meta launches apple while openai expands its platform

Key Takeaways

Meta records employee keystrokes to train AI models

Meta plans to use data from its own employees to train its artificial intelligence systems. The company will launch an internal tool called the Model Capability Initiative to capture mouse movements, keystrokes, and screenshots on work computers. Meta states this data helps AI agents learn how humans actually use computers for everyday tasks. The company assures employees that sensitive content will be protected and the data will not be used for performance evaluations.

Meta MCI tracks employee actions for AI training

Meta's Model Capability Initiative now tracks mouse clicks, keystrokes, and screen snapshots to improve AI capabilities. Some employees express discomfort with the constant monitoring, though Meta promises the data remains anonymous and is not used for disciplinary actions. The initiative is part of Meta's broader goal to build more sophisticated AI models that can replicate human behavior. Despite assurances, some staff members remain skeptical and call for more transparency regarding how the data will be used.

Meta ramps up workplace tracking for AI training

Meta is installing new software on U.S. employee computers to capture mouse movements and keystrokes for AI training purposes. The Model Capability Initiative will run on work-related apps and websites to help AI agents duplicate human behavior. Meta CTO Andrew Bosworth stated the goal is for AI agents to perform most of the work while humans supervise. While Meta claims this is not increased surveillance, some experts worry about the blurring lines between voluntary data sharing and business data collection.

Meta tracks employee computer activity for AI agents

Meta's Model Capability Initiative records mouse activity, keystrokes, and screenshots to train its AI agents. Meta CTO Andrew Bosworth announced plans to increase internal data collection to help agents replicate human work better. A spokesperson confirmed the data will not be used for performance assessments and that safeguards protect sensitive content. Meta aims to refine its models so they can handle tasks like navigating dropdown menus and using keyboard shortcuts more effectively.

Meta monitors employee inputs for AI training

Meta plans to monitor employee mouse movements and keystrokes through a tool called the Model Capability Initiative. CTO Andrew Bosworth shared a memo stating the goal is to make AI agents better at replicating human work. Meta assured employees that the collected data will not be used in performance reviews. The initiative aims to help AI agents improve at tasks they currently struggle with, such as using keyboard shortcuts.

Meta tracks worker keystrokes for AI training

Business Insider reports that Meta sent an email to employees about starting to track their computer interactions for artificial intelligence training. The company will monitor keystrokes and mouse movements to help its AI agents work more efficiently. Meta stated that privacy protections are in place and the data will only be used for the stated purpose of training models. Employees have expressed discomfort with the new monitoring tool.

Meta tracks employee data for agentic AI training

Meta introduced a new monitoring tool that tracks keystrokes and mouse movements of U.S. workers to generate data for AI models. Many employees expressed discomfort and asked how to opt out, but Meta confirmed there is no option to opt out of the Model Capability Initiative. A spokesperson stated that privacy protections exist and the data will only be used for training. The company hopes these models will help AI agents work more efficiently.

Meta employees unhappy with new AI tracking

Meta installed software on U.S. employee computers to track mouse movements, clicks, and keystrokes for AI training. A memo stated the tool applies to full-time employees and contingent workers. Employees reacted negatively, with many expressing discomfort and asking how to opt out. Meta CTO Andrew Bosworth confirmed there is no option to opt out of the tracking on work laptops. The tracking is restricted to a pre-approved list of work-related applications.

Meta logs employee keystrokes for AI agent training

Meta told U.S. employees it will collect mouse activity and keystrokes on work systems to train artificial intelligence models. The Model Capability Initiative runs across work apps and websites to record clicks, cursor movement, and typing. The system can also capture periodic screenshots to improve areas where models struggle with routine computer behavior. Meta CTO Andrew Bosworth described a broader program called the Agent Transformation Accelerator focused on AI agents doing more of the work.

HP and Lenovo race for AI dominance

HP and Lenovo are launching new AI assistants to compete in the market. Lenovo's Qira assistant allows conversations to continue across devices and offers features like cross-device summaries and real-time transcription. HP's HP IQ aims to provide a universal experience across all its devices with a common user interface. Both companies face challenges because they do not own complete ecosystems like Apple or Samsung, which limits their ability to integrate AI across all user devices.

HP and Lenovo struggle with AI ecosystem gaps

HP and Lenovo are betting on on-device AI assistants but face significant challenges compared to true AI players. Lenovo's Qira and HP's HP IQ aim to provide a universal experience, but neither company owns the full ecosystem needed for seamless integration. Lenovo's Motorola line lags behind Apple and Samsung in the U.S., while HP does not make phones. Experts worry that advanced AI models will likely run in the cloud, reducing the benefits of on-device processing for tasks requiring external data.

