Why do LLMs get sometimes simple tasks wrong?
Why do LLMs fail at simple tasks? It usually isn’t a lack of intelligence, but a lack of context. Learn how tokenization affects performance and how a single sentence—asking the model to “think step-by-step”—can dramatically improve the accuracy of your AI results.
The Overestimation Problem
Behavior is not cognition, and imitation is not insight. Explore the history of AI overstatements—from chess-playing computers to modern chatbots—and why maintaining perspective is the only way to think clearly about the reality of machine intelligence.
Is Forgetting the Secret to Mastery?
We often treat forgetting as a human flaw, but in both biological and artificial intelligence, it is a vital feature. Discover why intelligence isn’t about storing more data—it’s about the ruthless deletion of noise. From AI pruning to the struggle of retaking a driving test, this post explores how what we discard defines the quality of our expertise.
Are AI’s answers reality — or just Plato’s shadows on the cave wall?
We don’t just use AI; we project our own humanity into it. By mistaking statistical shadows for conscious minds, we fall into a trap of our own making. Discover why the real danger of AI isn’t the machine itself, but our irresistible urge to attribute intention and reasoning to mere probability.
Could a whale ever make Chomsky wrong?
AI predicts text; it doesn’t experience the world. This post explores the “illusion of depth” created by fluent language and why mistaking correlation for judgment is the biggest hurdle in modern business and research. To use these systems effectively, we must first learn to see the shadows for what they are.
#12 AI Data Quality: Crap in – Crap out
AI Data Quality – Any AI project is based on data used to train the model. Unlike what we would imagine, getting the right data in the right shape is far from easy or obvious. Building a quality dataset is an engineering work. This paper covers the various steps of this job.
#11 AI: Fixing the Training gone Wrong
Building on Paper #10’s AI training pitfalls—underfitting (too lazy), overfitting (too rigid), high bias (skewed guesses), and high variance (wild swings)—this paper offers practical fixes for our smell detector. We explore three levers: boosting network capacity, extending training with more epochs, and enriching data for smarter learning.
#10 AI Training going wrong
This paper explores why the model might fail in practice: underfitting (too simplistic), overfitting (too rigid), and the underlying issues of bias and variance. Through examples, we show how underfitting leads to random guesses , while overfitting causes oversensitivity. We introduce bias (consistent errors) and variance (prediction variability).
#9 AI Training & Back Propagation
AI Training & Back Propagation – In order to use a Digital Neural Network, we need to train it. In this paper we present how we can “train” one using supervised training and backpropagation. By comparing the model’s output with the value that we know to be correct, we can tune the parameters and make it solve the problem at hand.
#8 – AI Forward Propagation
AI Forward Propagation – AI Neural networks mimic the neural network of the brain. In this paper we present what is happening inside a digital neural network from data entry to result. We study the various mathematical steps in their simplest format to allow global understanding of the inside mechanisms. The end-to-end process is called Forward Propagation.