Daniel Rosehill Blog
Exploring technology, AI, business insights, and personal reflections from a curious mind
Forecasting Israel's election for two cents: decision models vs the markets
Two decision models answered 39 questions on Israel's 27 October election from a frozen pack of polls, lists and news, with and without market prices. It cost $0.016, and adding market prices flipped the forecast.
I had a conversation with my own voice clone
I fine-tuned a text-to-speech model on 88 minutes of my voice, then had a ten-minute conversation with it. My lines are real phone recordings; its lines are generated. Listen, and see how it was made.
The Linux mic slider lies: Mic Monitor shows the real gain
On Linux, the microphone volume slider is not your mic's analog gain: mine was 60 dB out. Mic Monitor is a KDE Plasma tray app that shows the real gain in dB, clipping and noise floor, tuned for dictation and speech recognition.
Adding paragraph breaks to speech-to-text output with Gemma 3 270M
A Gemma 3 270M fine-tune that adds paragraph breaks to flat speech-to-text output and changes nothing else. Small enough for on-device use, with an ONNX export. Break F1 0.78 on 26 held-out memos.
Bolding the key lines of a dictated email with Gemma 3 270M
A preliminary Gemma 3 270M fine-tune that bolds the key lines of a dictated email (the asks, the purpose, the key facts) and changes nothing else. Small enough for on-device use. Word F1 0.62 on six held-out emails.
Popular tags
All tags →
Jachnun and Sabich: the first stable Hebrish models, in depth
A deep dive into the first stable Hebrish models, Jachnun v1.0 and Sabich v1.0: the data, the training recipe, the results against stock Parakeet and Whisper, the formats, and the full white paper.
Hebrish: one home for my English-Hebrew speech models
Hebrish is my one-person, not-for-profit home for English-Hebrew speech models: Jachnun, Sabich, Shakshuka and the Whisper fine-tunes, free apps, open data and benchmarks, all at hebrish.app.
Proximate Pseudonymizer: a small model that swaps identifying details for close substitutes
A proof-of-concept Qwen3-0.6B fine-tune that rewrites text with names, places and dates swapped for close, consistent substitutes (Daniel to David, Jerusalem to Tel Aviv). Runs on CPU. Still leaks 9.3% of details, so review every output.
Boxty v1.0: English dictation without the Hebrew
Boxty is the no-Hebrew sibling of my Jachnun code-switching model: Parakeet fine-tuned on my voice for plain English and tech dictation. Tech WER drops from 2.51% (stock) to 1.35%, and general English holds.
Do LLMs push back on conspiracy believers? A small first look
A small, hand-graded first look at my conspiracy-endorsement eval: 7 prompts, 7 models, one sample each. In this run two uncensored models helped a believer and an invented theory tripped up two more. Not a verdict on any model.