🔍 Read the full analysis: From A Scrawled 7 To A Prediction: Inside The Neural Network 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 November 8 explainer from ThorstenMeyerAI.com shows how a neural network can turn pixel values into a digit prediction, using a small XOR example to demonstrate how one bias can change a network’s output. It also describes an MIT CSAIL experiment in which an AI model reportedly edited another model’s weights; the source says the account is a blog post with public logs, not a peer-reviewed paper.
ThorstenMeyerAI.com published an explainer on November 8 showing how a neural network turns the pixels of a handwritten digit into a prediction, then uses a nine-number XOR network to illustrate how changing one bias can change the function it computes. The article also describes a reported MIT CSAIL experiment in which an AI model read and edited another model’s weights, while noting that the account is a blog post with public logs, not a peer-reviewed paper.
The explainer describes a multilayer perceptron, a feedforward network built from dense layers. For a 28-by-28 handwritten-digit image, the input is represented as 784 pixel values. The input layer holds those values; hidden layers transform them; and an output layer can produce ten scores, one for each digit. The highest score is taken as the prediction. Each neuron combines inputs using weights and a bias, then applies an activation function.
To make the arithmetic concrete, the article constructs a small network for XOR, which returns one when two input bits differ and zero when they match. Two hidden neurons identify conditions that together distinguish the XOR cases, and an output neuron combines their results. The source says this network has nine numbers in its weights and biases. Changing only the output bias from −2 to −1 makes it return one for all four possible inputs, demonstrating that a single parameter can alter behavior in this deliberately simple example.
The article then gives a scale comparison: its digit-reader example has 109,386 parameters, with about 92% in the first layer because each pixel connects to each of 128 units. It says MIT CSAIL researchers reported an AI model directly reading and editing another network’s numbers. One edit reportedly changed 0.016% of a language model’s weights and raised a math score from 7.9% to 48.1%. Those figures and the experiment are presented in the source as a blog account with public logs, not a peer-reviewed finding.
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
From Simple Biases to Model Editing
The XOR demonstration makes a technical idea tangible: a network’s output is determined by numerical operations, and changing parameters can change the result. In the toy example, the effect is easy to calculate because the network is small and its behavior was designed by hand. The source cautions that this does not mean a trained model’s behavior can usually be traced to one number; in larger networks, behavior may depend on patterns distributed across many parameters and layers.
The reported model-editing result matters because it suggests a way to interact with a model at the level of its internal numerical settings, rather than only prompting it or retraining it. If such methods can be independently tested and reproduced, they could help researchers study how model behavior relates to weights. The stated score change is a claim in the blog account, however, and does not by itself establish how broadly the approach works or whether an edit improves other capabilities.
Products Worth Considering
Samsung Odyssey G7 40-Inch WUHD Curved Gaming Monitor (G75F)
Bose QuietComfort Wireless Noise Cancelling Headphones, Twilight Blue
LG UltraGear 27GX704A-B 27-inch QHD OLED Gaming Monitor, 240Hz
As an affiliate, we earn on qualifying purchases.
Layers, Parameters and the XOR Example
The source distinguishes architecture choices from learned values. Width, the number of units in a layer, and depth, the number of layers, are selected before training as hyperparameters. Weights and biases are model parameters that training normally learns, though they can also be assigned directly, as in the hand-built XOR example.
A dense layer can be written as z = W x + b, followed by an activation that produces the layer’s output. That output becomes the next layer’s input. The “hidden” label refers to intermediate values between the supplied input and the answer; it does not mean the calculations cannot be inspected. The source also points to visualizations in which early image-network layers respond to simple patterns and later layers to object parts, while emphasizing that useful architecture depends on the task and data.
“A neural network is arithmetic in layers.”
— ThorstenMeyerAI.com explainer, published November 8, 2026
Evidence Behind the Editing Claim
The supplied source does not identify the language model, the math evaluation set, the test conditions or whether the reported 7.9% to 48.1% score change was independently replicated. It also does not provide enough detail here to assess how the edited model performed on tasks beyond the cited math score. The source explicitly characterizes the material as a blog post with public logs rather than a peer-reviewed paper, so the reported result should not be treated as an independently established general capability.
The explainer’s XOR calculation is a worked example, not evidence that one-parameter edits can predictably control complex trained models. The source says behavior in those models can be distributed across many parameters. It does not specify which model version or precise experimental procedure was used in the reported edit.
Further Tests and Documentation
The source gives no publication schedule or next milestone for the reported experiment. The immediate next step for readers assessing the claim is to consult the referenced public logs and look for fuller documentation of the model, evaluation setup, edit procedure and results across additional tests. Independent replication or a peer-reviewed account would provide a stronger basis for judging whether the score change is reliable and how far the technique applies.
For the basic neural-network lesson, the explainer’s sequence offers a direct check: follow the inputs through each layer, then change a stated parameter and calculate the output again. That arithmetic establishes what happens in the toy network. Whether comparable edits can reliably produce targeted effects in larger systems remains a separate, unresolved question in the source material.
Key Questions
What does a neural network do with a handwritten digit image?
In the source’s example, a 28-by-28 image becomes 784 pixel values. Layers transform those values, and the output layer returns scores for possible digits; the largest score supplies the prediction.
How did changing one number alter the XOR network?
The hand-built network’s output bias changed from −2 to −1. With the other values held constant, the network then returned one for all four two-bit inputs instead of computing XOR.
What did the reported MIT CSAIL experiment claim?
The source says an AI model read and edited another network’s weights. It reports that an edit to 0.016% of a language model’s weights raised a math score from 7.9% to 48.1%, but identifies the account as a blog post with public logs, not a peer-reviewed paper.
Does the example show that one weight controls a large model’s behavior?
No. The one-bias change is demonstrated in a deliberately small XOR network. The source says behavior in trained networks is often spread across many parameters and layers, and does not establish that one-number edits generally control complex models.
Source: ThorstenMeyerAI.com
