🔍 Read the full analysis: Tracing The Data Behind AI’s Recognition Of A Handwritten 7 on ThorstenMeyerAI.com
Get the little things that make your day 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
A November 8 explainer traces how a neural network can classify a handwritten 7, from pixel values through weighted layers to its output scores. It also reports a small digit reader with 109,386 parameters and describes separate, preliminary work in which an AI model edited another model’s weights.
A ThorstenMeyerAI.com explainer published November 8 traces how a neural network turns the pixels of a handwritten 7 into a digit prediction, showing the layer-by-layer arithmetic rather than treating the process as an invisible judgment. The post says a small digit-reading network contains 109,386 learned parameters and connects that example to reported research in which one AI model directly changed another model’s weights.
The explainer uses a multilayer perceptron, a feedforward network made of dense layers, to describe the basic process. A 28-by-28 image can be represented as 784 pixel values. The input layer holds those values and passes them forward; it does not perform a transformation. Hidden layers then calculate weighted sums, add biases and apply activation functions. The final layer can return a score for each of the ten digits, with the highest score used as the prediction.
In a dense layer, each output unit receives connections from the inputs. The explainer summarizes the calculation as z = W x + b, followed by an activation function. Here, weights determine how much inputs contribute, while biases shift when units respond. The hidden layers’ outputs become the next layer’s inputs. These intermediate values are called “hidden” because they are neither the supplied image nor the final answer, not because their arithmetic cannot be inspected.
To show how sensitive a network can be to its parameters, the post builds a small nine-number XOR network. It returns one when two input bits differ. Changing only the output bias from −2 to −1 makes the network return one for all four possible inputs, according to the worked example. The post then scales the discussion to a digit reader, saying about 92% of its 109,386 parameters are in the first layer, where each pixel connects to each of 128 units.
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 show both what can be understood directly and what makes larger networks difficult to interpret. In the XOR demonstration, a single bias has an obvious effect because the network is small and its logic was designed by hand. In a trained digit reader, a prediction can depend on many weights working together, so changing one number may have a less predictable effect.
The post links this issue to what it calls weight-space programming: changing a model by editing its parameters rather than only giving it new instructions as text. It reports that MIT CSAIL researchers had an AI model read and edit other networks’ numbers. In one cited example, an edit changed 0.016% of a language model’s weights and the reported math score rose from 7.9% to 48.1%. The result, if accurately reported, illustrates why the ability to locate and alter parameters could matter. The source describes the material as a blog post with public logs, not a peer-reviewed paper, so the result should not be treated as independently established by this source alone.
Products Worth Considering
Hyperice Venom 2 Heated Vibration Back Wrap
Withings BeamO Contactless Smart Thermometer, ECG Monitor, and Stethoscope
Fitbit Sense 2 Health and Fitness Smartwatch, Shadow Grey/Graphite
As an affiliate, we earn on qualifying purchases.
From Pixels to Network Weights
The article’s central question is what happens between an image entering a model and a digit label coming out. It answers at the level of a specific feedforward architecture, rather than claiming every image-recognition system works in exactly the same way. Its example flattens a 28-by-28 image into a list, sends the values through dense layers, and reads the output scores for possible digits.
It also distinguishes architecture choices from learned parameters. Width—the number of units in a layer—and depth—the number of layers—are selected when the network is designed. Weights and biases are parameters that training ordinarily learns, though they can also be assigned or edited directly. The explainer notes that layer size and depth depend on the task, data and architecture; neither greater width nor greater depth is automatically better.
The source is an educational post by Thorsten Meyer, accompanied by a video described as a hand-drawn explanation of the path from a handwritten 7 to a bias change in an XOR network. The post says the video’s narration uses an AI voice clone. That presentation is distinct from the separate research claim about AI systems editing model weights.
“A neural network is arithmetic in layers.”
— Thorsten Meyer, ThorstenMeyerAI.com
Limits of the Reported Results
The explainer does not identify the specific digit-recognition model, its training data, test set or evaluation conditions in the supplied material. It gives the parameter count and layer connection description for a small example, but those details do not establish how every system recognizes handwriting or how accurately this particular example performs.
The reported weight-editing result also needs qualification. The source says it comes from a blog post with public logs and is not peer-reviewed. The supplied account does not name the language model, explain the math benchmark or provide enough detail to independently assess how the score was measured. It is not clear whether the change generalizes beyond that reported test, or how the researchers verified that the edit caused the reported difference.
The XOR example is a transparent demonstration created by assigning numbers to a network. It illustrates that changing one bias can change behavior in that toy setup, but it does not show that one parameter will have an equally clear or predictable effect in a large trained model.
What Verification Would Require
The next step for readers evaluating the weight-editing claim is access to the underlying logs, model details and benchmark methods, followed by independent attempts to reproduce the result. The source material does not provide a publication date for a research paper or identify a scheduled peer-review milestone, so no such next event can be confirmed here.
For the handwriting example, a fuller technical account would specify the network’s complete architecture, training procedure and performance on a named test set. Until those details are available, the article’s strongest supported point is narrower: a dense network processes image values through explicit arithmetic, while the reported intervention on a language model remains a claim attributed to a non-peer-reviewed blog post.
Key Questions
How does the example network recognize a handwritten 7?
It represents the image as 784 pixel values, passes them through hidden layers that calculate weighted sums and apply activation functions, then produces scores for possible digits. The highest output score is the prediction. The source does not give the full architecture or evaluation results for a particular 7 image.
What is a neural-network weight?
A weight is a parameter that determines how much an input contributes to a unit’s calculation. A bias is another parameter that shifts the unit’s response. Training typically learns both, though they can also be set or edited directly.
What does the 109,386 figure describe?
It is the parameter count the explainer gives for a small digit reader. The post says about 92% are in its first layer, where each pixel connects to each of 128 units. It is not a general count for all handwriting-recognition models.
Is the reported AI weight-editing result peer-reviewed?
Not according to the source. Thorsten Meyer describes it as a blog post with public logs, not a peer-reviewed paper. The supplied material does not include enough information to independently verify the reported benchmark result.
Does the XOR demonstration prove that one weight controls a large model’s behavior?
No. It shows that changing one bias flips the output of a small, hand-designed network. The source says behavior in trained networks is often distributed across many parameters and layers, making larger-model interventions harder to interpret.
Source: ThorstenMeyerAI.com
