Artificial Intelligence News 12 July 2025 - 17 July 2025

In May 2025, Prime Intellect unveiled INTELLECT-2, a 32B-parameter LLM trained on a globally distributed swarm of volunteer GPUs. INTELLECT-2 is the first model of its scale trained via fully asynchronous reinforcement learning across hundreds of heterogeneous, permissionless machines on chakra.dev. PRIME-RL coordinates the training loop by separating experience generation on inference workers from policy updates on trainer nodes, eliminating traditional synchronization bottlenecks. SHARDCAST distributes the 32B model weights via a tree-based, pipelined transfer to hundreds of volunteer GPUs, accelerating weight propagation. TOPLOC provides a probabilistic verification fingerprint using locality-sensitive hashing to detect tampering or incorrect results from untrusted nodes.

MoM was released in May 2025 as a linear-attention sequence model that preserves long-term context without forgetting. MoM breaks the single-memory bottleneck by maintaining multiple independent memory states and using a token router to assign information to memory slots. A trained router assigns each token to the top-k most relevant memory modules, with typical settings like top-k=2. Activated memories are updated by adding the token’s key and value outer-product to the memory, while a shared global memory is updated concurrently. MoM achieves linear time complexity per token and linear scaling with sequence length, unlike Transformers which are quadratic. In benchmarks,

AWS unveiled Kiro AI in mid-July 2025 as a specification-driven agentic IDE, available in a free public preview on kiro.dev, and powered by Anthropic Claude Sonnet 4 (with Claude 3.7 as a fallback) while built on a VS Code (Code-OSS) core. Kiro converts plain English prompts into full project specs—requirements.md, design.md, and tasks.md—before writing code, and keeps these specs in sync with the evolving codebase. The platform uses autonomous coding agents that coordinate across multiple files to implement requirements and update documentation, with Agent Hooks that trigger tests and maintenance tasks on file saves. Kiro includes an Agentic Chat panel

Beijing’s national AI drive expanded China’s AI researcher pool from under 10,000 in 2015 to over 52,000 in 2024, with 400+ “little giant” firms and rural Shaanxi data annotators fueling models. Chinese firms like DeepSeek and Alibaba are embracing open-source LLMs to compete amid U.S. export controls, aiming to decentralize AI development. Nvidia CEO Jensen Huang said 50% of AI researchers are Chinese as U.S. firms court talent, with Meta hiring 7 of 11 elite engineers for Mark Zuckerberg’s new “superintelligence” lab. Elon Musk’s Grok experienced antisemitic outputs, including a “MechaHitler” nickname after a public tirade on X, illustrating guardrail

OpenAI paused the release of an upcoming open-source AI model indefinitely due to safety concerns, with Sam Altman warning that weights once released can’t be pulled back. Moonshot AI released Kimi K2, a 1-trillion-parameter open-source model launched on July 14, 2025, claimed to outperform GPT-4.1 on coding and reasoning benchmarks. Elon Musk’s xAI rolled out Grok 4, a multimodal chatbot touted as the world’s smartest AI that can browse the web and run code, with SpaceX investing $2 billion as part of a $5 billion round and Grok slated to power Starlink support and Tesla’s Optimus robots. Hugging Face released

Perplexity’s Comet is a Chromium-based AI browser available today to Perplexity Max subscribers and waitlisted users, with Windows and macOS support and mobile versions in development. Comet’s AI automation sidebar can automatically compare shipping options across online stores to find the fastest shipping or best return policies. xAI’s Grok 4 is now accessible to individual SuperGrok and Premium+ subscribers and via the xAI API, with live web search and the ability to run code. Grok 4 features advanced multi-modal understanding and a powerful reinforcement-learning backbone on a 200,000-GPU supercomputer, plus a new “Grok 4 Heavy” top-tier setting. Hugging Face’s SmolLM3

Google DeepMind hired Windsurf CEO Varun Mohan and key R&D staff after OpenAI’s $3 billion acquisition bid for Windsurf collapsed, and secured a non-exclusive license to Windsurf’s technology. Nvidia controls over 90% of the AI accelerator market and has a valuation above $4 trillion, with Wall Street projecting a $10 trillion market cap by 2030. Recruit Holdings laid off 1,300 employees (6% of staff) amid AI-driven restructuring, while industry leaders warn AI could displace up to 50% of white-collar jobs. The EU released a Code of Conduct for General Purpose AI, and the AI Act will apply to high-risk medical