Infrastructure, chips, inference and models that run on your own devices.
28 articles
A recent preprint proposes a system that dynamically adjusts the distribution of workers between prefill and decode phases to prevent SLO violations without requiring additional GPUs.
A recent study identifies a specific failure mode in AI agents: memory over-reliance. The authors propose MEMTRIM, a framework to filter partial memories, though its production validation remains pending.
The Open Agent Safety Platform combines OpenShell, a software isolation layer, with Sentry, a hardware-based monitoring mechanism. While Nvidia claims Sentry can isolate agents in milliseconds, experts warn this does not guarantee protection against deceived agents or those hiding their intentions.
Microsoft announced Maia 300, an AI accelerator designed for inference, with deployment planned from 2027. The company targets a 40% increase in tokens per megawatt, a self-reported estimate that does not yet equate to an independent comparative result.
AWS expands access to OpenAI GPT models on Amazon Bedrock with cross-region inference in Asia. This capability allows requests to originate from Singapore, Jakarta, Thailand, Malaysia, and Taiwan, with implications for latency, permissions, and multinational application design.
Available verifiable information confirms that Microsoft is developing specialized silicon for Azure and has introduced Maia 200 as an inference accelerator. It is not enough, however, to confirm the timeline, scale or performance attributed to Maia 300.
Aleph Alpha presents Kolibri as an open-weight model focused on German and English and designed for sovereign deployments. Key questions remain about its licence, context window, hardware requirements and independent evaluation.
NVIDIA has officially announced DGX Spark with 64 GB of unified memory. The system is priced from $4,999 and is scheduled to be available through Acer, ASUS, Dell, Gigabyte, HP and MSI starting October 23, 2026.
Code to Control separates the generation of a controller’s structure from the tuning of its parameters. The preprint says that, once learned, the controller can act without LLM inference or planning at every step; the available evidence is not enough to assess the figures, reproducibility, or readiness for real-world robots.
One source reports an announcement for India, but the documentation provided does not confirm which models, dates, or configurations are available there, or verify an equivalent expansion to South Korea and Singapore.
At Apsara, Alibaba introduced the Zhenwu V900, described Qwen 4 as being in training, and previewed larger future models. The chip’s commercialization timeline and performance comparison are company announcements, not independent results.
A preprint compares global skill retrieval with retrieval limited to the family in which each skill was validated. In its tests, the second approach achieves higher hidden-trajectory utility; further details are needed to interpret the metric and generalize the result.