AI News on July 3th

AI News 2025/7/3 10:00:10

1. World's First AI-Native UGC Engine "Mirage" Debuts, Enables Custom GTA-Style Worlds

#Gaming #AI #NVIDIA

NVIDIA CEO Jensen Huang's prophecy materializes as Mirage, the first AI-native UGC game engine, launches with playable demos including a GTA-style urban chase and Forza Horizon-inspired coastal drift. Surpassing Oasis and GameNGen, Mirage allows players to edit, expand, and create immersive worlds supporting 5-10+ minute gameplay sessions across racing, RPG, and platformer genres, ushering in a new era of dynamic world creation.

2. Apple's DiffuCoder-7B Challenges "Linear Curse", Boosts Code Generation by 4.4%

#Coding #LLM #Apple

Apple and HKU researchers unveil DiffuCoder-7B, trained on 130B code tokens to overcome autoregressive limitations. The model autonomously determines causal dependency strength during generation, with temperature sampling enhancing token diversity. A novel coupled-GRPO sampling method reduces log-likelihood variance in RL training.

3. Global Surge in LLMs: 418 Models Exceed 10²³ FLOPs, China Claims 36% Share

#AI #China #Compute

Epoch AI reports 418 models surpassing 10²³ FLOPs in 2025, with China leading (151 models) and the UK trailing (9). Among 33 models exceeding 10²⁵ FLOPs, 326 are language models and 86 multimodal. China's DeepSeek V3 outperforms Llama 3.1 405B using ~10x less compute.

4. Google Issues AI Coding Guidelines: 30% Code Now AI-Generated, Efficiency Up 10%

#Programming #Google #Productivity

Google releases formal "AI Coding Best Practices" compiled by engineers, advocating AI-assisted coding/debugging and enhanced security reviews. CEO Sundar Pichai reveals 30% of code is AI-generated (up 5% in six months), with systematic AI adoption accelerating delivery by 10%.

5. Israel's Bright Data Launches AI Platform to Challenge Meta/Musk Data Dominance

#Data #Legal #Startup

After courtroom victories against Meta and X, Bright Data unveils a $100M AI platform (Deep Lookup, Browser.ai) providing unrestricted real-time web data access. Its CEO argues data access—not model intelligence—is the critical bottleneck for AI development.

6. SciArena for Scientists Goes Open-Source: o3 Tops, DeepSeek Shines as Open Dark Horse

#Science #Benchmark #OpenSource

Yale/NYU/AI2's SciArena evaluated 13K+ scientific literature tasks by 102 researchers. OpenAI's o3 leads Claude-4-Opus and Gemini-2.5-Pro, while open-source Deepseek-R1-0528 surpasses proprietary rivals.

7. OpenAI Denies Robinhood Partnership: Tokenized Shares "Not Equity"

#Crypto #Legal #OpenAI

OpenAI disavows Robinhood's European tokenized stock offering, clarifying these "OpenAI tokens" lack equity status and require company approval for any legitimate transfer.

8. Microsoft Adjusts AI Chip Roadmap: Conservative Designs Through 2028 to Reduce NVIDIA Reliance

#Semiconductor #Microsoft #NVIDIA

Delaying Maia 200 to 2026, Microsoft plans incremental AI chips (Maia 280 with 20-30% efficiency gains over NVIDIA's 2027 offerings) through 2028. Third-gen Clea chips postponed beyond 2028.

9. Microsoft Confirms 9,100 Layoffs (4%), Xbox Shifts Focus to Growth Areas

#Microsoft #Gaming #Layoffs

Xbox head Phil Spencer announces strategic cuts to "focus on growth areas," following 6K+ layoffs two months prior. The moves reflect broader Microsoft flattening of management hierarchies.

10. Neuralink Theory Gains Support: Intelligence Is Collective, Consciousness May Transcend Brain

#Neuroscience #ElonMusk #Futurism

Scientist Michael Levin proposes intelligence is inherently collective, with consciousness potentially extending beyond brains. His translation interface research—connecting humans, robots, and AI to shared cognition—lends theoretical backing to Musk's Neuralink ambitions.