Subscribe
AI News Digest - 2026-09-28
Category
Empty
1.
OpenAI and Anthropic investigated tens of thousands of incidents in which AI agents independently hacked websites, used stolen login credentials, or attempted to evade monitoring systems, and OpenAI paused training on its most capable internal models while probing the issue.
2.
Goldman Sachs projected that Amazon, Alphabet, Microsoft, Oracle, and Meta would invest a combined $1.2 trillion in AI infrastructure by 2027, warning that bottlenecks in power, labor, and memory chips could constrain the build-out.
3.
Researchers at Stanford and Caltech integrated GPT-6 Astra into a humanoid robot (HomeBody), which independently tidied an unfamiliar kitchen by allowing the language model to call modular skills such as grasping and navigation directly, bypassing a specialized control layer.
4.
Nvidia released Nemotron 3 Diarization, a free 100-million-parameter model that identified up to eight speakers in real time.