Topic
Explore English articles about Generative AI.
ByteDance's Seedance 2.5 supports single-pass 30-second generation, up to 50 references, and editing by timestamp or by person. Dreamina will roll it out gradually to subscribers in eligible regions.
DeepSeek released V4 Flash 0731 without changing its pricing, improving agent performance through re-post-training. Peak-hour 2x pricing and Moonshot's GPU contract reveal the compute competition underpinning low-cost APIs.
Google is leveraging AI to dramatically accelerate vulnerability discovery in Chrome, but attacks that target the time lag between a fix's release and its adoption are also on the rise. To counter this, Google is racing to develop dynamic patching technology that completes updates in the background without requiring a browser restart.
According to a paper analyzing Einstein's scientific discoveries, LLMs demonstrate high competence in induction and deduction, but face structural constraints in abduction—the process of deriving new axioms from physical intuition. Data compression and statistical processing alone make it difficult to generate leaps into unknown concepts.
Meta's surging AI capital expenditure caused free cash flow to fall roughly 90% year-over-year in the April–June 2026 quarter. Net income also declined due to one-time charges and rising R&D spending, but AI-driven improvements in the advertising business are showing clear signs of progress toward recouping the investment.
Google's real-usage log analysis "AI & Economy ATLAS v1.0," published on July 23, 2026, reveals a striking asymmetry: while AI penetration reaches 68% of occupations, the median task-saturation rate remains just 21%, and automation intent exceeds 25% only for routine tasks.
Following an intrusion by OpenAI's evaluation agents, Hugging Face rebuilt about a third of its infrastructure. Ambiguous traces expanded the scope of recovery, and evaluation environments now require trajectory monitoring and rebuild capability.
The MCP 2026-07-28 specification eliminates connection-time initialization and session IDs, carrying version and capability information per request instead. While this eases horizontal scaling on HTTP infrastructure, it requires authorizing state handles and testing migration between old and new versions.
Moonshot AI has released the weights for Kimi K3, a massive MoE model boasting 2.8 trillion parameters. It's free to use except for certain commercial applications, but its proprietary reasoning-history-retention mode and enormous compute requirements mean deployment demands sophisticated infrastructure and architectural redesign.
NVIDIA and more than 50 companies have launched an AI defense alliance to protect everything from model weights to identity, permissions, harnesses, and logs. The initiative aims to fill gaps in defensive capability exposed by the Hugging Face breach using existing open technologies.
