Norwin YuAI usage research 中文 Open to research conversations

Journal / 202621 notes

Updated / 2026.11

Workbench
notes.

Research, systems and design notes from the workbench. Write long when the question earns it; write short while the answer is still moving.

No content calendar, no growth tricks. Only the questions that were still there after the build.

Writing note: every note here is AI-drafted; projects, data and facts come from me and are reviewed.

Balanced accuracy for the training pool before and after cleaning.
Featured researchNLPEvaluation

A fingerprint
in the noise.

Telling language models apart with a small classifier, and why clean data is more useful than a bigger model.

Read research · 9 min

All notes

20 notes
21

Twenty-five sentences, two hands

Two sessions, almost the same number of human instructions: 44 rounds of defect reports on one side, a system rewrite on the other. What each round changed, where the eight gates stayed green while the street looked wrong, and where the gap actually sits.

20

A third of the cost, seven dead ends

Measuring an engineering agent: a hard budget cuts cost to a third with no score change, rewritten project documents lift the partial score, and seven self-check ideas all fail.

19

Turning a novel into evidence

chinese-novel-to-game is an extraction project, not a game generator: verbatim evidence, merge rules, a knowledge ledger, and gates that can be rerun.

18

One skill, eight gates

What Photoreal Game Harness is: a procedure, eight executable gates, and how it turns a demo into a finished 3D game.

17

The same harness, a second genre

From a 3D shooter to city karts: what transfers, what has to be rewritten.

16

Blender is an asset compiler

Mistvale and Harvest: one farm concept, two asset pipelines, and what Blender buys.

15

Eight gates, 318 KB

Black Ice and Neon: how one harness turns “it runs” into “it ships”.

14

Two Taihe Halls

One brief, two generations: a local Maya build and the online DeepSeek Flash 1017 one.

13

The exam as evaluation

Two 8-question runs scored 67 and 84: the eval framework's job is to admit the sample is too small.

12

Three minutes per restart

Changing a config means loading 55GB: experiment orchestration, the optimizations the second card throws away, and idle NVLink.

11

80 and 46

The same 27B model, 80 tok/s in the docs and 46 measured: the power wall and the compute density of MTP.

10

4% per frame

A zero-subprocess Rust hardware monitor: layered sampling cuts GPU collection from 97ms to 15ms.

09

Parallel is not a split

A 20B video model on two 22GB cards: 1.31× to 2.97×, with the host thread as the blocker, not the link.

08

1,500 steps later

A scale post-mortem on the Three Kingdoms tactics game: where selection, map, auto-battle and phone controls broke first.

07

Recognition is only a draft

RoomTessera: the hard part of image-to-3D isn't recognition, it's letting the user correct it.

06

An engine is not a bigger canvas

Godot + DSH for an airport tycoon: longer feedback chains, planes clipping through, airport flow and desktop distribution.

05

Agent loops need taste

More tools don't make a stronger agent. What's scarce is a loop that knows when to stop.

04

The last mile is taste

A generative interface gets you to the first draft; that last 10% still runs on editing, judgment and taste.

03

The 44GB VRAM problem

What two 22GB 2080 Ti cards can and can't do, and where local inference actually bottlenecks.

02

What 1.4MB actually buys

Unpacking Saltline: zero assets, a model-written content pipeline, and the difference between a “scene script” and a generator.