.
The source also describes a report from MIT CSAIL researchers about an AI model that read and edited other networks’ numbers. According to the source, one edit altered 0.016% of a language model’s weights and raised a math score from 7.9% to 48.1%. The source characterizes this as a blog post with public logs, not a peer-reviewed paper. Those results should be understood as a reported demonstration, not as independently established findings in this material.
Popular right now
Breville Barista Express Espresso Machine with Built-In Grinder, Brushed Stainless Steel
Ninja Luxe Café Mini Espresso and Drip Coffee Machine, Stainless Steel, ES301
Dyson Supersonic Nural Hair Dryer
As an affiliate, we earn on qualifying purchases.
From Pixel Inputs to Learned Weights
The article focuses on a multilayer perceptron, a useful way to explain dense, feedforward computations. A flattened digit image is a list rather than a grid in this example. Each layer receives numbers, applies its weights and biases, and passes its output onward. The output layer’s scores provide the basis for choosing a digit. This is a simplified account of one kind of network, not a description of every image-recognition system.
The XOR example is a textbook demonstration of why a single straight-line decision rule cannot represent every pattern in the task. The explainer uses two hidden neurons to represent the conditions “at least one bit is on” and “at most one bit is on.” Together, those conditions identify inputs where the bits differ. The hand-set weights show how a network can be programmed directly, while the article notes that training normally learns weights and biases. Width and depth—the number of units and layers—are architectural choices set before training, rather than parameters learned in the same way.
The source includes a video described as a hand-drawn explainer, about seven and a half minutes long, covering the handwritten digit example and the XOR bias change. It says the narration uses an AI voice clone of Thorsten Meyer. The article’s stated aim is to connect these simple calculations to the idea of one AI model changing another model’s weights, while distinguishing the small demonstration from the larger and less transparent task.
“A neural network is arithmetic in layers.”
— ThorstenMeyerAI.com explainer
Limits of the Demonstration
The source does not identify the specific architecture, training setup, dataset, or evaluation conditions behind the 109,386-parameter digit reader. Its figure that about 92% of parameters sit in the first layer is presented for that example and should not be treated as a general ratio for neural networks.
Details of the reported MIT CSAIL result are also limited in the supplied material. It does not name the language model, math benchmark, number of tests, or evaluation protocol behind the change from 7.9% to 48.1%. The source says public logs are available and that the work was presented in a blog post rather than a peer-reviewed paper. The figures are therefore attributed to that report here; the material does not establish how broadly the result applies or whether it has been independently replicated.
The XOR example establishes that changing its output bias changes its outputs for all four inputs. It does not show that editing a single parameter in a trained digit-recognition model would reliably change only one particular prediction. The source says behaviors in trained networks may be distributed across many parameters, and offers no specific test of a handwritten seven under a weight edit.
What to Check in Follow-Up Reports
The immediate next step for readers is to follow the explainer’s video and inspect the arithmetic in the XOR example: its output can be checked for each of the four input pairs. For the reported weight-editing work, a fuller account would need to specify the model, benchmark, evaluation method, and the public logs the source says accompany the blog post.
Any follow-up claims about the MIT CSAIL demonstration should also be checked for later technical documentation, peer review, or independent replication. The supplied source gives no publication schedule or next research milestone, so it is not clear when further evidence will appear. Until then, the digit example is best read as an explanation of how layered calculations work, while the weight-editing figures remain claims attributed to the cited blog report.
Key Questions
How does a neural network recognize a handwritten 7?
In the example, the image is converted into 784 pixel values. Layers transform those values using weights, biases, and activation functions. The output layer produces scores for the possible digits, and the highest score is the prediction.
Does the network contain a rule describing what a seven looks like?
The explainer says no explicit rule is written for the shape of a seven. Instead, the network’s weights and biases determine how inputs are transformed into output scores.
What does changing a bias do in the XOR example?
Changing the output bias from −2 to −1 makes the illustrated network return one for all four possible input pairs, rather than returning one only when the bits differ.
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. It reports a math score change from 7.9% to 48.1% after an edit to 0.016% of a language model’s weights, but the supplied material does not provide the evaluation details needed to assess the result independently.
Source: ThorstenMeyerAI.com
