GPUs, infra, regulation
The US government launches Genesis Mission, leveraging DOE computing power and federal data to advance 'AI for Science,' drawing comparisons to the Manhattan Project.
The race for longest context and sparse attention heats up. Mechanisms like DSA/NSA reduce long-text inference costs, becoming key themes in papers and products.
ByteDance and NTU research optimizes multimodal models' on-demand search strategy, cutting query volume by 30% while enhancing accuracy. This addresses AI industry extension topics for 2025-2026.
OpenAI reveals governance changes, outlining the for-profit subsidiary's transition to a public benefit corporation (PBC), aligning with its broader restructuring narrative.
Meta released Llama 4 (MoE), upgrading multilingual, multimodal, and long-context capabilities. Controversy followed over evaluation methods and 'special edition' models.
Lightelligence reveals new photonic-electronic hybrid computing hardware, aiming to overcome energy efficiency bottlenecks in matrix operations for AI.
Open-source video generation breakthrough: Domestic teams and the global open-source community advance billion-parameter models and low-cost training, narrowing the gap with closed commercial solutions.
Nvidia's Blackwell architecture enters mass delivery: GB200 systems begin bulk supply. Data center deployment narratives intensify alongside power and cooling constraints.
US and Chinese cloud providers rapidly adopt DeepSeek. Microsoft Azure, Nvidia platforms, and domestic clouds roll out hosting and access capabilities.
US AI chip stocks saw wild swings as markets debated whether efficient open-source models erode Nvidia's moat, leading to historic single-day market cap losses followed by a recovery.