🎬 OpenAI Shutters Sora, Pivots to 'Spud'

OpenAI has discontinued its Sora video generation app and terminated its partnership with Disney, which included a $1B investment and a 200-character licensing agreement. The company is now developing a replacement model codenamed 'Spud' while simultaneously raising another $10B, bringing its latest funding round to approximately $120B total.

Key Insights:

• Sora launched six months ago and has now been shut down entirely, with OpenAI citing strategic realignment toward a new video generation approach

• The Disney partnership, which involved a $1B investment and character licensing deal, was terminated after approximately three months

• OpenAI is raising an additional $10B in funding, pushing its latest round to around $120B total valuation

• The OpenAI Foundation, now holding approximately $130B in equity, has committed over $1B this year to AI-driven scientific discovery and societal resilience initiatives

• Sam Altman is shifting focus from safety oversight to data center development and fundraising, while a new model called 'Spud' is in development

The Bigger Picture: The rapid shutdown of Sora and the Disney deal signals that OpenAI is making hard strategic pivots rather than iterating on existing products. For AI practitioners, this suggests the video generation landscape remains unsettled - what works in demos may not translate to sustainable products at scale. The $120B funding round and Altman's shift toward infrastructure also indicate OpenAI is betting its future on compute capacity and capital deployment, not just model improvements.

Read More →

🎙️ Gemini 3.1 Flash Live Drops Latency

Google has released Gemini 3.1 Flash Live via the Gemini Live API in preview, targeting developers building real-time voice and vision agents. The model delivers improvements in latency, reliability, and natural-sounding dialogue specifically designed for conversational AI applications.

Key Insights:

• The model achieves higher task completion rates in noisy, real-world environments by filtering background sounds like traffic or television while maintaining responsiveness to instructions

• Instruction-following capabilities have been significantly boosted, allowing agents to stay within operational guardrails even during unexpected conversational turns

• Latency improvements and better recognition of acoustic nuances like pitch and pace make real-time conversations feel more fluid and natural compared to the previous 2.5 Flash Native Audio model

• The model supports more than 90 languages for real-time multi-modal conversations

• Developers including Stitch are already using the API to build voice agents that enable users to design with their voice while the agent 'sees' the canvas and selected screens

The Bigger Picture: Real-time voice AI has been bottlenecked by latency and reliability issues that break conversational flow. Gemini 3.1 Flash Live's focus on acoustic nuance recognition and noise filtering addresses the gap between demo-quality voice agents and production-ready systems that work in messy, real-world conditions. For developers, this preview signals that multi-modal conversational AI is shifting from research novelty to deployable infrastructure - if the latency gains hold up at scale.

Read More →

🖥️ Claude Can Now Control Your Mac

Anthropic has updated Claude to autonomously control macOS computers through its Code and Cowork AI tools, allowing the chatbot to open files, use browsers and apps, and run developer tools even when users are away. The feature is available as a research preview for Claude Pro and Max subscribers.

Key Insights:

• Claude can now execute tasks autonomously by directly controlling your browser, mouse, keyboard, and display ‘with no setup required' on macOS devices

• The feature builds on autonomous capabilities introduced with Claude 3.5 Sonnet in 2024, now extended to the Code and Cowork AI agents for programmers

• Claude prioritizes connectors to supported services like Slack and Google Workspace first, but will control your computer directly if no connector is available

• The system requires the Claude desktop app running on macOS paired with the mobile app, and 'always asks for your explicit permission' before taking control actions

• Anthropic acknowledges limitations: 'Complex tasks sometimes need a second try, and working through your screen is slower than using a direct integration.'

The Bigger Picture: Anthropic is pushing beyond chatbot interfaces into autonomous computer control - a capability that blurs the line between AI assistant and AI operator. The fact that Claude will 'always ask for permission' suggests guardrails are still being worked out, and the admission that screen control is 'slower than direct integration' reveals this is an intermediate solution, not the final architecture. For developers, this signals a shift from building AI features into apps to building apps that AI can operate directly - a fundamentally different design paradigm.

Read More →

📰 Arm Ships Its First In-House CPU, Meta Signs On

After decades of licensing chip designs, Arm launched the AGI CPU for AI inference with Meta as lead partner and co-developer. The chip delivers twice the performance per watt of traditional x86 CPUs with up to 136 cores per CPU and 64 CPUs per air-cooled server rack.

Read More →

📰 Gimlet Labs Raises $80M to Fix AI Hardware Waste

The Stanford spinout built software that splits AI workloads across diverse hardware types simultaneously, claiming 3x to 10x inference speedups. The company estimates that AI apps use deployed hardware only 15-30% of the time, and launched publicly in October with at least $10 million in revenues.

Read More →

📰 Apple to Open Siri to Third-Party AI Chatbots

iOS 27 will let users choose which AI chatbot powers Siri responses through a new Extensions feature, according to Bloomberg. Google Gemini and Anthropic Claude will join the existing ChatGPT integration, working across iPhone, iPad, and Mac.

Read More →

📰 Google Simplifies Switching to Gemini from Other AIs

Gemini now offers Import Memory and Import Chat History tools that let users copy preferences and conversation history from competing chatbots. Users paste prompted outputs or upload exported chat files to continue conversations in Gemini without retraining the AI on their preferences.

Read More →

🔍 Kensho Grounding Framework

Kensho built a multi-agent framework using LangGraph that turns S&P Global's fragmented financial data into a single natural language interface. A central router directs queries to specialized Data Retrieval Agents - each owned by different data teams handling equity research, fixed income, or macroeconomics - so analysts can skip the database schema gymnastics and get citation-backed answers from verified sources. It's the pattern you need when your data estate is sprawling, and your users demand both speed and audit trails.

Try it →

🧪 Deep Agents Eval Philosophy

LangChain's Deep Agents team breaks down how they build evals that actually improve agent performance: catalog the production behaviors you care about (like multi-file retrieval or composing multiple tool calls in sequence), write self-documenting tests tagged by category, and trace every run to LangSmith so the whole team can analyze failures together. The core insight - more evals don't make better agents, targeted evals that measure real-world capabilities do. It's a practical blueprint for teams drowning in test suites that don't move the needle.

Try it →