🔍 Read the full analysis: The Journey Of A Handwritten 7 Through An AI Neural Network 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
ThorstenMeyerAI.com published an explainer tracing how a multilayer neural network can turn the pixels of a handwritten seven into a digit prediction. It also describes a separate MIT CSAIL report on an AI model editing other networks’ parameters; that report 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 processes a handwritten seven, from pixel values to a digit prediction. The article uses a small, hand-built example to make the arithmetic visible, then describes a separate MIT CSAIL report in which an AI model read and edited other networks’ weights—a development relevant to efforts to understand and modify how models behave.
The digit example starts with a 28-by-28 image, represented as 784 pixel values. Those values enter a multilayer perceptron, or MLP, through an input layer that holds and passes them along. In each dense layer, neurons multiply inputs by weights, add biases and apply an activation function. The resulting values become inputs to the next layer. For digit recognition, the output layer can return ten scores, one for each digit; the highest score is the network’s prediction.
To demonstrate how a parameter can affect behavior, the article builds a separate nine-number XOR network. XOR returns one when two input bits differ. The network uses two hidden neurons and a step activation to represent the conditions “at least one bit is on” and “at most one bit is on.” Both conditions hold when exactly one bit is on. Changing the output neuron’s bias from −2 to −1 makes the network return one for all four possible inputs. This is a deliberately small example, not a description of how a trained digit reader stores its behavior.
The article says a small digit-reading network has 109,386 parameters, with about 92% in its first layer, where each pixel connects to each of 128 units. It also summarizes an MIT CSAIL report: 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 are attributed to the report; the source characterizes it as a blog post with public logs rather than a peer-reviewed paper.
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 Direct Weight Edits Matter
The examples address a basic question about AI systems: how can small numerical changes alter what a network does? In the XOR demonstration, the effect is fully traceable because the network has only nine numbers and the desired behavior was encoded by hand. That makes the relationship between one bias and the output easy to calculate.
A trained model is more difficult to interpret. Its behavior may depend on patterns spread across many parameters and layers, rather than one isolated number. The reported MIT CSAIL work matters because it describes an AI model working directly with another model’s weights. If such methods can be tested and reproduced, they could offer researchers another way to examine or alter model behavior. The reported score change is a result attributed to the source, not independent confirmation of a general capability.
Products Worth Considering
Samsung Galaxy Tab A11+ 11-Inch Wi-Fi Tablet, 6GB RAM, 128GB Storage, Gray
Sony WH-1000XM6 Wireless Noise-Canceling Headphones, Black
Apple MacBook Air 13.6-inch Laptop with M5 Chip, 24GB Memory, 1TB SSD, Midnight
As an affiliate, we earn on qualifying purchases.
From Pixel Inputs to Learned Weights
An MLP is a feedforward network made of dense layers. In a dense layer, each output unit receives connections from the layer’s inputs. A layer 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 input layer itself does not transform the data; hidden layers produce intermediate values, and the output layer provides scores or another answer.
The article distinguishes architectural choices from learned values. Width and depth refer to the number of units and layers and are selected as part of the model’s design. Weights and biases are parameters that training typically learns, although they can also be assigned directly in a constructed example. The video accompanying the source is described as a roughly seven-and-a-half-minute hand-drawn explainer; its narration uses an AI voice clone of Thorsten Meyer.
“A moment later, it says “7.””
— ThorstenMeyerAI.com, describing the digit example
Limits of the Reported Model Edit
The source does not identify the full evaluation setup behind the reported math-score change, including the benchmark, test conditions or whether the result was independently reproduced. It describes the work as a blog post with public logs, not a peer-reviewed paper. The figures therefore should not be treated as proof that small weight edits will reliably produce similar results in other models or tasks.
The explainer also does not provide the architecture, training details or exact output scores for its digit-reading network. Its 109,386-parameter count is a stated example, not a universal size for handwritten-digit systems. The XOR network illustrates how a single bias can flip a simple function; it does not show that one isolated edit will predictably control a complex trained model.
What Further Evidence Could Show
The source points readers to the MIT CSAIL post’s public logs but does not announce a publication date for a peer-reviewed paper, follow-up experiment or independent replication. The next evidence to watch for is a fuller account of the model, evaluation method and repeated tests of the reported edit, along with results from researchers outside the project.
For the handwritten-seven example, the immediate next step is the sequence itself: pixel values pass through the network’s layers, and the output scores determine the predicted digit. The article’s demonstration explains the mechanics at a high level, but it does not report a new digit-recognition test or claim that the accompanying video establishes a new research result.
Key Questions
What does a neural network receive from a handwritten-seven image?
In the example, a 28-by-28 image is flattened into 784 pixel values. The input layer holds and passes those numbers to the network’s first computational layer.
What do weights and biases do?
Weights determine how strongly inputs contribute to a neuron’s total. A bias shifts that total, and an activation function converts it into the value passed onward.
What does the XOR example prove?
It shows directly that changing one number can change a small network’s output function. It does not prove that a single edit will control behavior in a large trained model.
Is the reported MIT CSAIL result peer-reviewed?
The source describes it as a blog post with public logs, not a peer-reviewed paper. The reported score change remains a claim attributed to that post.
Source: ThorstenMeyerAI.com
