🔍 Read the full analysis: How Neural Networks Turn Handwriting Into Predictions on ThorstenMeyerAI.com
Get smart everyday buys 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 neural network classifies handwriting by passing pixel values through layers of weighted calculations and returning scores for possible digits. A November 8, 2026, explainer uses a small XOR network to show how changing one bias can alter behavior, then reports on a separate, unreviewed claim that an AI model directly edited another model’s weights.
A November 8 explainer from ThorstenMeyerAI.com lays out how a neural network turns a handwritten digit image into a prediction: it converts pixels into numbers, processes them through layers of weighted calculations and selects the highest-scoring digit. The article also describes a separate MIT CSAIL report in which an AI model reportedly read and edited another network’s weights, a development that points toward direct inspection and modification of model behavior.
For a 28-by-28-pixel image, the network receives 784 pixel values as its input. In the multilayer perceptron example, the input layer simply holds and passes along those values. Hidden layers transform them, and an output layer returns 10 scores, one for each digit from zero through nine. The highest score supplies the prediction. A score is the model’s output for a class; it is not, by itself, proof that the handwriting has been recognized correctly.
Each neuron in a dense layer multiplies its inputs by weights, adds a bias and applies an activation function. In compact form, a layer calculates z = W x + b, then applies the activation to produce its output. That output becomes the next layer’s input. The explainer stresses that this process is arithmetic, not a written set of rules defining what a seven looks like. The difficulty is the quantity of calculations and learned values involved.
To show how a single value can matter, the article constructs a tiny network for XOR, a function that returns one when two input bits differ. Its two hidden neurons test whether at least one bit is on and whether at most one is on; both tests hold when exactly one bit is on. The network uses a step activation and nine numerical values. Changing the output bias from −2 to −1 makes it return one for all four possible input pairs instead of only the two that differ.
The explainer says a small digit reader has 109,386 parameters, with about 92% in its first layer because each pixel connects to each of 128 units. Those parameters are weights and biases learned during training, although they can also be assigned directly. The article further says MIT CSAIL researchers reported an AI model 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 reported work; the source describes it as a blog post with public logs, not 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 Weight Changes Matter
The examples connect a familiar task—reading a handwritten digit—to the less visible mechanics behind model predictions. A network does not need a human-written rule for every possible handwriting style. Instead, training adjusts many parameters so that combinations of pixel patterns produce useful output scores. The XOR demonstration makes that idea tangible: even a small bias change can change what a network computes.
Directly editing weights could offer another way to investigate or alter a model, beyond prompting it or retraining it. But the XOR case has only nine values and a deliberately designed behavior. The reported language-model intervention concerns a small fraction of a much larger system, and the source’s score change alone does not establish how broadly the method works or whether it transfers to other tasks. The reported result is a reason to examine the method, not evidence that complex models can routinely be controlled with one simple edit.
Products Worth Considering
Logitech MX Creative Console with 9 Customizable Keys, Graphite
Apple MacBook Neo 13-inch Laptop with A18 Pro Chip, 8GB Memory, 256GB SSD, Liquid Retina Display, and 1080p Camera – Blush
Dell UltraSharp U2424H 24-Inch FHD Monitor
As an affiliate, we earn on qualifying purchases.
From Pixels to Learned Parameters
The explainer focuses on a multilayer perceptron, a feedforward network made of dense layers. It is a useful teaching example because the calculations can be written as matrix operations. It is not a claim that every handwriting-recognition system uses this exact architecture. The input, hidden and output layers describe the roles values play in this example: raw numbers enter, intermediate representations are transformed, and class scores come out.
The article distinguishes choices made before training from values learned during it. Width and depth—the number of units in a layer and the number of layers—are architectural choices. Weights and biases are parameters normally learned from training data. The article says wider layers can represent more features in parallel and additional layers can combine earlier features, while the suitable design depends on the task and data. Its video is described as a roughly seven-and-a-half-minute hand-drawn explainer narrated using an AI voice clone of Thorsten Meyer.
“A neural network is arithmetic in layers.”
— ThorstenMeyerAI.com explainer, published November 8, 2026
Limits of the Reported Edit
The source does not identify the language model, the exact math evaluation, the test conditions or the number of runs behind the reported 7.9% and 48.1% scores. It also does not provide enough detail here to assess whether the improvement held on other benchmarks, persisted after further testing or came with trade-offs. The stated 0.016% is a share of weights changed; the source gives no broader comparison baseline for interpreting that figure.
The account is explicitly described as a blog post with public logs, rather than a peer-reviewed paper. The source does not establish whether the method generalizes to other models or whether the changed weights produced the intended behavior beyond the reported math score. Nor does it say that a handwritten-digit network was the model edited in that experiment: the digit network is an explanatory example, while the reported intervention concerns a language model.
What Further Testing Could Establish
The next step for readers evaluating the weight-editing claim is to consult the public logs cited by the source and check the model, evaluation setup and exact edits. Independent replication and results across multiple tasks would help establish whether the score change is repeatable and how narrowly or broadly the technique works. The source provides no timetable for a peer-reviewed publication or further testing, so those milestones remain unknown.
For handwriting predictions, the practical sequence remains straightforward: pixels become input values, layers calculate intermediate outputs, and the final layer scores candidate digits. The explainer’s XOR example shows why inspecting a small network can clarify the role of parameters. Whether comparable interventions can reliably explain or change behavior in large models is still an open issue in the material provided.
Key Questions
How does a neural network recognize a handwritten digit?
In the example, the image is represented by 784 pixel values. Layers multiply and combine those inputs using weights and biases, then the output layer returns scores for 10 possible digits. The highest score is the prediction.
What are weights and biases?
Weights set how much input values contribute to a neuron’s calculation. A bias shifts the calculation before an activation function turns it into an output. Training normally learns these parameters from data.
Why does the XOR example change when one bias changes?
The small network’s output depends on the bias as well as its inputs. Changing the output bias from −2 to −1 makes the output neuron fire for every possible input pair, changing the function the network computes.
Has the reported AI weight-editing result been independently confirmed?
The supplied source calls the account a blog post with public logs, not a peer-reviewed paper. It does not provide information about independent replication, so that confirmation is not established in the material.
Source: ThorstenMeyerAI.com
