🔍 Read the full analysis: How AI Processes A Handwritten 7 Before Giving Its Answer on ThorstenMeyerAI.com
Get audio and creator gear 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, 2026, explainer from ThorstenMeyerAI.com traces how a neural network can classify a handwritten seven using arithmetic across layers. It also shows, with a tiny XOR example, how changing one bias can change a network’s behavior, and reports a separate MIT CSAIL demonstration of AI editing other models’ weights.
ThorstenMeyerAI.com published an explainer November 8 showing how a neural network turns an image of a handwritten seven into a prediction: it represents the image as pixel values, processes those numbers through layers of weighted arithmetic, then scores possible digits. The piece uses a small network whose calculations can be followed by hand to show how even one changed bias can alter a model’s behavior.
For the example, a 28-by-28 image is flattened into a list of 784 pixel values. The input layer holds and passes those values along; it does not itself perform the transformations. In a dense layer, each neuron multiplies its inputs by weights, adds a bias and applies an activation function. The resulting values become inputs to the next layer.
The explainer describes the calculation as z = W x + b, followed by an activation. Here, the weights determine how much each input contributes, while the bias shifts the neuron’s response. At the output, a digit classifier can produce 10 scores, one for each possible digit; the largest score is used as the prediction. The source does not provide a specific trained model’s scores for a handwritten seven.
To illustrate how a parameter can matter, the article constructs a small network for XOR, a function that returns one when two input bits differ. The network has nine numbers in its weights and biases. Changing its output bias from minus two to minus one makes it return one for all four possible inputs, rather than only the two cases where the bits differ. The example is hand-designed, not a claim that a single parameter typically controls a complex trained model.
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 Parameter Edits Matter
The examples make a technical point relevant to understanding AI decisions: a model’s answer is produced by a sequence of computations, not by a human-written rule for every possible image. A handwritten seven is represented through numerical inputs and transformed into scores. Following those operations helps explain what a prediction means—and what it does not mean: the top-scoring class is the network’s selected answer, not by itself a guarantee that the classification is correct.
The source also connects the simple demonstration to weight-space programming, in which a model or person directly changes network parameters rather than only supplying a prompt or training examples. It says MIT CSAIL researchers reported an AI model reading and editing the numbers in other networks. According to the source, one reported edit changed 0.016% of a language model’s weights and raised a math score from 7.9% to 48.1%. That result is presented as a claim from a blog post with public logs, not as a peer-reviewed finding.
Products Worth Considering
Sonos Play Portable Wi-Fi and Bluetooth Speaker, White
Bose QuietComfort Wireless Noise Cancelling Headphones, Twilight Blue
Apple iPad 11-inch with A16 Chip, 512GB, Wi-Fi 6, and Apple Pencil (USB-C) — Blue
As an affiliate, we earn on qualifying purchases.
From Pixel Values to Learned Weights
The explainer focuses on a multilayer perceptron, a feedforward network made of dense layers. In such a layer, each output unit can receive input from every unit in the preceding layer. “Hidden” layers are called hidden because their intermediate outputs are neither the provided input nor the final answer; the term does not mean their arithmetic cannot be inspected.
The distinction between architecture and parameters is also central. Width and depth describe how many units and layers a network has; these are choices made when designing the model. Weights and biases are parameters that training normally learns, although they can also be assigned directly. The source says the example digit reader has 109,386 parameters, with about 92% in its first layer because each pixel connects to each of 128 units. Those figures describe the specific model in the explainer, not all digit-recognition systems.
The source includes a roughly seven-and-a-half-minute video showing the path from a handwritten digit through the XOR example. It identifies the narration as an AI voice clone of Thorsten Meyer. The accompanying article is dated November 8, 2026, and links to the video; it does not present the digit-network walkthrough as a new peer-reviewed research paper.
“A neural network is arithmetic in layers.”
— ThorstenMeyerAI.com explainer
Limits of the Demonstration
The source does not identify the exact trained digit model, its training data, evaluation method or confidence scores for the handwritten seven. It therefore illustrates a general processing route rather than documenting a particular prediction with independently reported performance results.
The MIT CSAIL example also has limits in the material provided. The source says the report appeared in a blog post with public logs and was not a peer-reviewed paper; it does not give enough detail here to assess the task setup, how the math score was measured, or whether the result holds across models and tests. The XOR example shows that a single bias can flip a deliberately constructed toy network, but it does not establish that one parameter change will predictably control a large trained model.
What the Source Documents Next
The source’s immediate next step for readers is the linked video, which walks through the digit example and the hand-calculated XOR network. It does not announce a follow-up study, publication date or additional milestone for the reported parameter-editing work.
For the MIT CSAIL claim, further evidence would be needed to establish how the edit was tested and whether the reported score change is reproducible under other conditions. For the digit example, a specific model’s architecture, training and output scores would be needed to verify how it classified a particular handwritten seven.
Key Questions
How does the network process a handwritten seven?
In the example, the image is converted into 784 pixel values. Dense layers transform those values using weights, biases and activation functions, and the output layer returns scores for possible digits.
Does the input layer analyze the image?
No. The source says the input layer holds the pixel values and passes them to the next layer. The following layers perform the weighted calculations and transformations.
Can one changed number alter a network’s answer?
Yes, in the hand-designed XOR example, changing the output bias from minus two to minus one changes the output for all four possible inputs. That toy result does not show that one edit will have the same effect in a large trained model.
What is the reported MIT CSAIL result?
The source says researchers reported an AI model editing other networks’ parameters; one edit changed 0.016% of a language model’s weights and raised a math score from 7.9% to 48.1%. It describes the evidence as a blog post with public logs, not a peer-reviewed paper.
Source: ThorstenMeyerAI.com
