The AI frontier just got more creative, contextual, and computationally lean than ever. OpenAI’s Sora redefines storytelling through reusable AI cameos, Google’s NotebookLM upgrades memory and persona depth, and Moonshot’s Kimi Linear model revolutionizes large-context reasoning, each reshaping how humans and machines co-create.

🎭 OpenAI Sora: Character Cameos Take the Stage – Sora’s new “character cameos” feature lets users craft reusable AI avatars, from pets to personas, for dynamic video creation. With tagging, clip stitching, and creator leaderboards, Sora turns AI video editing into a social, monetizable experience.

🧠 Google NotebookLM: Custom Personas Meet 1M-Token Memory, Google supercharges its research AI with a massive 1M-token context and personalized personas, allowing users to build domain-specific assistants for academia, journalism, and science. Powered by Gemini, it brings deeper recall, insight synthesis, and privacy-first memory per project.

⚡ Kimi Linear 48B: Long Context, Lightning Speed – Moonshot AI’s hybrid attention architecture, Kimi Delta Attention, slashes memory use by 75% and boosts decoding speed up to 6× while handling 1M-token contexts, merging performance and efficiency for long-form reasoning and large-scale AI applications.

From creative avatars to contextual intelligence and lightning-fast models, the new era of AI isn’t just powerful, it’s expressive, enduring, and engineered for imagination.

OpenAI Sora Adds Character Cameos for AI Video Creation

OpenAI's Sora app has introduced a new "character cameos" feature allowing users to create reusable AI avatars from pets, objects, original personas, or other subjects to include in AI-generated videos. This enhancement expands the previous ability to generate deepfake likenesses of people and enables tagging and display name assignment for characters, encouraging creative video production. The update includes video clip stitching and leaderboards showcasing popular content. OpenAI plans to allow monetization for creators through character cameos, with potential revenue-sharing for rights holders of popular characters. The rollout follows legal challenges over the "cameo" term from the celebrity video platform Cameo. Sora has grown rapidly since its launch, already surpassing a million downloads, mostly available in the US and Canada, and continues to refine controls on copyrighted and personal likeness usage to balance creativity with intellectual property protections.​

Google NotebookLM’s Major Upgrade with Custom Personas and 1M Token Memory

Google's NotebookLM has received a significant upgrade, greatly enhancing its AI-powered research and note-taking capabilities. The update introduces a massive expansion in memory with a 1 million-token context window, allowing the AI to maintain coherence and recall over much longer interactions—six times more than before—making it ideal for complex, sustained projects. Additionally, the new custom personas feature enables users to tailor the AI's style, tone, and expertise to better suit specific research needs, transforming it into a personalized assistant for fields like academia, journalism, or data science. Powered by the latest Gemini models, this upgrade also boosts response quality by 50%, supports seamless multiturn conversations, offers saved and secure chat history, and automatically synthesizes insights from multiple angles within users' documents, all while prioritizing privacy with notebook-centric memory isolated per project. These enhancements position NotebookLM as a powerful tool for professionals seeking deep, contextual AI assistance tailored to their unique workflows.​

Kimi Linear 48B: Fast, Efficient Hybrid Attention Model

The Kimi-Linear-48B-A3B-Instruct model by Moonshot AI is an advanced large language model built on a hybrid linear attention architecture called Kimi Delta Attention (KDA), which offers high efficiency and scalability with drastically reduced memory use—up to 75% less—and decoding speeds up to six times faster than traditional full attention models. It has 48 billion total parameters, with around 3 billion activated during each inference step, enabling it to handle extremely long contexts of up to 1 million tokens, suitable for long-form reasoning, document understanding, and reinforcement learning applications. The model combines linear attention with a smaller proportion of global multi-head latent attention, maintaining or even surpassing the quality of full attention across tasks. Open-sourced on Hugging Face, it supports efficient training and inference with PyTorch and Fast Linear Attention libraries, making it a versatile and powerful choice for demanding AI workloads requiring long context processing and high throughput.

Hand Picked Video

In this video, we'll look at BGBlur - an AI-powered video blur tool that automatically blurs faces, backgrounds, and objects in seconds.

Top AI Products from this week

• Pen Island Clean Drawings - Clean drawings, clean mind. A public art site protected by really really advanced computer vision trained to block unwanted content. We are incredibly passionate about sharing art on the internet. But the internet is full of trolls.

• Colab Jetpack - Research is important to product development -- research is also a pain! Conduct AI (or human) interviews with stakeholders/users and automatically generate comprehensive insights, requirements, and prompts at scale.

• Popstarz - Create an AI music artist with Popstarz. Generate your AI artist, AI music, and social media posts easily!

• ChartingLens - Track what Warren Buffett & Michael Burry are buying! Professional platform with real-time charts, 15+ indicators, auto pattern detection, insider trading P&L tracker.

• Baseten Training - We’re excited to announce that Baseten Training is now generally available (GA) to all businesses on Baseten! Baseten Training provides ultra-performant, developer-first infrastructure for training AI models destined for production.

• Inbox Detox - Inbox Detox uses AI to scan your Gmail, intelligently detect promotional emails you're ignoring, and unsubscribe you in one click. ✨ AI-powered inbox waste report.

This week in AI

• Cursor 2.0 Faster, Parallel AI Coding - Cursor 2.0 introduces Composer, a coding model 4x faster with low-latency multi-agent coding and a new interface for efficient parallel agent collaboration, code review, and testing.​

• Altman’s Merge Labs Noninvasive BCI - Altman hires biomolecular engineer Mikhail Shapiro to lead Merge Labs, focusing on noninvasive brain-computer interfaces using ultrasound and gene therapy.​

• Google Mixboard Expands Globally - Google Mixboard Expands to 180+ Countries, Enabling Creators Worldwide to Brainstorm Visually with AI-Powered Moodboards for Ideas and Projects.

• How OpenAI built OWL - OpenAI launched ChatGPT Atlas, a web browser with ChatGPT integrated for faster, smarter browsing. It features instant startup, agent mode for tasks, and maintains performance by separating Chromium from the main app.​

Paper of The Day

This paper presents an agentic AI Home Energy Management System (HEMS) using large language models (LLMs) to autonomously coordinate multi-appliance scheduling from natural language input to device control. It introduces a hierarchical architecture of one orchestrator agent and three specialist agents for washing machine, dishwasher, and EV charger scheduling, leveraging API data like electricity prices and calendar events. The system achieves 100% cost-optimal scheduling with Llama-3.3-70B, matching traditional optimization techniques but with the flexibility of natural language interaction and adaptive workflows. The research highlights model selection, prompt engineering, and deployment considerations for practical use, emphasizing the potential to overcome user complexity barriers and enable widespread HEMS adoption. This approach enables intuitive, conversational control over home energy usage without hardcoded workflows.

To read the whole paper 👉️ here