10-03-Daily
AI 日报中心

爱窝啦 AI 日报

每日 AI 行业动态、产品更新、前沿研究、开源项目与社媒信号

Today’s Digest

OpenAI cuts prices, Higgsfield locks in 80% gross margin with proprietary routing, Xiaomi's MiMo lead promoted to top tier in 11 months—B2B monetization and model routing are redefining AI moats
Open-source projects with almost zero code, pure Markdown workflow docs, hit 500K stars—agent value shifts from raw capability to process control and task decomposition
Start with Higgsfield's dynamic routing cost logic and two workflow doc projects today, then decide if your agent needs an SOP layer

🔥 Top 10 Today

1. Higgsfield Hits 80% Gross Margin on Enterprise Revenue

B2B accounts for over half. Higgsfield’s founder disclosed in an interview that B2B commercial client revenue slightly exceeds 50% , with high-frequency OPC clients contributing 40%. First-month churn sits around 30%, but by month 12, enterprise NRR tops 300%🔄; one cross-border e-commerce client scaled from $99/month to $6M annual contract. Core moat: proprietary commercial photography image models, deeply fine-tuned open models, and dynamic model routing—keeping blended gross margin above 80%. Relying solely on closed commercial LLMs would compress margin to 20–30%.

Higgsfield data screenshot
Higgsfield data screenshot

2. Xiaomi MiMo Lead Luo Fuli Promoted to Top Tier

Eleven months to level 22. Xiaomi MiMo lead Luo Fuli promoted to top tier —level 22, the ceiling in Xiaomi’s ladder. Luo joined Xiaomi in November 2025 after contributing to DeepSeek-V2 at DeepSeek. In those 11 months, MiMo went from a 7B reasoning model to trillion-parameter scale; MiMo-V2.6-Pro, open-sourced Sept 22, topped Artificial Analysis’s intelligence index for open weights📊. Developers watching domestic LLMs should keep an eye on MiMo-V3’s upcoming HySparse2 architecture.

Luo Fuli promotion notice
Luo Fuli promotion notice

3. Agent-Reach Lets AI Agents Pull Social Platform Content

Zero API fees across platforms. GitHub daily trending project Panniantong/Agent-Reach lets AI agents read and search Twitter, Reddit, YouTube, GitHub, Bilibili, Xiaohongshu, and more🔍. Written in Python, it picked up 696 stars in a day, now at 88,732 total. Perfect for developers needing agents to grab real-time social data via CLI, no platform API fees.

4. Grok Bot Adds Master Agent for Task Delegation

No more manual bot classification. Jike post shows Grok Bot rolled out a master Agent role that delegates tasks and spins up specialized bots🤖. Usability jumped—normal users don’t have to define multiple bots upfront. Tool is invisible; users had no intuition for task-based bot segmentation. If you’re using Grok Bot, try the new master Agent auto-delegation today.

Grok Bot update screenshot
Grok Bot update screenshot

5. ChatGPT Web Version Shows Ads

New ad slots in web UI. V2EX thread reports ChatGPT web version now carries ads. Screenshot shows ad content inside the conversation interface. No official word yet on format or targeting. Web users may see this change; ad-blocker rulesets might need updates📢.

ChatGPT ad screenshot
ChatGPT ad screenshot

6. Vacation Rental Dispute Sparks Consumer Trust Debate

Checkout refund process triggers backlash. Vacation rental dispute sparks consumer trust debate —a guest booked Aranya’s Wuling Mountain property at ¥1,300/night, discovered a leaking bathroom ceiling on check-in. Host offered room swap with ¥2,600 top-up or a discount to stay in the leaking unit. Guest chose refund, but no instant return—10 to 14 business days required. At 10 p.m. on Mid-Autumn night, nearby towns were sold out; the family drove back to Beijing overnight. Aranya offered no apology, no compensation. Cases like this remind travelers to review refund clauses and backup plans.

This generation of middle-class consumers no longer wants to open the blind box of trendy guesthouses
This generation of middle-class consumers no longer wants to open the blind box of trendy guesthouses

7. Two Open-Source Projects Teaching AI Agents Process Discipline

500K+ stars on workflow docs. AI Exploration Guide channel compiled two GitHub megahits—obra/superpowers and mattpocock/skills—together over 500K stars. Almost zero code, all Markdown docs📄, teaching AI how to work. Superpowers’ core: don’t let AI start coding on demand; make it understand, design, decompose, execute, test, review first. mattpocock/skills is more direct—TypeScript expert sharing his real agent SOP. Perfect for developers configuring workflows for Claude Code, Codex, or similar agents.

8. Personal Agent Positioning Sparks Entry Point and Runtime Debate

Hardware may be the stickiest gateway. Jike user Mozihaopeter shares that if a personal agent truly extends you, it shouldn’t have a separate persona, avatar, or identity. This wave sees many using computer use to call closed third-party software, but if you can computer-use someone else, why can’t they computer-use you? You can only wear one device in one spot at a time🎧, and hardware doesn’t get replaced lightly—so hardware may be the stickiest entry point. Another blind spot: why can’t agents run continuously in the real world, circling around you?

9. DeepSeek Harness Launches Desktop Version

macOS and Windows available. DeepSeek official site shows DeepSeek Harness Desktop is out, supporting macOS and Windows💻. Hacker News thread got 382 points and 205 comments. Developers using DeepSeek can download the desktop build today for easier local invocation.

10. Frontier Model Training Requires Composite Reward Schemes

Pure human preference leads to reward hacking. Arena.ai research shows relying solely on human preference is insufficient—models may generate visually appealing but factually wrong content or fall into “Reward Hacking”⚠️. Research adopted composite rewards, including faithfulness-based rewards evaluated via millions of human votes. This training lifted FLUX.2 by 69 points on the leaderboard, Ideogram 4 by 20 points. Offline ablation via Gemini 3.5 as judge showed that adding constraint rewards on top of preference significantly boosted win rate. Teams training generative models should consider this composite reward strategy.


⚡ Product & Feature Updates

openJiuwen X-Router Cuts Agent Invocation Cost by 50%

Self-evolving model routing debuts. QuantumBit reports openJiuwen X-Router’s self-evolving model routing tech launches, Ascend-optimized, making agents cheaper over time💰. Real-world tests show over 50% token reduction. Core idea: pick the right model per request, turn every feedback into the next smarter, cheaper choice. Developers optimizing agent costs should watch this tech.


⌘ Top Open-Source Projects

pbakaus/impeccable: Design Language for AI Tools

Make AI tools design-excellent. pbakaus/impeccable i

Last updated on