OpenAI speeds up AI agents with WebSockets

OpenAI is using WebSockets to make its Responses API 40% faster for AI agent workflows. The change eliminates unnecessary network hops and allows for persistent connections that cache conversation state. This optimization helps the new GPT-5.3-Codex-Spark model reach speeds of over 1,000 tokens per second. Companies like Vercel and Cursor reported substantial latency reductions after the rollout.

OpenAI cuts API latency using WebSockets

OpenAI is accelerating AI agent workflows by adopting WebSockets for its Responses API. This move addresses the bottleneck caused by traditional sequential API requests that add up to minutes of waiting time. The new approach uses persistent connections to cache conversational state and reduce redundant processing. Early prototypes showed up to 40% improvements in agentic workflows, and the production rollout achieved over 1,000 tokens per second.

Llama.cpp auto fit boosts local AI performance

The auto fit feature in llama.cpp allows 70-billion-parameter models to run on consumer hardware with as little as 8GB of RAM. This tool dynamically adjusts quantization parameters to fit models into available memory without manual tweaking. Users report that output quality holds up well against cloud-hosted counterparts on standard reasoning tasks. This development lowers the barrier for running advanced AI locally on devices with limited resources.

Newspapers use AI articles with human bylines

McClatchy Media is using an AI tool called the content scaling agent to create summaries of news stories. Some journalists are being forced to take partial bylines even when the AI writes the article. Unionized publications like the Miami Herald and Sacramento Bee use different byline formats, while non-union outlets sometimes omit the author entirely. A company leader stated that if a journalist cannot remove their name via contract, the company will use it.

Most employees need AI training at work

A recent survey shows that 83% of U.S. job seekers say companies need to formally train employees on AI tools. Only 36% of companies provide a list of approved AI tools, leaving many workers to navigate the technology on their own. Hiring managers agree that formal training should be a priority, with 86% supporting it. The data indicates that AI adoption is outpacing employee readiness and training programs.

Bots outnumber humans online for attention

AI agents wrote more content than humans last year, and this trend is accelerating. Experts discuss the need for robust methods to verify human identity in online interactions. Worldcoin proposes a Proof of Humanity system using iris scans to create a unique digital identity. This approach aims to distinguish humans from bots without revealing personal data. The conversation highlights the risks of unverified interactions in financial transactions and social media.

AI models struggle with memory and learning

Current AI models cannot learn or adapt after deployment because they rely on external memory systems. Researchers are developing continual learning to allow models to update their own parameters based on new experiences. In-context learning helps with existing answers but fails when genuine discovery or tacit knowledge is needed. The goal is to equip models with the ability to compress and internalize new information directly.

Microsoft AutoAdapt fixes LLM adaptation issues

Microsoft's AutoAdapt framework automates the process of adapting large language models for specific domains. It replaces manual trial-and-error with a structured workflow that selects the best adaptation steps like RAG or fine-tuning. The framework includes AutoRefine to optimize hyperparameters efficiently within budget constraints. Microsoft Research is making the framework open source to help teams deploy LLMs more reliably in sectors like healthcare and law.

Senate seeks funding for AI expansion

Senator John Smith is requesting a $2.8 million funding boost to expand the Senate's use of AI technologies. The initiative aims to enhance cybersecurity capabilities and improve efficiency in handling congressional business. The proposed funding will develop AI-driven tools for data analysis, document management, and predictive analytics. Both Democrats and Republicans support the plan as a step toward a more technologically advanced legislative environment.

Sources

• Meta will record employees' keystrokes and use it to train its AI models
• Meta starts recording employee mouse and keyboard actions for AI training
• Meta Updates Employee Tracking Capabilities, Putting AI Training Front and Center
• Now Meta will track what employees do on their computers to train its AI agents
• Meta will track employee mouse movements and keystrokes, report says
• Meta to track workers' keystrokes and mouse movements for AI training, Business Insider reports
• Meta tracks employee keystroke data for agentic AI model training amid privacy furor
• Meta is monitoring employee clicks and keystrokes for AI training, and they aren't happy about it
• Meta Plans To Log Employee Keystrokes For Training Workplace AI Agents: Report
• HP and Lenovo Are Vying for AI Dominance. Maybe They Should Just Stick to Hardware
• HP and Lenovo Are Vying for AI Dominance. Maybe They Should Just Stick to Hardware
• Speeding up agentic workflows with WebSockets in the Responses API
• OpenAI Slashes API Latency with WebSockets
• Llama.cpp’s auto fit feature is quietly reshaping what local AI inference can do on consumer hardware
• Newspaper Company Allegedly Puts Humans' Bylines on AI Articles Unless Contractually Prevented from Doing So
• 8 in 10 Employees Say They Need AI Training -- After Their Companies Already Rolled Out the Tools
• Bots Outnumber Humans Online: The Proof of Human Imperative
• AI's Memory Problem
• AutoAdapt: Microsoft's LLM Adaptation Fix
• Senate eyes AI expansion for congressional business - Live Updates