Reports

Beyond Speech-to-Text — what would actually be next-level for Bidet

2026-05-10

Discrete transcribe-then-clean stays right (output IS text). Single-Gemma-3n is the architectural ideal. Speech-to-speech skips text only when output is audio. BCI not faster than speech.

Brain dump retrospective — receipts for The brain dump has changed my life

2026-05-10

~49 hours of speech into AI tools across 6.6 months. The brain-dump habit produced the brain-dump tool (Apr 7 spec). 33 PRs in 2.5 days. With verbatim attributable quotes.

Kaggle writeup framework — 1000 raw + 500 clean = 1500 total

2026-05-10

Six brain-dump sections. Opens with The brain dump has changed my life. Closes with Take a brain dump Bidet AI cleans the mess. Includes the judge-specific Clean-for-others prompt.

Prompt design research — 2 final drafts + 5 presets

2026-05-10

PR 28 prompts directionally right; 2 gaps: input delimiter + length anchor. SLP framing internal-only. Vygotskyan diff is UI not prompt.

Accessibility options — top 3 to ship in v0.3

2026-05-10

Atkinson Hyperlegible new default font (not OpenDyslexic). Read-aloud TTS highest-evidence option. Text-spacing presets WCAG 2.1 SC 1.4.12.

Moonshine + sherpa-onnx + Unsloth integration research

2026-05-10

Pivot to Moonshine→Gemma routing is sound. Gemma 4 now Apache 2.0. Streaming partial transcripts officially supported. One risk: libonnxruntime.so cohabitation. ~0K stacked target.

Bidet AI promotion opportunities — 12 contests + events + venues

2026-05-10

Top 3 this week: Show HN Saturday, DEV.to Build-with-Gemma track stack, HF Space wrapper. Plus ATIA Jun 3, Tools Competition fall, CSUN Sept CFP. Ranked by fit not money.

State of On-Device STT 2026 — Verify Gemma 2x perf + better-than-Whisper survey

2026-05-10

Gemma 4 E2B at 2x realtime on Pixel 8 Pro CPU is expected (Google benchmarks confirm). Moonshine Tiny beats Whisper-tiny on accuracy AND speed. sherpa-onnx is 51x faster than whisper.cpp.

Gemma 4 audio cold-start — root cause + ladder

2026-05-09

Tensor G3 has no OpenCL — GPU backend silently hangs. Fix: Backend.GPU → Backend.CPU. Pre-warm + grayed Record. Whisper flavor fallback if needed.

Kaggle Gemma 4 — Verified Prize Tree

2026-05-09

$200K stackable. Cactus $10K is the find — fits Bidet exactly. Future of Education + Cactus locked. Deadline 5/18 23:59 UTC.

Bidet 31-min brain dump — Cleaned for Others

2026-05-09

The polished version Mark's phone owed him after the on-device generate hung. Origin → product → contest decision → video concept.

Gemini Deep Research — Bidet AI

2026-05-09

Cross-validation of Claude's deep research. Adds the SLP Receptive/Expressive taxonomy + Vygotskyan scaffolding recommendation.

Bidet AI - Deep Research Verdict

2026-05-09

Are you reinventing the wheel? Opinionated answer: no, but the gap is narrower than yesterday's dossier said. Keep the dual-axis.

Project Euphonia + Tab 4 redesign

2026-05-09

Euphonia is partial-integratable; Tab 4 should split into Read for me + Speak for me.

Brain Dump as a Tool

2026-05-09

135 years of evidence behind verbal-input + customizable-output. Contest-ready citations.

DEV vs Kaggle Comparison

2026-05-09

Side-by-side: DEV.to ($3K, 5/24) vs Kaggle ($200K, 5/18). Calendar to hit both.

AI Radar - 2026-05-09

2026-05-09

Weekly scan: top 3 actions, 8 ranked, verified sources.

Morning digest - 2026-05-08

2026-05-08

TP3 brain at 58,357 rows. Stack health, calendar, project briefs.