🔍 Read the full analysis: The Route From A Handwritten 7 To A Neural Network’s Answer on ThorstenMeyerAI.com
TL;DR
ThorstenMeyerAI.com published a walkthrough of how a neural network converts pixel values into a digit prediction, using a small XOR network to show how changing one bias can alter its output. The article also describes a separate MIT CSAIL report in which an AI model directly edited another network’s weights; the reported results come from a blog post with public logs, not a peer-reviewed paper.
ThorstenMeyerAI.com published an explainer on November 8 showing how a neural network can turn the pixel values of a handwritten digit into a prediction, using a small XOR network to make its calculations visible. The article also points to a reported MIT CSAIL experiment in which an AI model edited another network’s weights, while noting that the account is a blog post with public logs rather than a peer-reviewed paper.
The walkthrough follows a multilayer perceptron, or MLP, a type of feedforward network built from dense layers. For a 28-by-28-pixel image, the input can be flattened into a list of 784 pixel values. The input layer holds and passes along those values; it does not perform the layer calculations described in the article.
In a hidden layer, each unit multiplies incoming values by weights, adds a bias and applies an activation function. The resulting outputs feed the next layer. An output layer can produce 10 scores for the digits 0 through 9; the network’s prediction is the digit with the highest score. The article presents this process as arithmetic that can be inspected, even when a real network’s many calculations are difficult to follow.
To show how a single number can affect behavior, the article builds a small network for XOR, which returns one when two input bits differ and zero when they match. Its two hidden units check whether at least one bit is on and whether at most one is on. Together, those conditions identify inputs with exactly one bit on. The network contains nine numbers: its weights and biases. Changing the output bias from −2 to −1 makes the network return one for all four possible inputs, according to the explainer.
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 Weight Changes Matter
The XOR demonstration makes a specific point: a network’s output depends on its parameters, and changing even one bias can change the function it computes. In this small example, the behavior follows directly from the listed arithmetic. In a trained model, the article says, a behavior is often distributed across many parameters and layers, so locating an effective edit is harder.
That distinction matters to claims about AI systems that modify other AI systems. The source reports a separate experiment in which an AI model read and edited a language model’s weights. It says an edit affecting 0.016% of the weights was associated with a math score rising from 7.9% to 48.1%. Those figures are attributed to the source’s account; the material provided does not establish the test conditions or independently validate the result.
If such editing methods work reliably, they could offer a way to alter a model’s behavior by changing its internal parameters rather than only changing its prompts. The cited example is not enough to establish how general, repeatable or safe that approach is. The source itself describes the report as a blog post, not a peer-reviewed paper.
Popular right now
De'Longhi COM630BK All-in-One Coffee and Espresso Machine
iRobot Roomba Plus 678 Robot Vacuum and Mop with AutoWash Dock
Dyson Cool AM07 Bladeless Tower Fan, Iron/Blue
As an affiliate, we earn on qualifying purchases.
From XOR to Digit Recognition
The explainer connects a deliberately small, hand-built network to a much larger digit reader. In the XOR example, the author chooses the weights and biases by reasoning from the desired behavior. In ordinary training, by contrast, a model’s weights and biases are learned from data. The article distinguishes those learned parameters from width and depth, which are architectural choices set before training.
For the digit-reading example, the source gives a network with 109,386 parameters, about 92% in its first layer. It attributes that concentration to the layer connecting every input pixel to each of its 128 units. This count illustrates why the simple nine-number XOR example should not be mistaken for a full account of how a trained image model works: the larger network has far more values and interactions to track.
The source also links to a roughly seven-and-a-half-minute video that illustrates the journey from a handwritten 7 to the XOR bias change. It says the narration uses an AI voice clone of Thorsten Meyer. The video is an explanatory resource, not separate evidence for the reported model-editing results.
„A neural network is arithmetic in layers.“
— ThorstenMeyerAI.com explainer
Limits of the Reported Edit
The source does not provide enough detail to independently assess the reported 7.9% to 48.1% math-score change. The test set, evaluation method, number of runs and comparison conditions are not specified in the supplied material. It also does not explain how the researchers selected the edited weights or whether the result held across other tasks or models.
The source identifies the experiment as a blog post with public logs, not a peer-reviewed paper. Public logs may allow readers to inspect parts of the work, but the material here does not establish that the result has been independently replicated. The distinction between a tiny network whose behavior can be traced by hand and a trained model with more than 100,000 parameters also remains important: the XOR example illustrates parameter sensitivity, not a complete method for interpreting or safely editing large models.
What Further Evidence Would Show
The next step for readers assessing the model-editing claim is to examine the public logs and underlying evaluation details cited by the source, including the task, scoring method and conditions before and after the edit. Independent tests would help establish whether the reported change can be reproduced and whether it applies beyond the specific model and math evaluation described.
For the neural-network walkthrough, the linked video offers a visual companion to the article’s calculations. The source material does not announce a publication date for a peer-reviewed paper or a follow-up experiment. Until more details or independent results are available, the weight-editing figures should be treated as reported results, not a settled measure of what AI systems can do in general.
Key Questions
How does a neural network identify a handwritten 7?
In the example, a 28-by-28 image is represented as 784 pixel values. Network layers transform those values using weights, biases and activation functions, and the output layer returns scores for possible digits. The highest score determines the prediction.
What does changing one bias do in the XOR example?
Changing the output bias from −2 to −1 makes the nine-number network return one for all four two-bit inputs, rather than returning one only when the bits differ.
What does the source report about AI editing another model?
It says an AI model changed 0.016% of a language model’s weights and that a math score rose from 7.9% to 48.1%. The source describes the account as a blog post with public logs, not a peer-reviewed paper.
Has the reported weight-editing result been independently confirmed?
The supplied material does not establish independent replication. It also does not give enough detail about the test conditions, evaluation method or repeat runs to assess the result fully.
Source: ThorstenMeyerAI.com
