Topic
Explore English articles about AI Agents.
NVIDIA's Groq 3 LPX, now in mass production, splits AI inference work between GPUs and LPUs to reach 3,431 output tokens per second. Here's what that number means, where the benchmark falls short, and what's behind the $17 billion deal.
A peer-reviewed study testing 10 AI models found that majority-following behavior in AI agents varies with group size and model. The 1,000-agent figure is not evidence of complex cooperation but the upper limit tested in a binary-choice experiment.
Alibaba is accelerating its AI infrastructure investment and aims to shorten the payback period to 2.5 years by expanding the use of self-developed chips. Amid U.S. export restrictions, the company is strengthening its own silicon foundation to boost profit margins, though winning external customers and navigating international regulations remain key challenges ahead.
NVIDIA says Claude Opus 5, running on its proprietary AVO agent framework, achieved a perfect score on the public ARC-AGI-3 benchmark, versus roughly 30% for the model alone under different measurement conditions. We examine how this points to a structural shift in which harness design—not the underlying model—increasingly determines real-world AI performance, and why it also serves NVIDIA's push to expand GPU demand.
As AI chip TDPs exceed 1 kW and liquid-cooling adoption jumps from 6% to 53% in three years, thermal management has replaced raw performance as the top constraint in semiconductor design. From imec's STCO research to Samsung's zHBM concept and KAIST's AI design agents, here's a look at the front lines of the battle against heat.
Go 1.27 introduces generic methods for concrete types while excluding interfaces and reflection. Teams adopting it should also check how well their IDEs and analysis tools support the new syntax.
A paper titled "Lost in Compaction" from Pennsylvania State University researchers demonstrates via the COMPINT benchmark that AI context compaction causes an average of 83% of user instructions to be lost. The study also points out that Anthropic's official documentation does not guarantee automatic instruction retention.
Examining materials claiming Qwen3.8-27B inference at 30+ tokens/s on a hypothetical 64-core XuanTie C950 configuration. We unpack the SHL optimizations behind the performance, the undisclosed test conditions, and the distance to mass production.
Gartner forecasts that provider-side AI inference costs per agentic workflow will more than quintuple by 2028, shifting the focus from per-unit pricing to cost-per-outcome management.
DeepSeek has released an execution framework for AI agents that makes everything from the model layer to the UI swappable. Cordis's dynamic attach/detach mechanism and append-only logs shift operations from fixed products to composable parts.
