MiniMax announced the creation of ForgeTrain, the world's first production-grade large model pretraining framework fully written by AI, and successfully used it to train the MiniCPM5-1B on-device model. ForgeTrain outperforms NVIDIA's Megatron, achieving a 10% faster training speed on identical hardware and delivering an additional 10% acceleration on Huawei Ascend platforms. With only 1B parameters, MiniCPM5-1B surpasses all models under 2B in the AA-Index, supports 131K long context and mixed reasoning, and can be deployed across devices ranging from smartphones to servers. MiniMax introduced a new paradigm called Forge Engineering, where AI customizes dedicated code for different models, hardware, and tasks, compressing weeks of manual coding into tens of minutes. The related models and framework have been fully open-sourced, marking the transition of 'AI building AI' from concept to reproducible engineering.

Researchers from UC San Diego published a study in the Proceedings of the National Academy of Sciences (PNAS) conducting Turing tests on four large language models. In 1,023 five-minute text conversations involving 284 participants, GPT-4.5 was mistaken for a human in 73% of cases, officially passing the test. LLaMa-3.1-405B had a 56% identification rate, while older models GPT-4o and ELIZA scored only 21% and 23%, respectively. The study highlights that although AI is nearly indistinguishable from humans in short interactions, it raises concerns about 'fabricated humanity,' potentially affecting online trust, social dynamics, and employment. Researchers also questioned whether the Turing Test remains a valid benchmark for machine intelligence.

Google DeepMind released AlphaProof Nexus, a system combining Gemini 3.1 Pro with Lean formal verification, which successfully solved nine long-standing Erdős mathematical problems, two of which had remained unsolved for 56 years. The system also proved 44 integer sequence conjectures, resolved a 15-year-old algebraic geometry problem, and discovered new optimization theory parameters. All proofs were machine-verified for logical rigor, with each costing only hundreds of dollars in compute. Despite a success rate of approximately 2.5% across 353 attempts, the methodological implications are significant: AI transitions from a 'black-box oracle' to a traceable research collaborator. DeepMind CEO Hassabis emphasized the system is still 'far from' achieving AGI.

Reports indicate Anthropic completed over $30 billion in funding between May 26 and 27, reaching a $900 billion valuation and surpassing OpenAI to become the world's most valuable AI startup. Concurrently, Anthropic is preparing to expand commercial access to its most advanced model, Claude Mythos, via the Claude Code platform, launching a new version called Mythos 1. The previously initiated Project Glasswing has partnered with over 40 companies including AWS, Microsoft, and Google, identifying more than 10,000 high-risk vulnerabilities through Mythos. The Claude Security platform is being upgraded into a full vulnerability management tool, having already detected 1,596 vulnerabilities across 281 open-source projects.

Samsung Electronics announced it will fully roll out external generative AI services—including OpenAI’s ChatGPT, Google Gemini, and Anthropic’s Claude—across its Device Experience (DX) division starting in June. The company is adopting a dual-track strategy, combining its proprietary Samsung Gauss with external AI tools to enhance productivity across product planning, development, and marketing. A proof-of-concept trial conducted from April to May involved 2,500 employees, and a governance mechanism now requires staff to complete internal security training before using external AI. Samsung also plans to transform all global manufacturing sites into AI-driven smart factories by 2030, gradually deploying humanoid manufacturing robots.

Together AI launched and open-sourced OSCAR (Offline Spectral Covariance-Aware Rotation), a 2-bit KV cache quantization system for long-context LLM serving. OSCAR generates data-aware rotation matrices via attention-aware offline calibration, steering quantization errors into attention-insensitive directions, achieving near-BF16 accuracy at approximately 2.28 effective bits. Tested on models like Qwen3 and GLM-4.7-FP8, it shows minimal average accuracy loss under 32K context, up to 3x higher decoding throughput, and up to 7.83x job-level throughput in batch scenarios, reducing overall KV memory by about 8x. It has been integrated into the SGLang inference framework and supports paged attention and prefix caching.

AI inference routing platform OpenRouter announced a $113 million Series B round led by Alphabet’s CapitalG, with participation from Nvidia NVentures, ServiceNow Ventures, and MongoDB Ventures. The company provides a unified API enabling enterprises to access hundreds of AI models, with features including intelligent routing, cost control, and policy management. Its weekly processing volume has reached 25 trillion tokens—five times higher than six months ago—and serves over 8 million users globally. Its publicly available model usage data and rankings have become industry reference benchmarks. New funds will be used to expand routing, governance, and optimization capabilities.

Microsoft AI CEO Mustafa Suleyman unveiled the next-generation text-to-image model MAI-Image-2.5, which debuted third on the Arena text-to-image leaderboard with a score of 1,254—72 points higher than its predecessor. Microsoft AI becomes the newest lab, after Google DeepMind and OpenAI, to enter the top five in this domain. The new model excels in visual reasoning, delivering more accurate handling of object relationships, scene structure, lighting, proportions, and spatial relations, with notable improvements in text clarity and brand visual representation. Users can already try it on Arena.ai, with availability expected on MAI Playground and Microsoft Foundry next week.