TLDR AI 2025-09-03

Is your team exploring MCP integrations?

The Model Context Protocol (MCP) is quickly emerging as the go-to standard for connecting LLMs to external tools and data. But as adoption picks up, many teams are implementing MCP without a clear security playbook.

This practical guide breaks down real-world MCP security risks and offers actionable steps to help teams secure LLM integrations.

Inside the guide:

• Key risks with local and remote MCP servers
• Real-world threats like prompt injection and supply chain compromise
• Actionable steps for safely using MCP tools

Download the guide to get smart on securing MCP as adoption grows.

OpenAI acquired the A/B testing platform to strengthen its product experimentation capabilities as it scales ChatGPT to hundreds of millions of users worldwide.

Since raising at a $61.5 billion valuation in March, Anthropic has experienced explosive growth driven by its success in coding. Annual recurring revenue jumped from $1 billion to $5 billion in 2025, with Claude Code alone generating over $500 million in run-rate revenue.

Mistral has introduced "Memories" in Le Chat, a hybrid memory system designed around user needs for transparency, control, and task focus. The feature surfaces relevant information with visible recall, citations, and editable memory logs.

Expert forecasters and AI domain experts dramatically underestimated AI progress from 2022 to the present across multiple benchmarks, especially in mathematical reasoning. For the International Mathematical Olympiad, superforecasters and AI experts 8.6% expected models wouldn't achieve gold-level performance until 2030, which OpenAI and Google both achieved in July of this year.

This post aims to give a clear sense of direction for those who want to do mechanistic interpretability research. Mechanistic interpretability research is a high-leverage and impactful research area that can be learned on your own with short feedback loops and modest compute. The post guides users through how to learn the ropes and conduct research with projects. It can take a few months to learn, but there's a possibility of a career after becoming competent.

Company knowledge is scattered across countless tools and conversations. Slack's new ebook reveals how AI-powered search transforms how teams discover information—turning endless hunting into instant answers. Learn practical strategies to unlock your organization's collective intelligence and boost productivity across every department.

A study on how large vision language models (VLMs) rely on visual input. It proposes an extended heatmap visualization method to analyze visual token relevance and attention behaviors across different architectures.

Graph Transformers, as the next evolution of Graph Neural Networks (GNNs), are capable of modeling structured data through attention mechanisms rather than traditional message passing. This article highlights how many forms of business data, typically stored in relational tables, can be restructured as graphs to unlock deeper insights and enable more reliable AI models.

AI fluency will be critical by 2025, impacting both performance reviews and hiring processes. The number of tech GTM job postings requiring AI skills has surged, encompassing diverse roles from engineers to marketers. Employers are adopting frameworks and interview techniques to assess candidates' AI proficiency, with AI-centric assignments becoming a key part of the hiring process.

Four Apple employees have departed within the space of a week. Jian Zhang, a lead AI researcher in robotics, left for the Meta Robotics Studio; John Peebles and Nan Du will be headed to OpenAI; and Zhao Meng is joining Anthropic. It is unknown whether Apple is currently hiring. Apple's AI team was a fast-growing team in its own right.

Join open source workshops, Granite LLMs sessions, BeeAI demos, and hands-on content across Gen AI, governance, and hybrid cloud. Register now →

Dojo was supposed to be the cornerstone of Tesla's AI ambitions, but after six years of hype, Tesla has shut down the project and disbanded the team behind the supercomputer.

Ahead-of-time (AoT) compilation can significantly boost performance in ZeroGPU Spaces, which offer temporary access to powerful H200 GPUs.

OpenAI is waging a legal campaign against AI governance nonprofits, searching for secret links to rival billionaires like Elon Musk and Mark Zuckerberg, who've sued the company and poached employees.