Researchers have developed a novel approach to self-attention in Transformer models that significantly reduces computational costs. This method achieves constant cost per token, regardless of context length, by leveraging a symmetry-aware Taylor approximation.

The technique decomposes the standard self-attention formulation into tensor product expressions, exploiting their symmetry to map queries and keys to a minimal polynomial-kernel feature basis. This allows for an inverse relationship between cost and head size, enabling more heads per token.

The implementation has been empirically validated, demonstrating its correctness. This breakthrough promises to substantially lower the infrastructure and energy demands of large-scale Transformer models by enabling unbounded token generation at a modest, fixed cost. The mathematical innovations introduced are also considered to be of independent interest.

👉 Read the full article on arXiv

Apple has integrated AI agents into its integrated development environment (IDE), Xcode, with the release of Xcode 26.3. This update brings agentic coding capabilities by supporting agents from Anthropic (Claude Agent) and OpenAI (Codex).

These agents can autonomously perform complex development tasks such as exploring projects, understanding code structure, writing code, building projects, running tests, and iterating on fixes. They operate with transparency, providing step-by-step breakdowns of their actions. The integration utilizes the Model Context Protocol (MCP), allowing any compatible agent to access Xcode's documentation, tools, and project management features.

Apple is actively promoting this new functionality with workshops, signaling a significant push towards democratizing AI-assisted development. This move positions Apple to capitalize on the growing trend of AI-driven software development, potentially impacting the broader software industry.

👉 Read the full article on DIGITIMES Asia / The Tech Buzz / The Neuron / Gadgets 360

Fibr AI has secured additional funding from Accel, signaling strong investor confidence in its ability to transform static websites into personalized, one-to-one experiences using AI agents.

The company's technology leverages AI agents to dynamically adapt website content and interactions based on individual user data and behavior. This allows businesses to create highly tailored user journeys, moving beyond generic web presences to deliver deeply engaging and relevant content.

This investment is expected to accelerate Fibr AI's product development and market expansion, further solidifying the trend of AI agents personalizing digital interactions and enhancing customer engagement.

👉 Read the full article on TechCrunch