Large Language Model (LLM) inference presents unique challenges due to its autoregressive decoding phase, fundamentally differing from training.

The primary bottlenecks are identified as memory and interconnect rather than compute. Research opportunities include High Bandwidth Flash for a tenfold increase in memory capacity with sustained bandwidth, Processing-Near-Memory and 3D memory-logic stacking for high memory bandwidth, and low-latency interconnects to accelerate communication.

While the focus is on datacenter AI, the applicability to mobile devices is also being reviewed.

👉 Read the full article on arXiv

Multimodal Language Models (MLMs) excel at semantic vision-language tasks but struggle with spatial reasoning requiring another agent's perspective, indicating a persistent egocentric bias.

Inspired by human spatial cognition, this research introduces perspective tokens, specialized embeddings that encode orientation through either embodied body-keypoint cues or abstract representations supporting mental rotation.

Integrating these tokens into LLaVA-1.5-13B improves performance on level-2 visual perspective-taking tasks. Across synthetic and naturalistic benchmarks, perspective tokens enhance accuracy, with rotation-based tokens generalizing to non-human reference agents. Representational analyses suggest that MLMs contain precursors of allocentric reasoning but lack appropriate internal structure, indicating that embedding cognitively grounded spatial structure directly into token space is a lightweight, model-agnostic mechanism for perspective-taking.

👉 Read the full article on arXiv

Apple is planning a significant overhaul of Siri, transforming it into a ChatGPT-like AI chatbot to compete in the evolving AI landscape. Codenamed 'Campos,' this upgrade is slated for integration into iOS 27, iPadOS 27, and macOS 27, replacing the current Siri.

The revamped Siri will feature natural language conversational capabilities similar to ChatGPT, accessible via voice or text. It is expected to perform a wide range of tasks, including web searches, content generation (including images), coding assistance, summarizing information, and analyzing uploaded files. Furthermore, it may be able to access personal data on the device to complete tasks, recognize on-screen content, and adjust device settings.

The chatbot is anticipated to be powered by a custom model based on Google's Gemini. Apple is reportedly considering privacy measures, such as limiting the memory of past user conversations. This strategic move aims to leverage Apple's platform ownership and provide a more capable AI experience, addressing previous criticisms of Siri's limitations.

👉 Read the full article on MacRumors