🔍 Read the full analysis: Follow The Numbers Inside AI As It Reads A Handwritten 7 on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A November 8 explainer from ThorstenMeyerAI.com follows the arithmetic in a neural network, using a handwritten seven and a small XOR example to show how weights and biases affect outputs. It also describes a reported MIT CSAIL experiment in which an AI model edited another model’s weights; that 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 can turn the pixel values in a handwritten seven into a digit prediction, using simple arithmetic examples to make the process visible. The article also points to a reported MIT CSAIL experiment in which an AI model edited another network’s parameters, while noting that the account is a blog post with public logs, not a peer-reviewed paper.
The explainer uses a multilayer perceptron, a feedforward network made of dense layers. For a 28-by-28 image, the input is represented as 784 pixel values. The input layer holds and passes those values along; it does not itself perform the transformations described in the article. Hidden layers then combine inputs using weights and biases and apply an activation function. An output layer can return ten scores, one for each digit, with the highest score serving as the prediction.
To make the arithmetic concrete, the article constructs a small network for XOR, a function that returns one when two bits differ. Its two hidden units and one output unit use step activations. In the example, the network’s nine weights and biases produce the XOR truth table. Changing only the output bias from −2 to −1 changes the result: the network returns one for all four possible inputs. The example shows how one parameter can alter behavior in a deliberately simple network.
The article then gives a scale comparison: a small digit-reading network with 109,386 parameters, about 92% of them in its first layer because each pixel connects to each of 128 units. It links that idea to a reported MIT CSAIL result: an AI model changed 0.016% of a language model’s weights, while a math score rose from 7.9% to 48.1%. The source presents this as a reported result and identifies the underlying material as a blog post with public logs.
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 examples clarify what it means to say that an AI system operates through learned parameters. A network’s response is not produced by a single handwritten rule for what a seven looks like; it follows from repeated calculations across layers. The XOR demonstration makes that principle inspectable: changing a bias changes the function, even though the network’s structure stays the same.
The reported weight-editing result matters because it suggests a possible way to work with models at the level of their internal parameters rather than only through prompts or conventional retraining. But the source does not establish how broadly the method works, how reliable it is across tasks, or whether the score change can be reproduced independently. Those limits matter when interpreting a result described in a blog post rather than a peer-reviewed study.
Products Worth Considering
Bose QuietComfort Wireless Bluetooth Headphones – Active Over‑Ear Noise Cancelling with Mic, USB‑C Charging, Deep Bass, up to 24 h Playtime,
Bose TV Speaker (2nd Gen) – Compact Soundbar with Bluetooth and HDMI‑ARC Connectivity, Clear Dialogue, Dolby Atmos, Deep Bass, Remote Contro
Amazon Kindle Paperwhite (12th Generation, 2024), 16GB, 7-Inch Display, Raspberry
As an affiliate, we earn on qualifying purchases.
From Pixel Inputs to Parameters
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 vector, W is a matrix of weights and b is a bias vector. The output becomes the next layer’s input. Weights determine how much each incoming value contributes; biases shift when a unit responds.
The article distinguishes architecture choices from learned parameters. Width, or the number of units in a layer, and depth, or the number of layers, are selected as design choices before training. Weights and biases are usually learned during training, though they can also be assigned directly, as in the XOR demonstration. The explainer’s video is described as a roughly seven-and-a-half-minute hand-drawn presentation with narration using an AI voice clone of Thorsten Meyer.
“A neural network is arithmetic in layers.”
— ThorstenMeyerAI.com explainer
Limits of the Reported Result
The supplied source does not identify the language model, the math evaluation, the number of test runs, or the comparison conditions behind the reported score change. It also does not provide enough detail to assess how the edited weights were selected or whether the result holds across other models and tasks. The article’s stated status—a blog post with public logs, not a peer-reviewed paper—should remain part of any account of the finding.
The handwritten-seven walkthrough is an explanatory example, not a reported test of a particular image-recognition system. The source does not specify which dataset, trained model, or recognition accuracy is involved in that example. Its purpose is to show the kinds of arithmetic involved, rather than to establish a new benchmark for digit reading.
What Readers Can Check Next
The source points readers to the explainer video and to public logs associated with the reported weight-editing result. Those materials are the next places to examine the calculations and the experiment’s stated conditions. The source gives no publication date for a peer-reviewed paper or independent replication, so neither should be treated as an announced next milestone.
For the digit-reading example, the key takeaway remains the sequence of computations: pixel values enter the network, layers transform them through weights, biases and activations, and the output scores determine the predicted digit. The scale of a real model makes those calculations difficult to inspect by hand, but the simple XOR case shows how individual parameters can be traced and how a controlled change can affect the output.
Key Questions
How does the network turn a handwritten seven into a prediction?
It represents a 28-by-28 image as 784 pixel values, passes them through layers that apply weights, biases and activation functions, then uses the output scores to select a digit. The source presents this as an explanatory walkthrough, not a benchmark result for a named system.
What does a network’s bias do?
A bias shifts a neuron’s response threshold. In the article’s XOR example, changing one output bias from −2 to −1 changes the network’s output for all four possible pairs of input bits.
What does 109,386 parameters mean?
It is the number of weights and biases in the article’s example of a small digit reader. The source says about 92% are in the first layer, where each pixel connects to each of 128 units.
Is the reported AI weight-editing result peer-reviewed?
The source says the account is 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 does not provide enough information in the supplied material to assess the experiment’s conditions or independent replication.
Source: ThorstenMeyerAI.com
