🔍 Read the full analysis: What A Neural Network Calculates When It Sees 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
A neural network recognizes a handwritten digit by passing pixel values through layers of weighted arithmetic, then comparing output scores. A November 8, 2026, explainer uses a small XOR network to show how changing one bias can alter a network’s behavior; it also describes a digit reader with 109,386 parameters. The accompanying account of AI editing other models’ weights is attributed to a blog post, not a peer-reviewed paper.
A November 8, 2026, explainer traces how a neural network can classify a handwritten seven by converting image pixels into numbers, processing them through layers of calculations and selecting the strongest output score. The account uses a small, hand-constructed network to show how changing one bias can switch the function it computes, then gives a digit-reading model’s parameter count: 109,386 numbers.
For a 28-by-28-pixel image, the network’s input is a list of 784 pixel values. The input layer holds and passes those values along; it does not itself calculate a classification. In each dense layer, neurons multiply inputs by weights, add a bias, and apply an activation function. The resulting values become the next layer’s inputs. At the output, a digit classifier can return 10 scores, one for each digit, and predict the digit with the highest score.
The explainer makes the mechanism concrete with a tiny network for XOR, a function that outputs one when two input bits differ and zero when they match. Its two hidden neurons encode conditions on the bits, and an output neuron combines their results. The network uses nine numbers in total. With an activation that returns one for a total of zero or more, the stated weights and biases produce the XOR results for all four possible input pairs.
Changing only the output bias from −2 to −1 makes the same network return one for all four inputs, according to the walkthrough. This example is designed by hand; it is not a demonstration that a real image-recognition model’s behavior can generally be changed with one edit. The article says about 92% of the digit model’s parameters are in its first layer, where each input pixel connects to each of 128 units.
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 Model Weights Matter
The walkthrough offers a way to understand what a classifier is doing without treating its decision as magic: it is a sequence of numerical operations, even if the full calculation is too large to follow by hand. That distinction matters when people ask how a model turns an image into a label. The arithmetic is inspectable in principle, while interpreting the role of many interacting parameters can remain difficult.
The article also links this idea to weight-space programming: changing a model’s parameters directly to alter its behavior. It says MIT CSAIL researchers reported that an AI model read and edited the numbers in other networks. In one example, the post says an edit to 0.016% of a language model’s weights raised a math score from 7.9% to 48.1%. Those figures are claims reported by the source, not independently established findings in the supplied material. The source describes its account as a blog post with public logs, not a peer-reviewed paper.
If such methods prove reliable beyond demonstrations, they could offer another way to inspect or modify models. But the XOR example only establishes that a small, deliberately designed network can be understood and altered through its arithmetic. It does not by itself show that edits to large models are predictable, safe, or broadly useful.
Popular right now
Ninja Luxe Café Premier 3-in-1 Espresso Machine, Drip Coffee Maker, and Cold Brew Maker
De'Longhi Classic Signature Espresso Machine with Adjustable Milk Frother
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
As an affiliate, we earn on qualifying purchases.
From Pixel Inputs to Weights
The source focuses on a multilayer perceptron, or MLP: a feedforward network made from dense layers. A dense layer connects each of its units to every value from the preceding layer. Its calculation can be written as z = W x + b, followed by an activation: x is the input vector, W the weight matrix, and b the bias vector. The activation converts intermediate totals into outputs that move forward.
In this description, “hidden” refers to layers between the provided input and the result, not to inaccessible operations. Weights and biases are parameters learned during training or assigned directly; width and depth describe the architecture and are chosen before training. The source notes that early layers in visual models have been found to respond to simple patterns such as edges, while later layers can respond to more complex features. The digit example’s 109,386 parameters illustrate how quickly the number of values grows beyond the nine-number XOR demonstration.
The article was published by ThorstenMeyerAI.com and includes a roughly seven-and-a-half-minute explanatory video. The source identifies the narration as an AI voice clone of Thorsten Meyer. Its central material is an educational walkthrough rather than a report of a newly released digit-recognition system.
“A neural network is arithmetic in layers.”
— Thorsten Meyer, ThorstenMeyerAI.com
Limits of the Demonstration
The supplied material does not identify the exact digit model’s dataset, training process, test performance, or the conditions under which it classifies a handwritten seven. The 109,386-parameter count is given for a small digit reader, but the architecture and evaluation details are not fully specified in the excerpt.
The reported weight-editing result also lacks details here about the language model, the math benchmark, the evaluation setup, and whether the score change held across repeated tests or other tasks. The source says public logs are available, but the supplied account does not provide the logs’ contents or an independent assessment. It remains unclear how well the result generalizes and what effects the edit had beyond the reported score.
The one-bias XOR change should not be read as a general description of trained networks. The example is deliberately constructed so its logic is easy to trace. In larger models, behavior may depend on many parameters interacting across layers, as the source itself cautions.
What Further Evidence Would Show
The immediate next step for readers seeking to verify the model-editing claim is to examine the public logs cited by Thorsten Meyer and compare them with the source’s account of the test. The supplied material does not give a publication schedule, a follow-up experiment, or a peer-reviewed study, so no later milestone can be confirmed from it.
For the digit-recognition walkthrough, the arithmetic can be checked against the stated network design: each layer’s inputs, weights, biases, activation, and output can be calculated in sequence. To establish how well the 109,386-parameter model recognizes real handwriting, readers would need details about its data and measured performance that are not included in the source excerpt. Further evidence on the separate weight-editing example would need to clarify the benchmark, repeatability, and effects on other model tasks.
Key Questions
How does a neural network identify a handwritten 7?
In the described example, it receives 784 pixel values from a 28-by-28 image, transforms them through layers using weights, biases, and activation functions, then selects the highest score among outputs for possible digits.
Does the input layer recognize features?
No. In the explainer’s MLP, the input layer holds and passes on the pixel values. The hidden layers perform the transformations used to calculate the output.
What does the XOR example prove?
It shows that a small network with nine numbers can be designed to compute XOR and that changing one bias can change its output for all four input pairs. It does not establish that large trained models are as easy to alter predictably.
Is the reported AI weight-editing result peer reviewed?
Not according to the supplied source, which describes the report as a blog post with public logs, not a peer-reviewed paper. The details and broader reliability of the reported score change are not established in the material provided.
Source: ThorstenMeyerAI.com
![Rails World 2026 Opening Keynote [Video]](https://elfysworld.com/wp-content/uploads/2026/09/rails-world-2026-opening-keynote-video-featured-560x300.jpg)
