News, analysis and practical ideas for a fast-moving field.
Google announced Gemini 4 Argon with initial access for trusted cyber defenders. The release separates a capability announcement from broad availability and highlights the engineering demands of longer autonomous workflows.
NVIDIA released Kumo Tabular for classification and regression from labeled tables. Faster setup may be valuable, but data leakage, evaluation splits, and business thresholds remain the central work of deployment.
NVIDIA presented an open agent safety platform spanning software and hardware. Its architecture makes the location of enforcement a central question for teams operating agents with access to sensitive tools and data.
H Company released Holo4 models for computer use across multiple interfaces. The ability to switch between a screen and an API is promising, but consistent state and recovery become central engineering concerns.
Early measurements put Sonnet 5.5 near the top of an intelligence index, with unusually high output-token use. Effort settings and cost per completed task matter more than the model family’s positioning.
Opus 5.5 entered Text Arena at the top of its ranking. The result is useful evidence about comparative response preference, provided readers keep it separate from task completion and production reliability.
A physicist’s account of Claude completing a difficult scattering-amplitude calculation offered a concrete research case. The result is most informative when the task, computational workflow, and independent validation remain distinct.
Cognition reported more than $1 billion in annualized revenue run rate. The milestone signals commercial demand for coding agents, while leaving buyers to evaluate the economics of their own engineering workflows.
