A smaller speech model earns its place on the iPhone
Aidana transcribed the 47 matched clips 3.35 times faster, with observed English WER of 2.15% and Spanish WER of 5.37%.
4 min readThe experiment archive
The journal.
What I tried. What happened. What I’d do next.
Showing all 44 experiments
A coding agent that knows when to ask for help
Three live recovery scenarios passed, each finishing with 10/10 checks.
4 min readCorrective data meets a fresh gameplay gate
On Ubuntu CPU, the corrected policy passed 10/10 full-kitchen trials against a same-host retraining control at 7/10.
6 min readWhen the retention gate says stop
All three registered learning rates crossed the retention floor after eight steps and rolled back.
7 min readA second dish, a new frozen test
The corrected candidate passed 10/10 paired two-dish trials against a frozen baseline at 0/10.
6 min readEight clears are still not a promotion
The v3 protocol measured 8/10 full-stage clears versus 0/10 for its paired baseline; the strict gate rejected it.
6 min readLearning to play from pixels
The first-dish baseline produced six successful recorded evaluations across five starting positions, with exact replay.
4 min readThe reusable RGBA learning pipeline
Owned frames, bounded controls, saved state, teacher-data export and exact replay form one auditable pipeline.
9 min readWhat counts as a successful episode?
The evaluator tightened survival, terminal-state and exact-replay checks without rewriting historical experimental outcomes.
4 min readFinding a local model for the inbox
Bonsai-2-27B PQ2_0 with MTP passed 48/50 tasks with thinking on, at a 15.8-second median on an RTX 3060.
10 min readTeaching a 4B model to read tmux status
The classifier matched its teacher on 192/202 examples, with macro-F1 0.863.
7 min readAn image editor that respects the pixels you keep
A real masked object removal finished in 24.39 seconds and preserved every unselected decoded pixel exactly.
3 min readCharacter LoRAs: identity, motion, and continuity
Three initial adapters completed: 600 LTX steps and two Qwen identity runs of 800 steps each.
4 min readFifteen pictures, one photographic-style adapter
213 files became 97 unique decoded images; the curated random selection used 15 distinct photographs or photo-like images.
3 min readMaking the keyboard lock key really sleep the phone
A keyboard-specific handler preserved the lock gesture and the display reached OFF roughly 0.7 seconds later.
4 min readMasked cleanup before a fresh LoRA
25 distinct images were cleaned into lossless copies, preserving dimensions and all decoded pixels outside each mask.
3 min readMoving speech decode onto the GPU
Decode dropped from 3.36 seconds to 0.24; total render wall time dropped from 35.1 to 24.4 seconds.
2 min readPutting a watermark gate in front of training
The binary screen accepted 24/25 cleaned images; both known positive controls were detected.
3 min readA complete image stack on one 24GB GPU
Eight matched 1024-square scenes averaged 93.352 seconds with FP8 and 46.219 with opt-in INT8.
15 min readLong speech, bounded memory
A 551-character, eight-sentence input rendered in two chunks with 48.0 seconds of healthy audio.
2 min readRound-trip tests expose a broken encoder
The new 1.76GB FP8-E4M3 codec reached FP32-control parity in the measured round-trip test with its encoder intact.
2 min readThe clone bug was in the reference encoder
Using FP32 reference encoding lifted AWQ WavLM target similarity from 0.8475 to 0.9556 while retaining INT8 output decode.
3 min readThe faster image pipeline that produced nothing
The balanced two-GPU path took 48.1 seconds but every image was invalid flat gray noise.
2 min readThe same story, six precision combinations
Generator precision affected the measured ranking; changing the output codec shifted similarity by at most about 0.001.
3 min readBalancing GPU experts when the model will not fit
An explicit 12/12/12 GPU expert split plus 12 CPU blocks measured 31.1 tokens per second.
3 min readPipeline parallelism, INT4, and a 262K context
The patched vLLM stack measured 67.4 tokens per second for one stream and 73.1 aggregate for two.
3 min readA speech service that sleeps between requests
The recorded service used about 10.3GB resident memory; first render took 14.8 seconds and warm render 5.7.
2 min readA quantization job goes silent
The full-spec job stopped after six of 36 blocks; the monitor detected the dead wrapper and idle GPU.
2 min readAWQ speech: fast enough, with a memory catch
AWQ passed 33/33 prompts with pooled real-time factor 0.43 and 35.8 tokens per second.
2 min readDo you really need the bigger embedding model?
The 1B encoder embedded 1,004 documents per second; the 8B encoder reached 181, with no proven retrieval-quality win.
12 min readFP32, BF16, INT8: separating codec drift from speech
Pure TTS legs shared token sequences, and normalized transcription differences were zero across the tested codec legs.
3 min readKeeping the audio-sensitive layers intact
The HQ4 preview matched 96.5% of text tokens in the greedy test and passed three staged-memory probes at 6.7–6.8 GiB.
3 min readStaging the speech stack under a 12GB envelope
The staged run peaked at 6.63 GiB, down from 13.01 before staging; 252 INT4 projections passed the verifier.
4 min readThree draft tokens hit the sweet spot
Draft length three reached a short-prose mean of 88.16 tokens per second, versus 64.67 with drafting off.
8 min readFrom a model download to a working local service
The deployment notebook records the EXL3 serving family, cache choices and real runtime constraints.
7 min readThe EXL3 quantization journey
The notebook follows official BF16 weights through EXL3 conversion, artifact checks and deployment work.
4 min readA model proxy earns its promotion
The promoted merge passed 107 tests with one environmental skip.
8 min readThe best checkpoint was not the last one
Polish step 400 measured 6.1% word error rate; the final step 1500 measured 42.9%.
2 min readRebuilding calibration for an agentic quant
The clean agentic importance matrix recovered the hard T3 task; later draft decoding improved measured throughput.
7 min readA draft head, a rollback fix, and a midnight incident
The rollback fix brought draft decoding to parity at temperature 0.7 and a 24% gain at greedy sampling.
7 min readMaking room for a 200K context
The warm proxy decode median was 59.17 tokens per second; a 198,960-token request passed with GPU headroom.
4 min readDFlash2 on a single RTX 3090
Ornith measured 210.16 tokens per second on the realistic 256-token prompt and 270.40 on predictable prose.
5 min readLess thinking, fewer tokens: a template A/B
Sharp reduced completion tokens by 20.6%, 20.1% and 16.0% across the three deployments.
9 min readThe real cost of a five-second video
At 960×544, dual INT8 completed a 124-frame, 20-step run in 354.580 seconds.
5 min readNo matching experiments.
Try another term or reset the filters to open the complete notebook.