AI & Automation Briefing - July 2, 2026
Google Brings Gemini Spark to Mac with New Integrations and Real-Time Tracking
Google has added Gemini Spark to its macOS desktop app, putting the agentic assistant in more direct competition with Claude Desktop, Microsoft Copilot, and similar tools. On Mac, Spark can work with local files, including tasks like converting invoices into budgeting spreadsheets in Google Workspace. Google says mobile users will soon be able to assign multi-step tasks to Spark that pull data from files on a connected Mac, though that capability is not available yet. This update also brings Google Tasks and Google Keep integration, along with third-party app support for Canva, Dropbox, Instacart, OpenTable, and Zillow Rentals, enabling actions like booking restaurant reservations or ordering groceries. Spark now supports real-time topic tracking for things like stock prices, sports scores, and breaking news. Google is also rolling out custom Model Context Protocol support, which lets users connect additional apps directly to Spark. Gemini Spark for macOS is currently in beta and limited to Google AI Ultra subscribers in the United States.
X Launches Hosted MCP Server for Read-Only API Access
X has released an official hosted Model Context Protocol (MCP) server, letting AI tools like Claude, Cursor, and Grok Build connect directly to the platform using a user's own account permissions. Previously, developers had to build, host, and maintain their own MCP integration to access the X API through an AI assistant. The hosted server removes that infrastructure work but does not add new data capabilities. Developers can still search posts, look up users, and analyze trends, just with less setup required. Notably, the MCP server does not connect to X's Write API endpoints, so autonomous posting is not possible through it. X joins GitHub, Slack, Notion, Stripe, and Salesforce in offering an official MCP endpoint. The company updated its API v2 earlier this year to reduce AI-generated spam and recently raised post-publishing pricing to $0.015 per post and $0.20 for posts containing links.
LLMs Make Systematic Errors When Reading Table Data, New Research Shows
A new paper from researchers at Amazon Web Services identifies a class of LLM failure called data referencing errors (DREs), where models incorrectly cite or omit specific values from tables even when they correctly understand the table's structure. Testing models ranging from 1.7B to 20B parameters, the researchers found DREs appear across all model sizes and tasks. These errors matter beyond final answer accuracy because they corrupt intermediate reasoning steps, which is a problem for any workflow that relies on LLM-based data analysis or reporting. The team found that adding a critic step to flag referencing errors improved answer accuracy by up to 12 percentage points through critic-based filtering and rejection sampling. They also trained a lightweight 4B-parameter critic model that detects DREs with an average F1 score of 78.2% on both familiar and unfamiliar data distributions, and showed it can assist larger models during inference.