AI News on July 1th

1. From OpenAI to DeepMind: Meta Poaches Top Researchers to Form Super Intelligence Lab

#TechIndustryMoves #AITalentWar

On June 30, Meta announced the launch of the "Super Intelligence Lab (MSL)," led by Alexandr Wang, former CEO of Scale AI. According to Zuckerberg's internal memo, the newly recruited team members include top researchers from Google, OpenAI, and other firms, such as Trapit Bansal, the pioneer of OpenAI's reinforcement learning "thought chain"; Shuchao Bi, the creator of GPT-4o's voice modality; and Jack Rae, the head of Google DeepMind's Gemini pretraining technology. The lineup covers core technical talents in fields like large language models and multimodality.

2. Apple May Abandon In-House AI Model! In Talks with Anthropic/OpenAI, Claude Could Power Siri

#TechStrategyShift #AICollaboration

On July 1, Bloomberg reporter Mark Gurman reported that Apple is considering dropping its in-house AI model and is in talks with Anthropic and OpenAI to evaluate integrating their large language models into Siri. Apple has requested customized versions of the models tailored to its cloud infrastructure to protect user privacy. Siri lead John Giannandrea led tests showing Anthropic’s Claude outperformed ChatGPT. Apple’s VP of Corporate Development has negotiated with Anthropic, but high licensing fees remain a hurdle. If talks fail, Apple may turn to OpenAI. Internal AI projects are progressing slowly, and the external partnership strategy has caused team unrest.

3. Zuckerberg Bets Big on AI! Poaches OpenAI Talent but May Abandon In-House Llama Model

#AIStrategyControversy #CorporateR&DDynamics

The New York Times reports that Meta is considering adopting OpenAI or Anthropic’s AI models instead of its in-house Llama, as Llama is open-source while others are proprietary. However, Meta claims it remains committed to Llama and plans multiple new versions this year. Despite heavy investments in poaching top AI talent from OpenAI and forming the "Super Intelligence Lab," investors question the sustainability of this "mercenary" strategy. Llama 4’s third version was delayed due to uncertain improvements.

4. Microsoft’s MAI-DxO: AI Diagnosis Accuracy 4x Higher Than Doctors, Costs Drop 70%

#AIinHealthcare #TechnicalBreakthrough

July news: Microsoft launched the AI diagnostic system MAI-DxO, which achieves 4x the accuracy of experienced doctors in complex cases while reducing costs by nearly 70%. Evaluated using the "Sequential Diagnosis Benchmark (SDBench)," which simulates real diagnostic workflows, the system was tested on 304 complex case reports from The New England Journal of Medicine. The AI and doctors both started with case summaries and asked questions for information, with the AI generating realistic synthetic test results.

5. AI’s New Frontier: Shifting from Prompt Engineering to Context Engineering, Key to Agent Success

#AITechnologyTrend #DevelopmentParadigm

The AI field is shifting focus from "prompt engineering" to "context engineering." This concept involves providing large language models with complete contextual environments, including system prompts, user history, long-term knowledge bases, and more. Dynamic integration of information and tools enhances model performance. Research shows AI agent failures often stem from insufficient context, not model capability. Compared to single-prompt design, context engineering focuses on building dynamic systems that deliver critical information at the right time and format, a core challenge in AI application.

6. Baidu Open-Sources ERNIE 4.5 Series: 10 Models Fully Released, Outperforming GPT-4o

#OpenSourceAI #ChineseTechProgress

On June 30, Baidu open-sourced its ERNIE 4.5 series, including 47B and 3B parameter Mixture of Experts (MoE) models and a 0.3B dense model, among 10 total models. Pre-trained weights and inference code are fully open. Developers can download and deploy them on platforms like PaddlePaddle’s Xinghe Community and Hugging Face. Baidu’s Qianfan Big Model Platform also offers API services for these models. Baidu had previewed this release in February.

7. AI Takes Over YouTube! 4 of Top 10 Channels Are AI-Generated, Creators Flee to TikTok

#AIinContentCreation #PlatformTransformation

YouTube is experiencing an AI content surge, with four of May’s top 10 subscriber-growth channels fully reliant on AI-generated videos. AI content’s high update frequency squeezes professional creators, despite some fake engagement data. Meanwhile, top YouTube creators are migrating to TikTok and Instagram. Data shows these AI channels never post fully human-made videos, with relentless updates driving their success, signaling a fundamental shift for the world’s largest video platform.

8. Alibaba International AI Launches Multimodal Model Ovis-U1: Breakthrough in Cross-Modal Processing

#MultimodalAI #ChineseTechInnovation

Alibaba’s international AI team recently released the multimodal model Ovis-U1, the latest in the Ovis series. With 300M parameters, it unifies multimodal understanding, text-to-image generation, and image editing for the first time. Its innovative architecture—featuring a visual tokenizer, visual embedding table, and large language model—efficiently aligns visual and text embeddings, excelling in tasks like mathematical reasoning.

9. Zhipu Draws OpenAI’s Attention: Chinese AI Firm Exports Tech Globally

#GlobalAICompetition #ChineseFirmOverseasExpansion

OpenAI’s Global Affairs account published "Chinese Progress at the Front," highlighting Chinese AI firm Zhipu. Founded in 2019, Zhipu has made "notable progress" in the global AI race, exporting infrastructure and tech support to Vietnam, Indonesia, and others. Its "localized Chinese version of OpenAI" gained traction before Western competitors entered these markets. Despite being added to the U.S. entity list, OpenAI noted its overseas growth, sparking jokes about a "reverse advertisement."

10. Cursor Launches Web Version: Manage AI Coding Agents Directly in Browser

#AIDevelopmentTools #ProductUpdate

On June 30, AI coding tool Cursor’s parent company Anysphere released a web version, allowing users to manage AI programming agents via browsers. The web app lets users give natural language commands for tasks like writing features or fixing bugs, monitor agent progress, and merge code changes.