Deepson Acharya

Undergraduate researcher · Caldwell University

Hover any word. The arcs are scaled dot-product attention — real softmax over real dot products, computed live on toy embeddings of these words. It is the same arithmetic on page 8 of the paper below, and the same mechanism every language model runs billions of times a second.

Embedding space · drag to rotate

About

I'm a Computer Science undergraduate at Caldwell University, studying how language models decide what to say next. My research with Dr. Sushil Poudel takes the attention mechanism apart symbol by symbol and puts it back together — and my companion project, Attention Atlas, lets anyone do the same by hand, in a browser.

I'm drawn to the places where AI is still uneven: what models are confident about versus what they actually get right, and how poorly today's tokenizers handle languages like Nepali. I like building things that make hard ideas feel reachable, and I'm impatient to work on problems I don't yet know how to solve.

  • Studying B.S. Computer Science, Caldwell University — expected May 2028
  • Researching Transformer interpretability and next-token uncertainty in LLMs
  • Building Open, interactive tools that make the mathematics of AI hand-checkable
  • Next Tokenization and uncertainty in Devanagari-script languages
Deepson Acharya, standing outdoors in a dark blazer and white shirt.
Caldwell, New Jersey

Research & artifacts

Two things, built to go together.

Paper · manuscript, 2026

Every Symbol Explained: A Dependency-Ordered Path to the Attention Mechanism

With Dr. Sushil Poudel. Every operand of the attention equation is introduced, exercised on hand-checkable numbers, and flagged for its later role before the equation appears — so when it arrives, it contains no unexplained symbol. Paired with an empirical study of next-token uncertainty in GPT-2: how entropy moves with context, what temperature does to it, and how wide the nucleus really gets.

Preprint soon

Software · archived, citable

Attention Atlas — an interactive course on how transformers work

Thirteen lessons and seven live instruments, from vectors and dot products through to a full transformer. Open source, MIT licensed, no build step, runs anywhere. Swing two vectors and watch the dot product cross zero; drag temperature and watch a distribution flatten.

doi:10.5281/zenodo.22005900

Open ↗

Measured in the paper, on GPT-2:

  • 0.36 → 14.19 bitsnext-token entropy across a single temperature sweep
  • 63×variation in nucleus (top-p) pool size from one context to another
  • “2+2=” → “ 3”the model's single most likely completion

From the paper

How sure is the model?

Before every word, a language model holds a probability over every token it knows. Temperature reshapes that distribution. Entropy measures how spread out it is, in bits. Nucleus sampling keeps only the smallest set of tokens that covers 90% of the mass — the bright bars. Drag the temperature and watch all three move at once.

The logits here are toy values chosen to be legible. The measured GPT-2 numbers — including why formulaic openers turn out to be the sharpest category — are in the paper.

NEXT TOKEN?

softmax(z / T) over eight candidate tokens · hover a bar for its probability

Entropy
bits
Nucleus, p = 0.9
tokens
Top token

What I'm working on next

Not every language costs the same.

GPT-2 and most models like it read bytes, not characters. A Latin letter is one byte. A Devanagari character is three. Before a model has learned anything at all, Nepali already costs roughly three times as much to represent as English — which eats context, inflates cost, and shows up as worse performance for the languages that can least afford it.

Type anything below. The numbers are exact UTF-8 byte counts, computed in your browser.

English
Nepali

Contact

Say something.

Research questions, collaborations, or anything about the Atlas — this reaches me directly. If you'd rather not use a form, email works just as well.