Blog

A strange thing happens in software teams: we often know some code should be deleted, but we leave it in. Sometimes it’s commented out, sometimes moved to a “backup” file, sometimes just left there quietly collecting dust. We tell ourselves it might be useful later — but deep down, we know it won’t. So why do we keep it? Not because it’s important, but because deleting it feels risky. What if it breaks something we’ve forgotten? What if it was written by someone senior? What if we just don’t fully understand it anymore? Sometimes it’s just easier to ignore it than to clean it up. ...

Here’s a fun thought experiment: can we spot the human in a room full of randomness? 🧠 Imagine two people in separate rooms, each controlling a light bulb. One flips a coin to decide when to switch it on or off, embracing true randomness. The other tries to act “randomly” based on intuition. Could we figure out who’s the human just by watching the patterns? Simulating True Randomness To explore this, I ran a simulation of 10 million coin flips, grouping them into 5-bit sequences (like 00000 to 11111). With 32 possible combinations, each sequence ideally appears about 62,500 times. Interestingly, the counts were close to this target, showing “frequency stability” — a hallmark of true randomness over time. ...

When you type a question and hit “go,” you’re triggering something remarkable—years of research, training, and technology come together to generate a response in seconds. Tools like ChatGPT or DeepSeek-R1 don’t “understand” your question like a person would. Instead, they look at your words and predict what should come next, one piece at a time. This ability comes from three key stages: Pretraining, Supervised Fine-Tuning, and sometimes Reinforcement Learning. It starts with your question getting broken down into small parts—called tokens. These could be words or even pieces of words. The model treats this as the beginning of a sentence and starts filling in the rest, one token at a time. The response you see isn’t pulled from a database or a Google search. It’s generated on the fly, based on how the model has been trained to “sound” helpful. ...