🧭 An annotated reading list

World over Words

"Training world models over word models."— Saining Xie (谢赛宁), founding manifesto of AMI Labs

Every paper, model and system Saining Xie names in his first-ever interview — a 6h45m marathon with Zhang Xiaojun (张小珺) — sorted into three tiers, each linked to the exact second it's mentioned in the video.

17SAINING CALLS GREAT
10SAINING'S OWN WORK
35WORKS MAPPED
6h45mSOURCE INTERVIEW

Saining Xie: website · 𝕏 · GitHub   |   Zhang Xiaojun (张小珺): 𝕏 · YouTube

Tier 1 — Papers Saining calls great

代表作 · 16

Saining's spontaneous enumeration of the deep-learning canon — the "~20–25 papers that deeply shaped deep learning and AI." Saining keeps his own work out: asked "DiT 不算吗?" he answers "算 0.25" — only a quarter.

🔍 Decoding the garbled names

"Tension Solution Unit" → an attention paper  ·  high confidence on "attention"

This transcript reliably renders self-attention as software attention / Software Tension (2:58:30, 2:58:45) — so "Tension" = attention. Since Transformer is listed separately, the best fit is the original attention paper, Bahdanau et al. 2014. ⚠️ Confirm by ear at ~2:22:04 — issues & PRs welcome.

Tier 2 — Saining's own work

10

Covered in depth across the interview — but he keeps these out of the "great" tier (he rates DiT "0.25"). MoCo & MAE are Kaiming He–led, with Saining as a co-author.

Tier 3 — Other works & systems referenced

RL · world models · systems

Mentioned in passing or as influences — the RL & world-model lineage, contemporary systems, and concepts.

Systems: Sora · SeeDance · Nano Banana · Gemini / ChatGPT · likely Veo · PyTorch Concepts: Neural Architecture Search · Scaling Law · Marr Prize nomination

✨ How this was built

Compiled from the full verbatim transcript — 4,926 timestamped segments — not secondhand summaries. That precision matters: popular recaps listed ConvNeXt and V-JEPA as mentioned, but those names never appear in the actual audio (0 hits). They were imported from Saining's CV.

Deep-linked timestamps

Every ▶ pill jumps straight into the Bilibili video at that exact second.

Three clear tiers

Separates "papers Saining calls great" from "Saining's own work" from "merely referenced."

Garbled-name forensics

ASR mistakes decoded with evidence and the precise second to verify by ear.

Full-transcript coverage

All 4,926 segments scanned and read in context, not skimmed.