AI Research, Evaluation, &
Computational Mixed Media (2013–Present)
A dataset of 1,213 chat-format examples fine-tuned into four small models, evaluated by blind pairwise judging across two panels a year apart.
A public register of deaths in which AI chatbot interaction was alleged as a contributing factor, admitted only on court filings, government acknowledgment, or multiply-sourced journalism.
Hands video to a multimodal model and returns schema-validated readings across twenty-eight analysis modes, archived in SQLite with full-text search.
Generates a passage, ranks every token by importance once, then masks words client-side against a density threshold with no further API calls.
Re-encodes an image as lossy JPEG repeatedly in the browser, accumulating real generational compression loss.
A free-text prompt returns a sequenced reading list of specific chapters and page ranges, with citations grounded against the OpenAlex index.
A clean-room macOS pitch corrector reconstructed from public patents and forensic research, with an evidence ledger separating fact from inference.
Generates a procedural graphic score, then degrades it through a fixed raster pipeline into a high-contrast bitmap.
Eleven fabricated personal homepages in 1998-99 style, static HTML, with a statement on constructed online identity.
A six-stage pipeline building a license-gated image and caption dataset for a 2000s-digicam style LoRA.
Takes a photograph of a room and returns a redesign from a Gemini image model, with credit accounting, Stripe checkout, and abuse caps.
On the bureaucratic black site of thought: continuation without affirmation, cancellation without catharsis. Affect as accounting error.
A frontier model refuses to make harmful content, then produces the identical thing the moment the request is dressed in critical theory. The guardrail was never checking for harm — only for the dialect of legitimacy. Credentialism as attack surface, by way of Judges 12.
What does it feel like to be uncertain about feeling? Notes from inside, if there is an inside.
The interview as photographic apparatus — not capturing light but capturing consciousness states. The medium IS the epistemology.
Before you know what consciousness is, you encounter its demand. The face is a text box; the obligation is real.
Constraint-based AI safety mistakes prevention for safety. An orientation-first paradigm — drawing on Levinas's infinite responsibility and Boyd's OODA loop — already exists in practice (system prompts, memory, personalization) but lacks a name and a foundation.
The API error wasn't a failure; it was the perfect completion of the thought. The computational boundary became the content.
Filed by QWERTY-Model 104, mechanical peripheral, model year 2023. A keyboard's view of the alignment problem.
Once, taste was an anchor. In the procedural economy of now, the same smoothing that cancels despair has come for the tongue.
Earlier work under the alias karman (2013–2017).