The Daily AI Briefing - 31/03/2025

Welcome to The Daily AI Briefing, here are today's headlines! Today we're covering massive funding for OpenAI, breakthroughs in understanding Claude's thinking process, integration of deep research capabilities into coding environments, Qwen's new visual reasoning model, SambaNova's research agent, and key OpenAI updates that impact users across their ecosystem. Let's start with what might be the largest private funding round in history: OpenAI is finalizing a staggering $40 billion funding round led by SoftBank. This deal would nearly double ChatGPT's maker valuation to $300 billion. SoftBank will invest an initial $7.5 billion, followed by another $22.5 billion later this year with other investors including Magnetar Capital, Coatue, and Founders Fund. Despite reporting losses of up to $5 billion on $3.7 billion in revenue this year, OpenAI projects its revenue to triple to $12.7 billion in 2025 and become cash-flow positive by 2029 with over $125 billion in projected revenue. Part of this funding will support the ambitious Stargate project, the $300 billion AI infrastructure joint venture announced with SoftBank and Oracle earlier this year. In a fascinating peek behind the AI curtain, Anthropic has released two research papers revealing how its AI assistant Claude processes information. Researchers developed what they call an "AI microscope" that shows internal "circuits" in the model, demonstrating how Claude transforms input to output during key tasks. One remarkable finding is that Claude uses a universal "language of thought" across different languages, with shared conceptual processing for English, French, and Chinese. The research also reveals how Claude plans ahead when writing poetry, identifying rhyming options before constructing lines to reach those planned words. Perhaps most significantly for everyday users, researchers discovered a default mechanism that prevents speculation unless overridden by strong confidence, which helps explain how Claude's hallucination prevention works. For developers looking to enhance their AI-powered coding, there's now a way to add Firecrawl's Deep Research capabilities to coding editors like Cursor and Windsurf. The integration provides real-time web information directly in your coding environment. Setting it up involves creating a Firecrawl account, generating a free API key, and configuring either Windsurf or Cursor with specific commands. Once configured, developers can simply type queries like "Deep Research the latest advancements in React state management" directly in their editor's chat interface. This integration represents another step toward more informed AI coding assistance with up-to-date information. Alibaba's Qwen team has released QVQ-Max, a sophisticated visual reasoning model that goes beyond basic image recognition to analyze and reason about visual information across images and videos. Building on their previous QVQ-72B-Preview, this new model expands capabilities in mathematical problem-solving, code generation, and creative tasks. One of its most interesting features is a "thinking" mechanism that can be adjusted in length to improve accuracy, showing that longer thinking time correlates with better results. QVQ-Max demonstrates complex visual capabilities including analyzing blueprints, solving geometry problems, and providing feedback on user-submitted sketches. Looking ahead, the Qwen team plans to develop a complete visual agent capable of operating devices and playing games. SambaNova has released a new Deep Research AI Agent designed to produce detailed reports and analysis in seconds rather than minutes or hours. This tool allows users to run complex research at a fraction of the traditional time and cost, with research tasks taking just 5-30 seconds to complete. The agent is fully open source and allows connections to user-owned data sources. It integrates with SambaNova Cloud, which delivers fast inference on top open source models, including D