TLDR AI 2025-10-13

30% of employees are secretly using AI, and while it may make them more productive, unsanctioned AI use creates serious data and compliance risks.

👉 What you can do: Use proven playbooks for responsible AI rollout. Use the AI Governance Kit as your starting point, with ready-to-use templates including:

• A step-by-step, long-term governance plan
• An internal AI use policy template
• A scoring matrix to prioritize high-impact use cases
• Tools to build a program that improves over time

Download the Enterprise AI Governance Kit from you.com

He had previously turned down an offer from Meta for $1.5 billion over six years, and Thinking Machines CEO Mira Murati had declined an acquisition offer for an undisclosed amount.

Veo 3.1 marks a noticeable step forward in prompt fidelity and visual/audio quality. It doesn't have the same issues as Veo 3, like occasional oddities with object proportions. The model demonstrates a better understanding of nuance, creating videos that closely match prompt intent. Traces of the model have appeared in Vertex AI and Google Vids. An official release is likely in the coming weeks.

xAI has joined Meta and Google in the race to develop AI systems that can navigate and design physical environments. The company is building world models with the goal of applying them in gaming. These models could be used to generate interactive 3D environments and unlock uses for AI beyond software and computers in physical products. Elon Musk announced on X that xAI would release a 'great' AI-generated game before the end of next year.

Not yet, but it's close. GenAI is a demand-led, capital-intensive boom with revenues doubling annually and expected to hit $100 billion by 2026. Capital expenditure to revenue ratios (six times) are high, exceeding both railroads (two times) and telecom (four times) at their peaks, but there are other indicators: economic strain above 2% of GDP, industry strain with capex over 10x revenue, revenue growth slowing below 50% annually, price-to-earnings ratios hitting 50-60, or internal cash covering under 25% of capex. Until two of these are met, the current risk, while large, is manageable.

Embeddings are a fundamental component of every modern retrieval augmented generation system. This post shows what computations are required to produce an embedding and how this can be used to calculate the true dollar cost of processing a token. The cost of processing a token is minuscule, and all embedding models converge to similar semantic representations, so no provider really enjoys a sustainable moat. This results in prices racing to the bottom as the underlying low costs are being passed directly on to consumers.

OpenRouter gives you a single set of endpoints to run ALL the major models: GPT, Claude, Gemini, Llama, Mistral, and more. Swap models at request level, route across providers by price or latency, and have fallback options to survive outages. Get started in 30 seconds

Meta has introduced "early experience," a training approach using data from an agent's own interactions without external rewards. It improves policy learning via implicit world modeling and self-reflection.

Anthropic has introduced plugin support in Claude Code, allowing users to install custom slash commands, agents, MCP servers, and hooks with a single command. These modular extensions simplify setup sharing and enable flexible customizations for development workflows.

ReasoningBank is a novel memory framework that distills generalizable reasoning strategies from an agent's self-judged successful and failed experiences. Agents can store and retrieve memories from ReasoningBank to inform their decisions. They can integrate new learnings and become more capable over time. The better memory enables more effective scaling. ReasoningBank consistently outperforms existing memory mechanisms that store raw trajectories or only successful task routines across web browsing and software engineering benchmarks.

InferenceMAX continually re-benchmarks the world's most popular open-source inference frameworks and models to track real performance in real time. It captures progress in near real-time. A live dashboard is available for free.

A market scan of 110 AI companion platforms reveals stark differences: 52% are US-based, targeting young men with hypersexualized anime AI girlfriends monetized through NSFW premium features. Chinese platforms target adult women aged 25-40 with dynamic AI boyfriends featuring card-collecting mechanics and WeChat integrations, shaped by China's marriage crisis (20% drop in 2024) and gender imbalance.

OpenAI just launched apps inside ChatGPT where you can say "Spotify, make me a playlist" or "DoorDash, order my usual" and it handles everything conversationally. This looks like early iPhone web apps, not true native experiences yet, but the trajectory is clear. As OpenAI expands its SDK to give developers access to user memory, persistent state, and action execution, we'll see actual native ChatGPT applications designed for conversational interfaces.

By 2028, 1.8 billion agents will be in operation across industries. In this paper, IDC analysts provide guidance on selecting the right agent-building platforms, overcoming common deployment challenges, managing safety and compliance concerns, and addressing the AI skills gap. Download the IDC Spotlight

Nathan Calvin, a lawyer shaping the policies around AI technology, says that OpenAI used the pretext of its lawsuit against Elon Musk to intimidate its critics.

OpenAI's top customers are Duolingo, OpenRouter, and Indeed.

OpenRouter has changed the pricing of its 'bring your own key' plan so that every customer gets a million requests per month for free.

Cisco launched the P200 chip to connect long distance AI data centers.

OpenAI spent roughly $7 billion on compute in 2024, with $2 billion going to inference and only $480 million to final training runs for released models like GPT-4.5 and o3.