Shengwen Yang

@turingbook Hold on — Professor Liu Zhiyuan leads the ModelBest team, not Moonshot AI. Are you sure about that?

面壁智能刘知远老师团队聚焦端侧模型，现在也是世界顶级了。

@_LuoFuli So the actual gift is $0.72 cash + a full month of MiMo-V2.5-Pro access. Not bad honestly, just terrible onboarding. You have to find the plan yourself and create your own API key. If you applied, check: platform.xiaomimimo.com → Console → 订阅管理(Subscription Management)

@_LuoFuli After going through the entire process — sending the request, getting approved, and registering the account — my final balance shows only $0.72. Is this what the Xiaomi Mimo 100T actually offers to developers?

Just dropped two open-source models: MiMo-V2.5-Pro (Code Agent, 1T total) and MiMo-V2.5 (Multimodal Agent, 310B total). Oh and one more thing — we're giving devs & creators 100T tokens on us. Go build something cool 🛠️ 🎁 100T Free Token Grant for Builders 100t.xiaomimimo.com

@_LuoFuli Update: turns out the $0.72 isn't the whole story. Logged into the console and found a free 1-month Standard Token Plan sitting in 订阅管理 (Subscription Management) — 200M Credits, $16/mo value. The approval email never mentioned it.

Stop building Data Agents. Start building Data Skills + Data Semantics.

A few days ago we said: Stop building Data Agents. Start building Data Skills. Then we built one. What should’ve taken ~4 months → done in a few days with AI coding tools. Here’s what actually happened 👇

12/ Final take: - Agents are interfaces. - Semantics are assets. AI can generate the first. Not the second.

11/ Where this leads: Soon everyone can build a Data Agent in days. So differentiation shifts to: semantic layer quality

10/ Unexpected insight: Evaluation > Agent We built 14 gold questions. Every change must pass them. No vibes. Just signal.

9/ Most valuable system we built: Audit → Gap detection → Governance loop We can answer: “What does the agent fail at most?” And fix THAT.

8/ Search design: -exact match first (FTS) - vector fallback And always return: `path_used` Tiny detail. Huge impact.

7/ SQL safety lesson: It’s not regex. It’s a compiler. AST parsing + cost estimation + guardrails.

6/ Provenance changed everything: Every fact has: - source - confidence - evidence Result: - conflicts become visible - answers become explainable

5/ We tried letting AI generate metrics. It failed perfectly: - clean SQL - wrong business meaning So now: LLMs propose. Humans approve.

4/ Architecture that saved us: Scribe (offline) → writes semantics Reader (online) → serves Skills Rule: LLMs never write to the semantic layer.

3/ Biggest realization: Skills are NOT tools. They are a versioned semantic layer: - metrics - lineage - definitions - knowledge Stored in files. Reviewed in Git.

2/ We refined the idea: AI can generate: - Harness - Skill functions AI cannot compress: - semantics - conflict resolution - provenance - governance

1/ AI didn’t just make it faster. It exposed something: Code compresses. Judgment doesn’t.

Check out my latest article: AI Built Our Data Agent in Days. What It Couldn’t Build Matters More. linkedin.com/pulse/ai-built… via @LinkedIn

Check out my latest article: Stop Building Data Agents. Start Building Data Skills. linkedin.com/pulse/stop-bui… via @LinkedIn