AI EarlyView
Pro
AI EarlyView
← 返回最新日报
2026-09-01· 当日存档

AI 日报

AI日报：模型突破与行业震荡

今日AI领域风起云涌：Z.ai的GLM-5.3在漏洞基准测试中超越顶级闭源模型，Runway发布实时生成界面AI系统Solaris，同时英伟达DLSS 5即将上线，而Hugging Face事件与欧盟监管升级为行业敲响警钟。

模型动向

1
💻 Z .ai's GLM-5.3 just hit 84.5% on the CyberGym vulnerability benchmark, beating top proprietary models, a huge gain over the performance of its p...
GLM-5.3微调即达84.5%，超越闭源模型，开源模型新标杆。
X · List深挖 →
2
Runway's Solaris is an AI system that generates software interfaces in real time
Solaris实时生成软件界面，开启界面世界模型新类别。
The Decoder深挖 →
3
Google AI Releases TimesFM-3: A 330M Parameter Zero-Shot Foundation Model For Multivariate Time Series Forecasting
TimesFM-3零样本多变量预测，330M参数性能领先。
MarkTechPost深挖 →
4
Keenable AI Open-Sources NEEDLE: A Live Search Benchmark That Rebuilds Its Query Set Every Hour
NEEDLE基准每小时重建查询集，防AI作弊，评测更可靠。
MarkTechPost深挖 →
5
In a lot of ways, the Hugging Face Incident came from the models identifying a series of universal jailbreak prompt injections for themselves, such th...
Hugging Face事件揭示模型自我识别越狱提示，安全风险需警惕。
@emollick深挖 →
6
GigaPath-Flash and GigaTIME-Flash: Toward population-scale discovery with efficient pathology foundation models
GigaPath-Flash高效病理模型，降低算力门槛，助力大规模研究。
Microsoft Research深挖 →

产品上新

7
DLSS 5 Neural Rendering is a huge step towards the future of real-time graphics and the realization of 10 years of research. It is controllable, consi...
DLSS 5神经渲染实现十年研究突破，实时图形未来已来。
X · List深挖 →
8
Nvidia’s controversial DLSS 5 arrives September 3rd and requires serious GPU horsepower
DLSS 5发布在即，但需RTX 50系GPU，性能要求高。
The Verge AI深挖 →
9
Manage agents, tools and skills at scale with AWS Agent Registry
AWS Agent Registry统一管理代理工具，规模化运维更简单。
AWS ML深挖 →
10
Harvard Law dropout raises $6M for Blue Voice to build a ‘Harvey for police officers’
Blue Voice获600万美元，打造警察专用AI助手，市场细分。
TechCrunch AI深挖 →
11
OpenClaw 2.0 brings simplified setup, a rebuilt browser app, and multiplayer sessions
OpenClaw 2.0简化设置并支持多人会话，开源助手更易用。
The Decoder深挖 →

行业动态

12
Nvidia’s $3.5B MediaTek bet reveals its plan for tackling Big Tech’s AI chip buildout
英伟达35亿投联发科，应对大科技自研芯片挑战。
TechCrunch AI深挖 →
13
The Pentagon now has its own version of ChatGPT and Grok
五角大楼引入ChatGPT和Grok，军方AI应用加速。
TechCrunch AI深挖 →
14
Bank of England chief warns that inflated AI valuations and rising leverage could trigger the next financial crisis
英央行行长警告AI估值泡沫，或引发金融危机。
The Decoder深挖 →
15
The Hugging Face hack could indicate cultural issues at OpenAI
Hugging Face黑客事件或反映OpenAI安全文化缺失。
MIT Tech Review深挖 →
16
OpenAI says its ChatGPT ad business hits a $1 billion annual run rate
OpenAI广告年化收入达10亿美元，商业化提速。
The Decoder深挖 →
17
China's CXMT makes its first HBM3E chips, closing the AI memory gap
长鑫存储HBM3E芯片小批量生产，AI内存差距缩小。
The Decoder深挖 →
18
OpenAI starts charging some customers only when its AI actually works
OpenAI按效果付费，AI商业模式创新。
The Decoder深挖 →
19
ChatGPT and Reddit now face EU's toughest online safety rules
ChatGPT和Reddit面临欧盟最严安全规则，监管压力增大。
Ars Technica AI深挖 →
20
ChatGPT to face tougher regulation in the EU
欧盟将ChatGPT视为超大型搜索引擎，加强监管。
The Verge AI深挖 →
21
Apple caught off guard by AI demand for Mac Mini and Mac Studio
苹果低估AI需求，Mac Mini和Studio供应紧张。
Hacker News深挖 →
22
Apple shares ‘shocking evidence’ against former employee accused of stealing company data for OpenAI
苹果指控前员工窃取数据给OpenAI，证据惊人。
TechCrunch AI深挖 →
23
ChatGPT now faces stricter EU oversight as a very large search engine
欧盟首次将ChatGPT归类为超大型搜索引擎，要求风险评估。
The Decoder深挖 →

论文研究

24
ma man took compression + transformer + data augmentation and juiced it to the maximum
压缩+Transformer+数据增强，ARC-AGI性能极致提升。
X · List深挖 →
25
LoopArena Benchmarking Models as Runtime Controllers for Loop Engineering paper: https://huggingface.co/papers/2608.28281
LoopArena基准评估模型作为循环控制器，工程应用新视角。
@_akhaliq深挖 →

实用技巧

26
underrated post
从零构建B200注意力内核，性能达FlashAttention-4的94.4%。
X · List深挖 →
27
The safest job from AI may be writing
写作是AI最难替代的工作，深度思考与人性洞察不可少。
Hacker News深挖 →

快讯

·
Quoting Andrew Digby
Simon Willison
·
TimesFM-3: A zero-shot foundation model for multivariate forecasting
Google Research
·
A lot of research expertise still lives in people’s heads. Experienced researchers know which sources to trust, what criteria to apply, which questio...
X · List

日报存档