爱窝啦 AI 日报 AI Daily 2026/10/15: Domestic Non-GPU Accelerators Ship More Than GPUs for the First Time, Science Agent Skill Library Tops GitHub Trending Today
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 Summary
Domestic non-GPU accelerators surpass GPUs in shipments for the first time; Nvidia's market share drops to 34%, narrowing the gap with Huawei Ascend to just 6 percentage points.
The Agent toolchain is advancing on three fronts simultaneously: behavioral auditing, integrating existing tools without modification, and MCP plugin packaging with paid tiers.
Today's priorities are the iFixAi audit tool and CLI-Anything — teams actively deploying Agents can evaluate whether to add them to their workflows right away.🔥 Today’s Top 10 Highlights
1. Domestic Non-GPU Accelerators Ship More Than GPUs for the First Time
Domestic chips have pulled ahead. According to an IDC report , China’s data center AI accelerator shipments in the first half of 2026 totaled 2.008 million units, with non-GPU accelerators reaching 1.016 million units to surpass GPUs for the first time. Nvidia’s share fell to 34%, leaving just 6 percentage points separating it from Huawei Ascend, while Nvidia’s year-over-year shipment volume declined 44%. For enterprises currently procuring compute capacity, the window for evaluating domestic alternatives is now open.
2. Science Agent Skill Library Tops GitHub Trending Today
177 validated scientific skills are now publicly available. K-Dense-AI’s scientific agent skill library on GitHub gained 163 new stars today, bringing its total to 49,338. The library covers biology, chemistry, medicine, and drug discovery, and is compatible with mainstream tools including Cursor and Claude Code. Developers already building research automation pipelines can plug it directly into existing Agents for testing.
3. Ask Grok Directly to Replicate Its News Curation Process
Ask Grok — it will tell you exactly how it works. This write-up from the AI Exploration Guide channel demonstrates a simple idea: ask Grok for the search queries and decision logic it uses to surface AI news, then organize the process into a reusable Skill. Account categories covered include official lab accounts, key figures, and leak accounts, spanning scenarios like model releases, funding rounds, and acquisitions. The key difference is that GrokBot’s X API requires no extra payment, whereas similar functionality in other bots requires you to cover the API costs yourself.
4. AI Content Overload May Replicate the Path of the Atari Crash
Content supply is exploding, but attention is the real scarcity. Jike user JimHuiHui draws a comparison in this discussion , likening AI’s impact on games and film to the 1983 Atari crash: low-quality content floods the market and erodes consumer trust. AI lets one person produce what used to require a team of over a dozen, yet users still only have 24 hours in a day — attention does not expand with supply. For content creators, the real competitive moat is the execution to push quality 10x or even 100x higher, not the tools themselves.
5. Ambi Founder: Only Hardware Agents Can Truly Deliver on Promises
Always-on is a gap that software Agents simply cannot close. In the 42章经 podcast summary reposted on Jike, Ambi founder Zihao points out that current PA products universally lack offline context and natural interaction. Even when AI glasses are worn continuously, users still have to re-explain the background every time they issue a command — which fundamentally undermines the experience an Agent should deliver. Hardware also has structural advantages including independent distribution and exclusive access to a body part, meaning it does not need to buy traffic from closed platforms every time.
6. Gemini 4 Stalls for Thousands of Seconds, Developer Gives Up
A powerful model dragged down by poor stability. Developer 歸藏 documented in this Jike post that while testing Gemini 4 on Antigravity, the thinking process repeatedly stalled, with single pauses lasting thousands of seconds. Auto-approval was effectively useless, with every step still requiring manual intervention. The problem persisted after switching to a direct API connection. Teams planning to deploy Gemini 4 in production should currently hold off and wait for stability improvements.
7. Solo Developer Uses AI to Build a Detailed 3D Little Prince Game in 16 Days
One person plus AI produced a studio-quality 3D work. Jike user Alchian花生 documented this case : developer Sac used Opus 5.5 + Tripo3D + Blender to create a game based on The Little Prince, completing it in 16 days to widespread attention on X. Getting the red scarf to flutter realistically alone consumed a significant amount of time. The core of this story is not the tools — it is the determination and polish that pushed quality higher.
