AI EarlyView
Pro
AI EarlyView
← 返回最新日报
2026-06-25· 当日存档

AI 日报

芯片与模型双线突破

OpenAI联合博通发布首款自研推理芯片Jalapeño，剑指大规模LLM部署；百度开源3B参数Unlimited OCR模型，KV缓存恒定；GPT-5.6内部路径曝光，发布在即。

行业动态

1
OpenAI unveils its first custom chip, built by Broadcom
OpenAI首款自研芯片，专为推理优化
TechCrunch AI深挖 →
2
OpenAI and Broadcom announce chip designed for LLM inference at scale
与博通合作，瞄准大规模推理场景
Ars Technica AI深挖 →
3
OpenAI reveals its first AI processor: Jalapeño
芯片代号Jalapeño，计划年底部署
The Verge AI深挖 →
4
OpenAI and Broadcom unveil "Jalapeño," a custom chip built for LLM inference
定制芯片细节，博通制造
The Decoder深挖 →
5
Amazon ups India bet with fresh $13B AI infrastructure investment
亚马逊加码印度AI基建，投资130亿美元
TechCrunch AI深挖 →
6
We just crossed $100M annual run-rate. I know many AI companies are capturing much more $$$ these days, but still proud of the milestone! Maximizing s...
Hugging Face年收入破亿，坚持开源
X · List深挖 →
7
IBM claims world’s first sub-1 nanometer chip technology
IBM首发亚1纳米芯片技术
Ars Technica AI深挖 →
8
Europe is pushing back on Washington’s chip war
欧洲抵制美国芯片出口新规
TechCrunch AI深挖 →
9
Cerebras stock plunges after earnings as CEO says margin outlook was misunderstood
Cerebras财报致股价暴跌
TechCrunch AI深挖 →
10
The memory chip crunch is paying off for this US company
内存芯片公司受益行业复苏
TechCrunch AI深挖 →
11
Companies are scrambling to stop employees from maxing out AI budgets with small tasks
企业限制员工滥用AI预算
TechCrunch AI深挖 →

模型动向

12
Baidu Releases Unlimited OCR, a 3B Model That Keeps the KV Cache Flat for Long-Document Parsing
百度开源OCR模型，长文档解析利器
MarkTechPost深挖 →
13
GPT-5.6 spotted in internal model-access route. That GPT-5.6 is coming is no secret. The only question is when and how quickly it will arrive. My gues...
GPT-5.6内部路径曝光，发布在即
X · List深挖 →

产品上新

14
Google bakes computer control directly into Gemini 3.5 Flash, letting the model see and operate your screen
Gemini 3.5 Flash可直接操控电脑屏幕
The Decoder深挖 →
15
Gradium Launches stt-translate and s2s-translate, Real-Time Speech Translation Models Beating gpt-realtime-translate on Accuracy and Latency
实时语音翻译模型超越GPT
MarkTechPost深挖 →

论文研究

16
LLM Agent Memory 综述：从 RAG 到 KV Cache 到参数写入，100+ 篇工作的方法演进与真实取舍
LLM记忆机制综述，从RAG到参数写入
AutoML机器学习深挖 →
17
Real-Time Voice AI Hears but Does Not Listen
实时语音AI忽略语气情绪
arXiv深挖 →
18
Same Evidence, Different Answer: Auditing Order Sensitivity in Multimodal Large Language Models
多模态模型对输入顺序敏感
arXiv深挖 →
19
RAS: Measuring LLM Safety Through Refusal Alignment
基于内部表征的LLM安全评估
arXiv深挖 →
20
Agentic evolution of physically constrained foundation models
物理约束智能体实现自主硬件设计
arXiv深挖 →
21
Recommendation as Generation: Unifying Personalized Video Generation and Recommendation at Industrial Scale
生成式推荐与视频生成统一框架
arXiv深挖 →
22
DFlash Speculative Decoding Drafts Whole Token Blocks in Parallel for Up to 15x Higher Throughput on NVIDIA Blackwell
并行草稿生成提升吞吐15倍
MarkTechPost深挖 →
23
PolicyAlign: Direct Policy-Based Safety Alignment for Large Language Models
直接对齐自然语言安全策略
arXiv深挖 →
24
How agents are transforming work
AI智能体变革工作方式
OpenAI深挖 →

实用技巧

25
Authors Guild test finds some AI detectors perfectly identify human writing while others fail on every single text
AI检测器表现两极分化
The Decoder深挖 →
26
AI was supposed to kill engineering jobs, but new data suggests they’re the most resilient
工程师岗位未因AI减少
TechCrunch AI深挖 →

快讯

·
TED talks were a sneak preview of AI slop. 1000 people speaking in the same cadence and tone, optimized to sound profound, regardless of substance.
X · List
·
A list of definitely must-read papers of the week ▪️ World Action Models: A Survey ▪️ ImageWAM ▪️ EvoEmbedding ▪️ Training Open Models for Age...
X · List
·
There are exactly 5 (five) mentions of air conditioning in this 37 (thirty seven) page document. The plan’s own numbers show ~387 premature deaths la...
X · List
·
A lot of people who say they never use AI are using AI, but secretly. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5464215
@emollick

日报存档