Loading 3D model…
WHENLABS · INDEPENDENT STUDIO NODE: ARLINGTON.TX · MMXXVI

whenlabs.

me.

my cat.

my office.

STUDIO BRIEF
A one-person studio for small tools, open-source software, and papers about language models, attention, the slow web.
I’m an independent researcher and toolmaker. I write small, durable software and papers about language models, attention, the slow web.
She supervises every commit. Naps through the standups, alert for the treats, the laser, the late-night build.
One desk, one window, no meetings. A quiet room in Arlington for long thinking, slow code, warm light.
↘ SCROLL TO ENTER
// Buildv1.4.2 · MMXXVI
// ModeStudio · Stable
// LocatedArlington, TX
// Active11 projects
// NowDrafting P-03
— 01
/ 01

Hi, I’m Siddharth.

I’m an independent researcher and toolmaker. I run whenlabs as a one-person studio — small, durable software and short papers about language models, attention, and the slow web.

Most of what I make is open-source and gets used by a few hundred people who care about the same things I do. That’s the right size.

PREVIOUSLY: research · ml infra · type design.

WRITES · SHIPS
READS · ANNOTATES
TEACHES · OCCASIONALLY
/ 02

The workshop.

Tools shipping soon. In the meantime, see research for what’s live now.

// tools coming soon
/ 03

Research & preprints.

P-01
The introspective gap.
PREPRINT · LLM SELF-KNOWLEDGE
arXiv:2603.04412
2026 · MAR
P-02
Parametric divergence mapping.
PREPRINT · CROSS-MODEL EVAL
arXiv:2604.11827
2026 · APR
P-03
The rendering tax.
CONTROLLED STUDY · IN FLIGHT
EST. SUMMER 2026
2026 →
/ 04

Now, this week.

// Writing
Finishing the discussion section of P-03 — the rendering tax; running the last calibration sweep on the controlled-study data.
~ 4,200 words drafted of an estimated 6,500. Target submission: end of May.
// Building
A small tea-timer that sits in the menu bar. (yes, really.)
Side-quest from a real frustration with my afternoon Earl Grey.
v0.2 ships next week.
// Reading
Andy Matuschak on tools for thought. Re-reading Latour’s Pandora’s Hope.
+ a pile of papers on attention + self-rep in language models.
/ 05

About the studio.

WhenLabs is a one-person studio run by me out of Arlington, TX. The work splits, roughly evenly, between open-source software (the workshop), short-form research about language models and attention (the preprints), and the slow accumulation of a small body of writing about how we actually work with computers.

Everything here is made with the aim of being useful first, beautiful second, and clever a distant third. If something looks finished, it usually is. If it doesn’t, it’s probably v0.

// Available for: short consulting engagements, paper collaborations,
// and the occasional well-scoped commission.

/ 06

Let’s talk.

siddharth@whenlabs.org ↗

Direct

siddharth@whenlabs.org +1 · by request

Code

github / Caissaisdead github / WhenLabs-org

Words

blog research index arxiv profile

Located

Arlington, Texas CST · UTC−6