TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
A November 8 explainer from Thorsten Meyer AI follows a handwritten digit through a neural network and uses a nine-number XOR example to show how changing one bias can change a network’s output. It also cites an MIT CSAIL blog post reporting that an AI model edited another network’s weights; the post’s results are not a peer-reviewed finding.
The explainer describes a multilayer perceptron, a feedforward network built from dense layers. Its input layer holds the image’s pixel values; each hidden neuron multiplies inputs by weights, adds a bias and applies an activation function. The resulting values pass to the next layer. At the output, the network can provide ten scores, one for each digit, with the highest score taken as the prediction.
To make the arithmetic concrete, Meyer constructs an XOR network using two binary inputs and three step-activation neurons. XOR returns one when the inputs differ and zero when they match. Two hidden neurons represent the conditions “at least one bit is on” and “at most one bit is on”; the output returns one when both conditions hold. The example assigns nine numbers across the weights and biases.
The article then changes the output bias from −2 to −1. With that single edit, the output fires when either hidden condition holds, which is true for all four possible inputs. The network now returns one for every input rather than computing XOR. Meyer uses the example to illustrate how a parameter change can alter behavior, while noting that trained networks often distribute behavior across many parameters and layers.
A Handwritten Digit’s Journey to an Answer
A network turns image pixels into digit scores. A small XOR example shows how one bias change can alter the answer.
From pixels to prediction
Each layer transforms values using weights, biases and activation functions.
The input layer holds the image’s 784 numbers.
Neurons combine inputs with weights and bias, then apply an activation.
The highest of ten digit scores becomes the prediction.
A bias changes XOR
In the toy network, one parameter edit changes the function the network computes.
- Output bias: −2
- Returns one when inputs differ
- Computes XOR
- Output bias: −1
- Returns one for all four input pairs
- No longer computes XOR
Toy example, larger model
The XOR network makes every operation easy to follow; a digit reader has many more connections.
Two binary inputs
Three step-activation neurons demonstrate how a bias can change behavior.
Image-scale computation
The described model has 109,386 parameters, about 92% in its first layer.
A separate weight-editing report
The explainer also points to an MIT CSAIL blog post about an AI model editing other networks’ weights.
| Reported detail | What the source says |
|---|---|
| Weights edited | 0.016% of a language model’s weights |
| Math score | 7.9% to 48.1% |
| Evidence context | Blog post with public logs; not peer-reviewed |
Why Network Weights Matter
The example makes the relationship between a network’s parameters and its output visible: a weight or bias is part of the computation, and changing it can change the answer. This gives readers a small, inspectable case for understanding what researchers mean when they describe editing a model’s weights.
Meyer connects the demonstration to an MIT CSAIL blog post, which he says reports that an AI model read and edited the numerical weights of other networks. According to the source article, one edit changed 0.016% of a language model’s weights and raised a math score from 7.9% to 48.1%. The post is described as having public logs, but it is not a peer-reviewed paper; those reported results should be understood in that context.
Products Worth Considering
LG UltraGear 32GX850A-B 32-inch 4K UHD OLED Gaming Monitor
Apple iPad 11-inch with A16 Chip, 512GB, Wi-Fi 6, and Apple Pencil (USB-C) — Blue
Amazon Kindle Paperwhite (12th Generation, 2024), 16 GB, 7-inch Display, Black
As an affiliate, we earn on qualifying purchases.
From XOR to Digit Recognition
The XOR example is a teaching construction, not a digit-recognition model. It uses two binary inputs and a step activation so the reader can follow every operation. A digit image instead begins as a 784-value input vector, and a practical network uses many more connections.
The source article says its small digit reader has 109,386 parameters, with about 92% in the first layer because each pixel connects to each of 128 units. This count illustrates the scale difference between the hand-worked example and a model that processes image data. The article also describes an accompanying video of about seven and a half minutes, narrated with an AI voice clone of Meyer.
“A neural network is arithmetic in layers.”
— Thorsten Meyer AI
Limits of the Reported Results
The source does not identify the language model, the math evaluation, or the conditions behind the reported change from 7.9% to 48.1%. It also does not provide enough detail here to assess how the score was measured or whether the result generalizes beyond that evaluation. The cited work is characterized as a blog post with public logs, not a peer-reviewed paper.
The explainer’s XOR network is fully specified as a toy example, but it does not show that a single parameter edit can generally predictably change a complex trained model’s behavior. Meyer says behavior in trained networks is often spread across many parameters and layers.
What Readers Can Follow
The source provides an accompanying video that walks through the handwritten-digit example and the XOR bias change. Readers seeking to evaluate the separate weight-editing claim would need to consult the cited MIT CSAIL post and its public logs for the model, evaluation setup and reported procedures. The source material does not state a next research milestone or publication date.
Key Questions
How does a neural network identify a handwritten digit?
In the example, a 28-by-28 image is represented by 784 pixel values. Layers transform those values using weights, biases and activation functions; the output layer can assign scores to the ten possible digits, and the highest score is the prediction.
What does changing the bias do in the XOR example?
Changing the output bias from −2 to −1 makes the network return one for all four possible pairs of binary inputs. The network stops computing XOR, even though its architecture and other values stay the same.
Is the XOR network the same as the digit reader?
No. XOR is a small teaching example with two binary inputs and three neurons. The source describes a separate digit reader with 109,386 parameters.
What does the article report about AI editing model weights?
It cites an MIT CSAIL blog post that reportedly describes an AI model editing other networks’ weights. The source says one edit changed 0.016% of a language model’s weights and raised a math score from 7.9% to 48.1%; it characterizes the post as not peer-reviewed.
Source: Thorsten Meyer AI
