Chinese Startup launches DeepSeek R1 a rival to OpenAI o1 and it is free

Chinese AI lab DeepSeek released its R1 model family under an open MIT license, with its largest version containing 671 billion parameters, claiming performance on par with OpenAI's o1 simulated reasoning (SR) model. The release includes the main DeepSeek-R1-Zero and DeepSeek-R1 models, as well as six smaller "DeepSeek-R1-Distill" versions ranging from 1.5 billion to 70 billion parameters, capable of running on devices from laptops to high-performance servers.

The models gained attention for their advanced reasoning abilities, comparable to proprietary models, and their open-source availability, which allows for study, modification, and commercial use. DeepSeek claims R1 outperforms OpenAI's o1 on several benchmarks, though these results await independent verification. Unlike conventional large language models (LLMs), R1 employs an inference-time reasoning approach, simulating a human-like chain of thought to improve performance on complex tasks like math and coding.

Despite its open licensing, the cloud-hosted version of R1 adheres to Chinese Internet regulations, filtering responses on sensitive topics. However, local versions bypass these restrictions. The release highlights the proliferation of capable, locally runnable AI models, potentially reducing reliance on proprietary systems and centralized controls.

Link to article

OpenAI launches AI Agent called “Operator”

OpenAI has launched Operator, an AI agent capable of performing web-based tasks autonomously, like filling forms, ordering groceries, and creating memes. Powered by the Computer-Using Agent (CUA) model, Operator combines GPT-4o's vision capabilities with reinforcement learning to interact with graphical user interfaces (GUIs). It can "see" through screenshots, "interact" via mouse and keyboard actions, and self-correct errors.

Currently available as a research preview for Pro users in the U.S., Operator allows users to complete repetitive tasks efficiently, personalize workflows, and run multiple tasks simultaneously. It ensures user control by requesting assistance for sensitive tasks like logins and payments. Users can also add custom instructions or save prompts for frequent tasks.

Operator collaborates with companies like DoorDash, Instacart, Uber, and public organizations like the City of Stockton to improve accessibility, streamline workflows, and enhance digital engagement. Future plans include expansion to other user tiers and integration into ChatGPT.

Link to article

Perplexity Launches AI Assistant

Perplexity, an AI-powered search engine, has launched Perplexity Assistant, a tool that combines reasoning, search, and apps to help with daily tasks. Available for Android, it can perform "multi-app actions," such as booking rides, finding event dates, and adding calendar entries. It also supports multimodal interactions, using a phone's camera to answer questions and maintaining context between tasks, like researching and reserving restaurants. Initially free in 15 languages, the assistant has some limitations, as admitted by CEO Aravind Srinivas, with planned improvements underway.

The launch follows Perplexity's rollout of Sonar, an API for embedding its AI tools, and its acquisition of Read.cv, a professional networking platform. Founded in 2022 and valued at $9 billion, Perplexity handles over 100 million weekly queries and has raised $500 million in funding. However, it faces legal challenges from publishers like News Corp and The New York Times, accusing it of misusing their content. Despite offering a revenue-sharing program, Perplexity’s legal disputes highlight ongoing concerns about its practices.
Link to article

Anthropic Launches Citations for referencing sources

Anthropic has introduced Citations, a new API feature for Claude that enhances response reliability by grounding answers in source documents. This feature allows Claude to provide detailed references to the specific sentences and passages it uses, making outputs more trustworthy and verifiable.

Key Highlights:

• Trust through Verification: Citations addresses the need to verify AI-generated responses, eliminating the complexity of prompt engineering for sourcing information.

• Use Cases:

  • Document summarization: Summaries with traceable sources.

  • Complex Q&A: Answers tied to specific document sections.

  • Customer support: Cited responses based on multiple reference materials.

• How it Works: Source documents (PDFs or text) are chunked into sentences and provided to Claude. The model analyzes these and generates responses with precise citations, reducing hallucinations and improving accuracy.

• Pricing: Based on a token model; quoted text output is not charged.

• Customer Success:

  • Thomson Reuters: Leveraged Citations for their AI platform, CoCounsel, enhancing trust and accuracy for legal professionals.

  • Endex: Reduced hallucinations and formatting errors to 0%, improved references by 20%, and eliminated the need for complex prompt engineering.

Citations is available for Claude 3.5 Sonnet and Haiku on the Anthropic API and Google Cloud’s Vertex AI.

Link to full article

Trump announces $500 billion in AI Infrastructure managed by OpenAI, Oracle and SoftBank

OpenAI, SoftBank, and Oracle announced a joint venture called the Stargate Project to build large-scale AI data centers in the U.S., starting with a site in Texas. The initial investment is $100 billion, with plans to reach $500 billion over four years. The project aims to create hundreds of thousands of jobs and solidify U.S. leadership in AI.

Key details include:

• Partners and Contributions: OpenAI will manage operations, while SoftBank provides financial backing. Oracle, Nvidia, Microsoft, Arm, and others are involved as technology partners and investors. SoftBank also holds prior investments in OpenAI, totaling $2 billion.

• Infrastructure Plans: Stargate's first site in Abilene, Texas, will include 10 buildings, each 500,000 square feet, with a 1-gigawatt capacity by 2026. The project may expand to 20 data centers by 2029.

• AI Chips: OpenAI is working on custom AI chips with Broadcom and TSMC, expected by 2026.

• Challenges: Data centers face criticism for environmental impact, high power and water consumption, and limited job creation. However, demand for AI infrastructure is driving significant investments.

The announcement highlights the urgent need to scale AI infrastructure as global AI demand grows, with competitors like Microsoft and BlackRock also investing heavily in data centers.

Link to article

Andrew Ng new course on Anthropic Computer Use AI Agent

The course "Building Towards Computer Use with Anthropic" focuses on leveraging Anthropic’s AI models to build applications capable of interacting with computer interfaces. Taught by Colt Steele, the course covers Anthropic’s approach to AI research, safety principles, and the features of its models.

Key Learnings:

• Anthropic Models and Research: Understand their family of models, AI safety principles, and unique features.

• Prompting Techniques: Create effective prompts using templates, XML structures, and examples. Learn multi-modal prompting (combining text and images) and implement prompt caching to reduce costs and latency.

• Tool Use and AI Applications: Build workflows for chatbots that call tools and integrate streaming responses.

• Computer Use Capability: Learn how Anthropic’s models use images of computer screens to analyze, navigate, and execute tasks with clicks and keystrokes.

• Final Demo: Combine these concepts to build an AI assistant capable of interacting with a computer.

This course enables participants to effectively use Anthropic's models to create cutting-edge AI applications for multimodal and tool-based use cas

Link to course