Skip to content

DEV Community

LYR

Building LYR. Engineering notes on AI performance, latency, inference optimization, and production AI systems.

Tokyo, Japan Joined on Aug 2, 2026 https://lyr.jp/en/

LYR

Aug 3

Are you still picking that setting by gut feel?

#performance #testing #machinelearning #ai

5 min read

LYR

Aug 3

A general-purpose model is too wasteful a vessel for subtitles

#machinelearning #devops #llm #ai

18 min read

LYR

Aug 3

Generation speed is decided by the byte count of the weights, not the parameter count

#machinelearning #performance #testing #ai

7 min read

LYR

Aug 3

Let another AI write your teacher data

#machinelearning #ai

12 min read

LYR

Aug 3

A small specialist drew level with the strongest model in service

#machinelearning #ai

8 min read

LYR

Aug 3

Let the AI pick a translation better than the "right answer"

#machinelearning #testing #ai

10 min read

LYR

Aug 3

Swahili from 0% to 30%. But the value wasn't the score — it was that the failures changed

#machinelearning #testing #ai

11 min read

LYR

Aug 3

I cut the bottom-rank rate from 64% to 24%, and held the top-rank loss to 12pt

#machinelearning #testing #performance #ai

9 min read

LYR

Aug 3

"Add context and the model gets smarter" was half a lie

#llm #testing #ai

9 min read

LYR

Aug 3

The "97% improvement" I threw away

#machinelearning #testing #ai

7 min read

LYR

Aug 3

The AI that was smart on the bench suddenly started making mistakes in production

#machinelearning #testing #devops #ai

7 min read

LYR

Aug 3

Taking a small AI's accuracy from 42% to 84% — and on some tasks, level with the giants

#machinelearning #ai

7 min read

LYR

Aug 3

Compress a model to a quarter of its size and its smarts barely drop

#machinelearning #performance #ai

6 min read

LYR

Aug 3

When "which one is faster?" is the wrong question to begin with

#architecture #testing #ai

5 min read

LYR

Aug 3

Five rebuilds and it still never beat chance — the cause was the labels, not the model

#machinelearning #testing #ai

7 min read

LYR

Aug 3

Don't stop at measuring a single mode

#testing #architecture #ai

7 min read

LYR

Aug 3

Don't explain the "why" until you've taken a control

#testing #ai

6 min read

LYR

Aug 3

That "accuracy" is a definition you chose

#testing #ai

7 min read

LYR

Aug 3

A duplicated benchmark lies to you quietly

#testing #ai

8 min read

LYR

Aug 3

If you let an AI do the scoring, start by doubting the scores

#testing #llm #ai

7 min read

LYR

Aug 3

False fires on UI elements: 44% 22%, without adding a single outer if

#performance #machinelearning #testing #ai

9 min read

LYR

Aug 3

I cut OCR's power budget by an order of magnitude — not by making it faster, but by calling it less

#machinelearning #performance #testing #ai

8 min read

LYR

Aug 3

AI apps get easier once you think in CPU and RAM

#devops #llm #ai

11 min read

LYR

Aug 3

I moved one stateful component and perceived latency went from 469ms to 252ms

#performance #devops #testing #ai

9 min read

LYR

Aug 3

Geography is the latency floor you can't move

#performance #devops #testing #ai

4 min read

LYR

Aug 3

Don't let a benchmark decide what "fast enough" means

#performance #testing #ai

6 min read

LYR

Aug 3

You are the bottleneck

#testing #architecture #llm #ai

6 min read

LYR

Aug 3

I brought AI in-house: E2E ~500ms 120ms, cost down to 1/3–1/10

#devops #performance #machinelearning #ai

10 min read

LYR

Aug 3

Borrowed AI disappears overnight

#devops #machinelearning #startup #ai

6 min read

LYR

Aug 3

Some AI models are disqualified before you ever measure quality

#machinelearning #performance #llm #ai

6 min read

LYR

Aug 3

Last year's right answer becomes this year's wrong one

#performance #devops #testing #ai

6 min read

LYR

Aug 3

One setting took output tokens to 1/27, latency to 1/4, and cost to 1/5

#performance #llm #testing #ai

6 min read

loading...