8th June, 2026

What am I working on?

I am working on a commission this weekend. It is a small gift for someone's farewell and this piece is about all the memories between him and the team.

What am I reading?

Wash Day Diaries by Jamila Rowser and Robyn Smith is a vibrant, slice‑of‑life celebration of friendship and the intimate ritual of hair care. Through thoughtfully rendered panels, it follows four friends in New York City as they navigate family, relationships, mental health, and work; often through their group chat, and almost always while someone is tending to their hair.

The comic beautifully captures the deeply personal relationship one has with their hair, devoting entire sequences to the hours‑long ritual of washing, detangling, and styling. These moments are not just about beauty but they become acts of care and connection. Rowser ties these individual stories together with the thread of sisterhood, showing how these black women support one another while navigating life in America.

It’s a quick read: the illustrations are vibrant and detailed and the characters are drawn with such warmth that they feel like your own friends. Overall, it is a good read!

What am I watching/watched recently?

Robot Dreams is a tender, dialogue‑free animation that speaks volumes through its simplicity. With naive yet expressive 2D character design, the film opens on a lonely dog living in solitude—eating alone, playing Pong against himself, and drifting through life. Everything changes when he buys a robot from a TV commercial. Their bond is immediate, and together they explore New York City. The sequence is warm and the background song is playing on loop at my home. (September by Earth, Wind and Fire)

The turning point arrives after a long beach day where the robot’s system jams and he falls flat. Despite the dog’s desperate attempts to help, he cannot save him and promises to return, only to find the beach closed until the next season. What follows is a poignant exploration of separation: both dog and robot struggle to function alone, their dreams filled with imagined reunions.

Although kids will delight in the charm of the characters, the adults will recognize the ache of loneliness, the awkwardness of living alone again, and the bittersweet memory of relationships that shaped us but did not last. There are so many beautiful sequences—like a bird nesting beside the rusting robot, the dog’s playful snowmen, or the robot’s dream of rescue—capture emotions ranging from grief and jealousy to hope. The ending was surprising for me (I was hoping a reunion) as both dog and robot find new partners. Even though the quiet awareness of each other’s absence lingers, they give a chance to the new beginnings. The song they danced together on is now the song on which they dance with their new partners :,)

New thing I learnt lately?

With the increasing use of AI, I started thinking how does LLM grasp the meaning and depth of words? Words like 'Love' or 'Grief' that have such complexity that sometimes even we cannot really define it.

1- Tokenisation: The inputs that we give are broken down into tokens. It is like chopping a sentence into bits so that the model can start comprehending it.

2 - Embedding - These tokens get converted into a vector (a list of numbers that represent what the word means in a multi dimensional space relative to other words on different dimensions) So words similar to each other appear in a cluster. But does that convey depth of understanding of the model? That may just be statistical co-occurrence.

3 - Attention - Here is where it gets interesting. Attention lets the model focus on right context. It lets every word look at every other word and form the context. And all this happens simultaneously and sequentially.

Then comes Superposition. It refers to the ability of models to represent more concepts than there are neurons or dimensions, by overlapping them in shared slots. This allows nuanced distinctions—like differentiating “long‑lost love” from “unrequited love”—because meanings are represented as neighborhoods with multiple angles. This is what gives a model its depth of understanding. A model that would have cleanly boxed taxonomies for everything would rather be a worse model.

It has its own drawback as superposition makes it notoriously difficult for humans to look inside an LLM and understand exactly what a specific neuron is doing and figure out the reason for a particular output.

This has definitely intrigued me and there is so much more exploration to do: How does superposition manifest differently in different parts of an LLM? Through the neural networks, the model can create new richer representations and answer our queries. But real life is not that simple - can superposition subsume real life complexities and give informed outcomes?