AI News — Thursday, July 16, 2026

9
Microsoft Reportedly Training Salespeople to Undermine OpenAI and Anthropic

Microsoft is reportedly instructing its sales teams to downplay competitors OpenAI and Anthropic, indicating intensifying competition in the enterprise AI market.

TechCrunchindustry
8
OpenAI Launches $230 Keyboard for Codex Amid Hardware Legal Battle

OpenAI has released a $230 keyboard designed for its Codex AI, a move that comes amidst an ongoing hardware legal dispute, signaling a potential expansion into AI-specific peripherals.

TechCrunchproduct
8
GPT-Red: Unlocking Self-Improvement for Robustness

OpenAI introduces GPT-Red, a new framework focused on enabling self-improvement in models to enhance their robustness and reliability.

OpenAI Blogresearch
8
Expanding Managed Agents in Gemini API: Background Tasks, Remote MCP, and More

Google announces significant enhancements to the Gemini API, including support for background tasks and remote Managed Control Plane (MCP) for more robust and versatile AI agents.

Google AI Blogproduct
8
SynthDocBench: Controlled Benchmark for Long-Context Visual Document Understanding

SynthDocBench introduces a new controlled benchmark designed to rigorously evaluate the performance of models in long-context visual document understanding.

Hugging Faceresearch
7
The US Advances AI Safety Through State and Federal Action

OpenAI highlights ongoing efforts by both state and federal governments in the US to develop and implement policies aimed at ensuring AI safety.

OpenAI Blogindustry
7
Verifiable AI Inference

A new blog post explores the critical need and methods for ensuring verifiable AI inference, addressing trust and transparency in AI systems.

Lobste.rsresearch
7
Building an AI Agent That Knows When Not to Guess (Qwen + MCP)

This article details a practical approach to developing an AI agent using Qwen and MCP that can identify when it lacks sufficient information, improving reliability.

Dev.toproduct
7
Search Beyond What Can Be Taught: Evolving the Knowledge Boundary in Agentic Visual Generation

This paper explores methods for agentic visual generation to expand its knowledge beyond pre-trained data, pushing the boundaries of what AI can create.

Hugging Faceresearch
6
LLM Evals For Developer Tools: Useful, Correct, Safe

The article discusses the importance of robust evaluation methodologies for Large Language Models (LLMs) used in developer tools to ensure they are useful, correct, and safe.

Dev.toresearch
6
The Chatbot Was Easy. The Engineering Wasn't.

This piece reflects on the challenges of engineering robust and scalable chatbot solutions, emphasizing that the underlying infrastructure is often more complex than the AI model itself.

Dev.toindustry
6
LangSmith vs Traccia: Observe vs Enforce in Production AI Agents

An analysis comparing LangSmith and Traccia highlights their different approaches to managing production AI agents, focusing on observation versus enforcement mechanisms.

Dev.toproduct
6
Type-safe LLM Outputs with Zod: Stop Guessing What the Model Returns

This guide demonstrates how to use Zod for type-safe validation of LLM outputs, ensuring predictable and reliable data structures from AI models.

Dev.toopen-source
5
Post-Mortem: Building a Local MCP Server for Codebase Memory using Ollama and ChromaDB

A post-mortem details the process and lessons learned from building a local Managed Control Plane (MCP) server for codebase memory, leveraging Ollama and ChromaDB.

Dev.toproduct
5
Why Did My Benchmark Stop at N=22? A Debugging Story in Nine Bugs

A detailed debugging story recounts the process of identifying and resolving nine different bugs that caused an AI benchmark to halt prematurely.

Dev.toresearch