Samsung Electronics has announced the deployment of ChatGPT Enterprise and Codex to all employees in its DX division globally, including those in South Korea, covering technical and non-technical roles in software development, marketing, product R&D, and manufacturing. This marks one of the largest enterprise deployments by OpenAI to date. While the collaboration initially focused on supplying memory semiconductors for AI infrastructure, it has now expanded to enterprise-wide workforce transformation—signaling that this global tech and manufacturing giant views AI as a core platform for enhancing how its global employees work and innovate, rather than a tool confined to specific teams. The deepening strategic alliance reflects the broader trend of large enterprises fully embracing AI.

Machine learning pioneer Michael I. Jordan sharply criticized the current 「thought leader」 culture in AI, arguing that AGI is merely a public relations term that distorts research directions and misleads young people. He advocates redefining AI research through a triad framework of statistics, economics, and computer science, emphasizing that current AI narratives are too narrow—focusing solely on individual cognition while ignoring the social and collective nature of human intelligence. Jordan notes that foundational models exhibit the greatest deviations at knowledge frontiers without providing uncertainty estimates, and systems like AlphaFold produce severely inaccurate predictions on frontier scientific problems with scarce data. He criticizes polarized public discourses around utopian or apocalyptic scenarios for lacking economic reasoning and depriving young people of constructive role models.

According to first-half 2026 funding data, the embodied intelligence 「brain」 sector is experiencing intense capital enthusiasm, raising approximately 43.8 billion yuan, over half of which went to brain-centric companies. The average Pre-A round reached 700 million yuan and B rounds averaged 2.25 billion yuan—approaching C/D round sizes in other fields—with leading companies valued at up to 20 billion yuan. Nearly 80% of funded startups are working on world models, though Fei-Fei Li points out the term is widely abused and lacks industry consensus; most companies combine world models with Vision-Language-Action (VLA) systems rather than treating them as alternatives. About half of the founders come from universities (mostly Tsinghua), while others hail from autonomous driving firms such as Horizon Robotics, Huawei, and Baidu. Young entrepreneurs are particularly favored by investors. Investment dynamics have reversed, with star companies now leading and commanding higher valuations, yet generalization capabilities and business models remain unsettled, signaling significant bubble risks.

The Hoover Institution and Stanford HAI updated their white paper, tracking the career trajectories of 356 researchers behind seven core DeepSeek papers. Among the 271 researchers with institutional records, 145 had never worked at institutions outside China, yielding a purely domestic training rate of 53.5%. Even within the 31-member core team, 10 were entirely domestically trained—indicating that China’s local AI talent pipeline can now support cutting-edge model development. The 80 researchers with U.S. institutional experience achieved the highest academic impact (average 4,108 citations), but the most common path was 「China-U.S.-China」 (38.8%), with only 12.5% remaining in the U.S., reflecting America’s talent drain. The team exhibits a two-tier structure: a stable core with rapidly expanding periphery.

Meta’s internal morale has hit rock bottom following layoffs and AI-driven restructuring. An employee cursed at AI executives during an internal livestream, and Zuckerberg’s proposed AI hackathon was met with boos and resistance. CTO Bosworth admitted the reorganization was 「a mess」 and overlooked employee sentiment, promising improved communication, tighter management spans, and increased budgets for snacks and travel. However, around 6,500 employees were forcibly transferred into application AI teams to perform mechanical tasks like generating programming problems for model training, feeling their professional expertise is being wasted—a situation described as 「labor camp-style life.」 Employees broadly perceive management’s soft measures, such as hackathons and snacks, as fundamentally misaligned with the harsh realities of layoffs, forced reassignments, and repetitive labor.

Researchers introduced MonitoringBench, a benchmark containing 2,644 successful attack trajectories categorized by difficulty, designed to evaluate the effectiveness of monitoring systems for coding agents. The study found that decomposed red-teaming—separately optimizing strategy generation, execution, and refinement—produces stronger attacks than direct prompting, with post-hoc refinement being the most effective single technique, reducing detection rates by up to 40% even against the strongest monitors and showing strong generalization to unseen monitors. The paper proposes a three-axis attack taxonomy (technical, structural, evasion) to mitigate pattern collapse and classifies monitor failures into four types: undetected, partially detected, persuaded, and calibration failure. The authors emphasize that reporting only aggregate capture rates is insufficient for meaningful evaluation of monitors.

An AI weekly report revealed that DeepSeek has completed its first fundraising round, achieving a valuation of 400 billion yuan and raising 5.1 billion yuan under a unique structure: external investors must join as limited partners, with a five-year lock-in period and no voting rights. Only state-owned capital holds direct equity, reflecting founder Liang Wenfeng’s caution toward external funding. Concurrently, ByteDance’s AI strategy is shifting from consumer-facing to B2B; while DouBao generates less than 1 million yuan daily, Seedance achieves a 2-billion-dollar annualized revenue with a 70% gross margin. SK Hynix has eliminated educational requirements, hiring based on actual ability amid fierce competition in the HBM market, sparking concerns among small and medium enterprises about engineer attrition. Nvidia employees transitioning to civil service jobs sparked online debate, with Jensen Huang stating he strives to offer high salaries to retain talent.

Anthropic demonstrated Claude Cowork’s meeting preparation and scheduled task features. It consolidates context from calendar, Slack, and email into a unified pre-meeting workflow, automatically pulling in attendees, surfacing relevant Slack threads, reviewing past meeting notes, and generating formatted agendas that match existing folder structures. New information sources can be added mid-task without restarting the session. Background scheduled tasks can run hourly, daily, on weekdays, or manually, monitoring shared drives to attribute file edits by contributor and generate client-grouped summaries. However, scheduled tasks require the desktop app to remain open and the computer awake, with each run treated as a separate session. The system follows a human-in-the-loop handoff model: Claude handles preparation, but final documents belong to the user.

Logic Intelligence’s SE-Bridge-TTS system, accepted at ICML 2026, investigates the impact of synthetic data ratio on low-resource text-to-speech (TTS). The study finds that more synthetic data is not always better—beyond a 50% threshold, 「synthetic erosion」 occurs: although word error rate (WER) continues to decline, naturalness (MOS), speaker similarity, and token entropy significantly degrade while repetition rates rise. The paper introduces two preference alignment frameworks: DGSA leverages internal prosody-timbre disentanglement in the model, using a small set of real speech as positive references and applying dual-objective DPO to correct pronunciation errors and flat prosody; TDSC enables self-correction in extremely low-resource settings with almost no real speech anchors via multi-temperature sampling and self-filtering. On FLEURS Lao and Thai cross-lingual evaluations, the system achieved 83.4% accuracy, surpassing Higgs Audio v3 (78.2%).