Journal entry 2026/07/03

11:11

One of the most rewarding parts of learning ‘difficult’ concepts is when something suddenly clicks. The graph of an activation function shows how much learning can actually happen at different values of Z. Because Z is calculated as weight × input + bias, bigger weights push the value of Z farther out along the horizontal axis. In the flat “plateau” regions of the curve (the tails), the slope becomes almost horizontal. This means the gradient is nearly zero, so the network barely updates those weights during back-propagation, learning effectively stagnates in those areas. That’s the vanishing gradient problem in action. Taken from Andrew Ng online course on Deep Neural Network fundamentals

Journal Entry 2026/07/02 I’ve managed to play at least one chess game every day for the last 304 days, and I’m now heading toward 2,500 games. The chess algorithm is strategically brutal. It has a way of finding your weaknesses, exposing your blind spots, and forcing you into positions where ego becomes a liability.

Taken from ‘The Magic Of Code’ by S . Arbesman

The theoretical questions that humans can answer, but AI cannot, will be the questions that require the human mind to escape its own context window. Not just process information within a frame, but step outside the frame entirely.

The system robbed so many people blind that it managed to convince them money actually grows on trees.

‘It is not only monetary code , it is moral code. It does not judge your character, but over time it exposes it. Patience is rewarded. Humility matters. Low leverage matters. Truth matters. Proof of work matters. Ego gets punished. Fraud gets punished. Entitlement gets humbled. Impatience gets taxed. Bitcoin may not require virtue, but the last honest market has a way of punishing its absence.’ Thoughts from @jack mallers

🫶

Taken from ‘The Magic Of Code’ by S . Arbesman

Light travels 9,460,730,472,580,800 meters in one year.

‘In machine learning training, smaller batch sizes give fresher gradient updates, but they can make hardware less efficient by creating idle time known as pipeline bubbles. This idle compute does not have to be wasted. In theory, unused capacity between training steps could be redirected toward Bitcoin mining. Bitcoin mining turns wasted compute and energy into monetized value.’ Thoughts inspired from a conversation on The Dwarkesh Podcast

By Richard Brautigan.

Journal Entry 2026/06/30 Exited the gym at 06h00 and was greeted by the beauty of the moonlight 🫶

The most important equation in Machine Learning: θ_{t+1} = θ_t - η ∇L(θ_t) • θ — theta, represents all the learnable parameters • η — eta, that’s the learning rate • ∇ — nabla, the gradient symbol • L — L, the loss function

A note to the Average social media user… TikTok is a black box. The algorithms determining who sees what are not publicly transparent. The source code is not open source. That means users have no real visibility into how content is selected, prioritized, or suppressed. It’s entirely possible that the algorithm leverages vast amounts of behavioral and device-level data to shape what each user sees. Over time, this can construct highly personalized narratives — realities tailored to the individual. As a result, two people on the same platform can inhabit completely different worlds, shaped by the content fed to them. Same app. Different reality.

Public servants should serve…not be served.

“Information is eating the other sciences.” - Rizwan Virk (MIT computer scientist)