🔍 Read the full analysis: Follow The Math As AI Recognizes A Handwritten 7 on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
A November 8 explainer from ThorstenMeyerAI.com follows how a neural network turns pixel values into scores for handwritten digits, using a small XOR network to show how weights and biases affect behavior. It also describes a 109,386-parameter digit reader and reports a separate MIT CSAIL experiment in which an AI edited another model’s weights; that result comes from a blog post, not a peer-reviewed paper.
ThorstenMeyerAI.com published an explainer on November 8 tracing how a neural network can classify an image of a handwritten seven, from pixel values to an output prediction. The article uses a small, hand-built network to show how changing a single bias can change a model’s behavior, then describes a digit-reading network with 109,386 parameters.
The walkthrough describes a multilayer perceptron, a feedforward network built from connected layers. For a 28-by-28 pixel digit image, the input can be represented as 784 pixel values. The input layer holds and passes those values along; it does not itself transform them. Hidden layers calculate weighted sums, add biases and apply activation functions. The output layer can produce ten scores, one for each digit, with the highest score used as the prediction.
To make those operations easier to follow, the article constructs a network that computes XOR: it returns one when two input bits differ and zero when they match. Two hidden neurons detect whether at least one bit is on and whether at most one bit is on; an output neuron combines those conditions. The example uses nine numbers in total—weights and biases—and a step activation that returns one when its input is zero or greater.
The article then changes the output neuron’s bias from −2 to −1. With that one edit, the network returns one for all four possible pairs of input bits, rather than only when the bits differ. The example shows how parameters influence an output in a network small enough to inspect directly. It does not establish that a large trained model’s behavior can usually be traced to one isolated number.
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 Pixels to Model Parameters
The walkthrough gives readers a concrete account of what happens between an image and a classification. Rather than treating the answer “7” as a mysterious leap, it lays out the sequence: pixel values enter the network, layers combine them through arithmetic, and the output layer assigns scores to possible digits. That distinction can help readers understand what a model prediction represents—and what it does not.
The parameter count also shows why inspecting a realistic network is harder than inspecting the XOR example. ThorstenMeyerAI.com says about 92% of the digit reader’s 109,386 parameters are in its first layer, where every pixel connects to each of 128 units. In a larger network, behavior can be distributed across many weights and layers. A single-number demonstration is useful for understanding the mechanics, but it should not be mistaken for a general explanation of a trained system’s decisions.
The source also connects direct parameter edits to a reported MIT CSAIL experiment. It says an AI model read and edited another network’s weights, and that one edit changed 0.016% of a language model’s weights while its reported math score rose from 7.9% to 48.1%. Those figures are presented in a blog post with public logs, not a peer-reviewed paper, so they are a reported result rather than an independently established conclusion here.
Popular right now
Breville Bambino Espresso Machine, Brushed Stainless Steel
De'Longhi La Specialista Opera Espresso Machine with Coffee Grinder
Vitamix 64 oz Low-Profile Blender Container with SELF-DETECT, BPA-Free
As an affiliate, we earn on qualifying purchases.
How Layers Carry Pixel Values
In the article’s account, a dense layer can be written as z = W x + b, followed by an activation that produces the layer’s output. Here, x is the incoming list of values, W contains the weights, and b contains the biases. The result is passed to the next layer. Weights determine how much inputs contribute to a unit’s total; biases shift the point at which it responds.
The terms “hidden layer” and “parameter” have specific, limited meanings. Hidden layers are intermediate computations between the supplied input and the returned answer; they are not hidden because their operations cannot be described. Weights and biases are model parameters learned during training in the usual setup, though they can also be assigned directly in a deliberately constructed example. By contrast, width and depth—the number of units and layers—are architectural choices.
The article’s linked video is described as a hand-drawn explainer of about seven and a half minutes, moving from the handwritten-seven example to the XOR bias edit. Its source says the narration uses an AI voice clone of Thorsten Meyer. The material is an educational walkthrough, not a newly reported scientific paper establishing a new method for digit recognition.
“A neural network is arithmetic in layers.”
— ThorstenMeyerAI.com, describing the walkthrough
Limits of the Reported Examples
The source material does not identify the exact dataset, model architecture, training setup or evaluation procedure used for the 109,386-parameter digit reader. It also does not provide enough detail to assess how often that model recognizes handwritten sevens correctly, or how its performance compares with other systems. The number of parameters describes the model’s size, not its accuracy.
For the MIT CSAIL result, the source does not name the model, math benchmark, evaluation conditions or the researchers involved. It characterizes the evidence as a blog post with public logs, rather than peer-reviewed research. The reported score change and weight-edit percentage should therefore be attributed to that account; the supplied material does not establish whether the result has been independently reproduced.
The XOR network demonstrates a behavior designed into a tiny example. The source itself cautions that, in trained networks, behavior often depends on patterns distributed across many parameters and layers. It does not show that the same single-bias intervention can reliably explain or control a real digit classifier or language model.
Further Checks on Model Edits
The immediate next step for readers is to follow the linked walkthrough’s calculations and distinguish its teaching example from the reported model-editing experiment. The article provides no scheduled follow-up, new release date or forthcoming research milestone.
For the reported MIT CSAIL result, useful follow-up information would include the underlying logs, model and benchmark details, evaluation methods, and any independent replication or peer-reviewed publication. Until those details are available in the supplied material, the result remains a claim attributed to the source’s account rather than a settled finding. For the digit example, performance figures and test conditions would be needed to judge how well the network recognizes handwritten digits beyond explaining the arithmetic it performs.
Key Questions
How does the network turn a handwritten seven into a prediction?
The image is represented as 784 pixel values for a 28-by-28 image. The network transforms those values through layers of weighted sums, biases and activation functions, then returns scores for possible digits. The digit with the highest score is the prediction.
What does changing the XOR network’s bias demonstrate?
In the article’s nine-number example, changing the output bias from −2 to −1 changes the output for every possible pair of input bits. It illustrates how a parameter affects a deliberately simple network; it does not show that one number explains most behavior in a large trained model.
What is the 109,386 figure?
It is the number of parameters the source gives for a small digit-reading network. The article says about 92% are in the first layer, where each of 784 input pixels connects to 128 units. The count is not an accuracy score.
Has the AI weight-editing result been peer reviewed?
The source describes it as a blog post with public logs, not a peer-reviewed paper. It reports a math-score increase from 7.9% to 48.1% after an edit to 0.016% of a language model’s weights, but the supplied material does not give enough detail to assess or independently confirm the result.
Source: ThorstenMeyerAI.com
