OpenAI's "Shipmas" continues, Google launches massive updates to most models

OpenAI's "Shipmas" continues, Google launches massive updates to most models
View on Substack
OpenAI's "Shipmas" continues, Google launches massive updates to most models

what to know for now

🎨 Google shows prowess with Veo 2 and Imagen 3. Google DeepMind introduced Veo 2, a state-of-the-art video generation model producing 4K, cinematic-quality videos with improved realism, physics, and human movement. Imagen 3 enhances image generation, offering richer details, accurate prompts, and diverse art styles. Both tools are now integrated into Google Labs’ VideoFX and ImageFX, while Whisk, a new tool, enables image-based remixing with AI. Read more on Veo 2, Imagen 3, and Whisk

🎁 OpenAI's 'Shipmas' continues. Sora AI creates and extends videos from text prompts, featuring an Explore page and style presets. ChatGPT expands canvas to all users, integrates with Apple via iOS 18.2, adds video to Advanced Voice Mode, introduces Projects for organizing conversations, and rolls out SearchGPT globally with enhanced mobile performance. Read more

🤖 Gemini 2.0 enters the agentic era. Google DeepMind launched Gemini 2.0, enabling multimodal inputs and outputs with advanced reasoning and real-time tool use. The model powers new AI agents like Project Astra, Mariner, and Jules, enhancing tasks across browsing, coding, and virtual environments. Read more

🎧 NotebookLM upgrades interface, adds interactivity. Google Labs' NotebookLM introduced a redesigned interface, interactive Audio Overviews, and NotebookLM Plus for premium users. The update features a three-panel layout for managing sources, AI chat, and content creation, alongside voice-enabled interaction for Audio Overviews. NotebookLM Plus offers enhanced limits, team sharing, and enterprise-grade privacy for organizations. Read more

💻 Replit launches AI assistant. Agent enables end-to-end software creation through natural language interactions, while Assistant refines existing projects with direct modifications. Billing utilizes checkpoints, providing unlimited usage with monthly credits. Read more

🧪 AI Research of the Week

Clio: A system for privacy-preserving insights into real-world AI use
From Anthropic

Jake’s Take: The paper introduces Clio, a privacy-preserving analytics platform for analyzing large-scale AI assistant usage data (specifically from Anthropic’s Claude.ai chatbot). It aggregates millions of user interactions into high-level patterns and clusters (while claiming to maintain privacy through multiple safeguards).

Clio identifies popular real-world use cases like coding, research, and writing tasks, revealing language-specific trends and novel abuse attempts. The system was also employed to monitor safety concerns, detect scaled misuse, and evaluate safety classifier effectiveness. Despite its capabilities, Clio acknowledges limitations in detecting rare behaviors, intent ambiguity, and the inherent trade-offs between granularity and privacy.

Clio is the first time a foundation company has discussed its post-deployment monitoring in length. Its effectiveness points toward the need for broader, industry-wide adoption of similar tools to ensure AI systems align with societal safety standards.

what to know for later

⚛️ Willow achieves quantum breakthrough. Google’s new chip claims to reduce errors exponentially with increased qubits, overcoming a major quantum error correction challenge. It performed a computation in under five minutes, a task that classical supercomputers would require 10 septillion years. Read more

🧮 Microsoft talks Phi-4 SLM. Microsoft introduced Phi-4, a 14B-parameter small language model excelling in complex reasoning, particularly in math tasks. Leveraging synthetic datasets, curated organic data, and post-training innovations, Phi-4 supposedly surpasses larger models like Gemini Pro 1.5 in math competition benchmarks. Read more

🎬 Pika 2.0 expands AI video control. Pika’s updated AI video tool introduced “Scene Ingredients” for customizable characters, objects, and settings, improving control over generated clips. Enhanced motion rendering and text alignment further refine naturalistic movement and prompt accuracy. Read more

🔧 Devin now available for teams. Offered at $500 per month, Devin includes unlimited seats, Slack integration, IDE extensions, and API access. It supports tasks like frontend bug fixes, PR drafting, and code refactoring, with onboarding and support from Cognition's engineering team via app.devin.ai. Read more