DeepSeek Unveils Open-Source Math Powerhouse

First, imagine an AI that doesn’t just solve math problems—it proves theorems with precision rivaling top mathematicians. DeepSeek-Prover-V2-671B is here, and it’s rewriting the rules of formal reasoning. With 671 billion parameters and a recursive thinking pipeline, it's breaking down complex theorems into solvable pieces and achieving near-human results in elite math competitions. Whether you're a researcher or educator, this isn't just a model—it’s a shift in how we teach and explore mathematics.

Meanwhile, Google is quietly reshaping how we interact with the web. The newly expanded AI Mode in Search now puts smarter, more visual results directly in your hands. You can ask layered, complex questions, follow up naturally, and see product ratings, store hours, and local inventory—right in your results. It's not just search anymore—it’s personalized exploration, powered by one of the world’s largest real-time data engines.

And then there’s Claude. With Claude Integrations, Anthropic’s AI is stepping into your workflow—literally. Need updates from Jira, summaries from Confluence, or automations in Zapier? Claude now talks to your apps, conducts deep research across them, and delivers answers that actually move your projects forward. This is what it looks like when AI becomes a real teammate.

The takeaway? AI isn’t just evolving. It’s embedding itself in how we reason, search, and work—turning complexity into clarity, and potential into action.

Advanced AI for Formal Theorem Proving

DeepSeek-Prover-V2-671B is an advanced open-source large language model developed by DeepSeek AI for formal mathematical theorem proving in Lean 4. Leveraging a massive 671-billion-parameter Mixture-of-Experts architecture, it combines informal reasoning and formal proof construction through a recursive pipeline, where complex problems are broken down into subgoals and solved step-by-step. The model achieves state-of-the-art results in automated theorem proving, boasting an 88.9% pass rate on the MiniF2F-test and notable performance on challenging benchmarks like PutnamBench and AIME competition problems. Accompanied by the ProverBench dataset, which includes 325 formalized math problems from competitions and textbooks, DeepSeek-Prover-V2-671B is available on Hugging Face and represents a major advance in AI-driven formal mathematics, supporting research, education, and further innovation in automated reasoning.

Google Expands AI Mode with New Features and Wider Access

Google has expanded AI Mode in Search, making it available to all Labs users in the U.S. without a waitlist, and introducing new features to enhance how people interact with information. Now, users can ask complex, multi-part questions, use follow-ups to refine their queries, and discover new websites and businesses directly within Search. AI Mode now includes visual place and product cards, allowing users to quickly access ratings, reviews, opening hours, real-time prices, images, shipping details, and local inventory for businesses and products, making it easier to make decisions and take action. A new left-side panel on desktop helps users revisit past searches and continue ongoing tasks without starting over. These updates leverage Google’s extensive, real-time data and Shopping Graph, which tracks over 45 billion product listings and updates billions of entries every hour, ensuring users receive fresh, reliable information. Google is also beginning a limited rollout of AI Mode directly in Search for some U.S. users, while continuing to evolve the experience based on user feedback.

Claude Connects Your Apps for Smarter AI Research

Anthropic has launched Integrations for Claude, enabling users to connect the AI assistant directly to popular apps and tools like Jira, Zapier, Asana, Confluence, and more, vastly expanding what Claude can do within business workflows. With Integrations, Claude can access and act on real-time project data, automate tasks, and manage information across connected services, all through conversation. Alongside this, Claude’s new Advanced Research mode lets it search not just the web and Google Workspace, but also any integrated apps, conducting deep, multi-source investigations for up to 45 minutes before delivering a comprehensive, citation-backed report. These features are available in beta for Max, Team, and Enterprise plans (with Pro coming soon), and web search is now globally accessible to all paid users. Developers can build custom integrations using Anthropic’s Model Context Protocol, making it easy to extend Claude’s capabilities and tailor it to organizational needs.

Hand Picked Video

In this video, we'll look at comparing three AI models (Gemini 2.0 Flash, OpenAI o1 & o3 mini, and Deepseek r1 ) for deep research tasks, testing their speed and output quality through a practical demonstration.

Top AI Products from this week

• Raycast for iOS - Daytona Cloud reimagines infrastructure for AI agents with sub-90ms startup times, bare metal performance, and stateful execution—capabilities traditional clouds can't match. Create, control, and deploy AI agents with unprecedented speed and flexibility.

• LLMrefs - Increase your brand’s visibility in AI search. Track keyword rankings and optimize your brand's AI SEO performance in ChatGPT, Gemini, Perplexity & more.

• omiGPT - Find perfect lists of leads, candidates or anything in the web. Websets’ semantic search lets you ask in plain English, it agentically sources results and enriches them with you ask for - like emails, tags, and more - all in the time it takes to brew a coffee.

• FundSpark - Fundraise reimagined with AI. Founders: AI Research & Outreach, Pitch Analysis, Due Diligence Copilot, and Intelligent Networking, Cap Table, Investor Updates. Investors: AI Sourcing, Intelligent CRM, Manage Portfolios, LP reporting, DD Copilot.

• Playgent - Playgent suggests outbound sales plays to generate pipeline more effectively. Feed it your domain and it will serve up AI-picked plays based on your GTM motion, buying signals, target audience, and more.

• Sushidata - We are a Voice-of-the-Customer platform that leverages AI to analyze unstructured conversational data. We uncover valuable insights from your community's public data sources, focusing on identifying business opportunities, improving customer success, and more.

This week in AI

• Amazon Nova Premier - Amazon Nova Premier is a powerful AI model for complex tasks, handling text, images, and videos with 1M token context. It enables multi-agent workflows and model distillation for cost-effective AI.

• AI Image Editing & Animation - Edit images with AI prompts or quick actions in Gamma! Animation is now available for Pro users. All users can access editing-bring your visuals to life today!

• Xiaomi's AI Leap - Xiaomi unveils MiMo, its first 7B parameter AI model, claiming it outperforms OpenAI's o1-mini and Alibaba's QwQ-32B in math and coding. Stock rose 5.3% following announcement.

• Mastercard Agent Pay - Mastercard launches Agent Pay, integrating AI-powered agentic payments with Microsoft and IBM to enable secure, personalized, and seamless commerce in the AI era.

• AI Teaching Revolution - Alpha School uses AI for 2-hour core academics, replacing traditional teachers with "guides." Students spend afternoons learning life skills, showing increased engagement and independence.