8. Stepstar STEPX Neo Native Agent Phone Launches
A large-model native Agent phone priced at 3,999 yuan. A purchase announcement posted by Jike user 童欧巴 shows that the Stepstar STEPX Neo comes configured with 16GB + 256GB storage, with a full-stack price of 3,999 yuan; the author has already ordered one in carbon black. This is currently a purchase announcement only — no hands-on impressions are yet available. Interested users may want to wait for real-world test results before deciding.
9. User Posts AI Deepfake Video on X, Yuzuru Hanyu Himself Appears in the Comments
An AI video ran into the real person. 宝玉 documented in this X post : a user published an AI-generated deepfake video of Yuzuru Hanyu on X, only to have Hanyu himself respond in the comments in Japanese: “被中国人当AI玩具了” (“Being used as an AI toy by Chinese people”). This is a textbook case of AI-generated content and the real subject involved meeting directly on the same platform. Creators who publicly post AI deepfake content should think carefully about the risks and boundaries beforehand.
10. Renoise Launches Video Creation as an MCP Plugin
Video scripts and storyboards can now be completed directly inside an Agent. 向阳乔木 explains in this introduction that Renoise from the Youware team can be installed in Codex and Claude Code to assist with writing scripts and storyboards for product ads or short films. Calling the AI video generation and TTS features consumes credits, while basic features are free. This is a new distribution structure: packaging capabilities as an MCP plugin with premium features sold separately.
⌘ Top Open Source Projects
HKUDS/CLI-Anything: Let Any Software Be Called by an Agent
Existing command-line tools can be called by Agents without modification. HKUDS/CLI-Anything provides a universal interface layer that makes existing software natively accessible to AI agents, gaining 103 new stars today for a total of 52,497. The project includes a CLI-Hub aggregation directory to make it easy to find already-supported tools. Developers who want to reuse existing toolchains in Agent workflows without building API wrappers from scratch should look here first.
ifixai-ai/iFixAi: Audit Whether an Agent Followed Instructions in 120 Seconds
Agent behavior can now be independently audited 🔒. ifixai-ai/iFixAi supports audits initiated by either a human or the Agent itself, centered on one core question: is the Agent doing what it is supposed to do? It gained 2,109 new stars today for a total of 27,405, making it the fastest-rising AI project on today’s trending list. Teams that need to run compliance or security checks in increasingly automated Agent deployments should try this first.
vxcontrol/pentagi: Fully Autonomous AI Penetration Testing Agent System
Complex penetration testing tasks can be fully delegated to an autonomous Agent. vxcontrol/pentagi is implemented in Go and can autonomously plan and execute multi-step penetration testing tasks, gaining 1,025 new stars today for a total of 29,464. The project targets security researchers and red teams and is suited for automating reconnaissance and vulnerability validation workflows. Before connecting it to a real environment, it is advisable to evaluate its behavioral boundaries in an isolated test environment first.
◉ Social Media Picks
FishAudio Open-Sources Drama 3 — One Text Passage Can Voice 16 Characters
An open-source voice model that can simultaneously play 16 distinct voice roles 🔊. Gorden Sun explains in this hands-on summary tweet that FishAudio’s Drama 3 uses a 35B-A3B-scale Qwen model as its text backbone, with VRAM requirements significantly higher than comparable models. It supports inserting emotion tags for sighs and laughter, and allows local edits without regenerating the entire passage. The model is publicly available on HuggingFace for non-commercial use only.
😄 AI Curiosities
Sergey Brin Joins the Conversation in Person, Google Holds a “Parent-Teacher Meeting”
After Gemini 4 Argon launched, Google skipped the standard press conference and instead had Logan Kilpatrick sit down with Sergey Brin and several product leads to film an internal-style conversation about how they got here and where they are going . Brin appears on camera to talk through Argon’s origins and direction. The whole tone feels more like a technical retrospective than a product launch. After all, the model is named after argon — an inert noble gas — but Google is clearly not in the mood to be inert this time.
❓ Related Questions
Claude subscribers who want to use Claude Code for biology or pharmacy scripts — do they have to re-explain the research background every single time?
The problem is not that the background is too complex — it is that the domain context has not been locked down. Today’s scientific-agent-skills library includes ready-made Skills covering biology, chemistry, and medicine, so you can write the recurring research terminology and workflows into the Skill upfront. Claude Code runs directly in the terminal, cutting out the back-and-forth of re-explaining context every session. Once the workflow is stable, think about a long-term subscription — existing account holders can check out 爱窝啦·AI账号店的 Claude 续费 . Skills are about working method; renewal is just about keeping the tool available — handle the two things separately.