Opensource OpenAI DeepResearch

PLUS: Reasoning Augmented Generation, DeepSeek R1-like aha moment for vision models

Today’s top AI Highlights:

  1. Build RAG and AI agents with this comprehensive no-code visual platform

  2. Open source ReAG - No vector DBs, indexing, and chunking

  3. Anthropic challenges you to break their new AI Safety System for $15K

  4. DeepSeek R1-like aha moment of VLM with less than $3

  5. Free and opensource alternatives to OpenAI and Google Deep Research

& so much more!

Read time: 3 mins

AI Tutorials

For businesses looking to stay competitive, understanding the competition is crucial. But manually gathering and analyzing competitor data is time-consuming and often yields incomplete insights. What if we could automate this process using AI agents that work together to deliver comprehensive competitive intelligence?

In this tutorial, we'll build a multi-agent competitor analysis team that automatically discovers competitors, extracts structured data from their websites, and generates actionable insights. You'll create a team of specialized AI agents that work together to deliver detailed competitor analysis reports with market opportunities and strategic recommendations.

This system combines web crawling, data extraction, and AI analysis to transform raw competitor website data into structured insights. Using a team of coordinated AI agents, each specializing in different aspects of competitive analysis

We share hands-on tutorials like this 2-3 times a week, designed to help you stay ahead in the world of AI. If you're serious about leveling up your AI skills and staying ahead of the curve, subscribe now and be the first to access our latest tutorials.

Don’t forget to share this newsletter on your social channels and tag Unwind AI (X, LinkedIn, Threads, Facebook) to support us!

Latest Developments

Stack AI is a platform for building and deploying AI agents through a visual interface - no code required. The platform provides a drag-and-drop canvas where you can create AI workflows by connecting different nodes for LLMs, knowledge bases, and custom tools.

With built-in templates inspired by real business use cases, teams can quickly implement common scenarios or build custom solutions. The platform includes extensive data connectors for AWS S3, SharePoint, OneDrive, and Snowflake, making it easy to integrate with existing data sources.

Key Highlights:

  1. Visual Workflow Builder - Ditch the code and build complex AI workflows using Stack AI’s drag-and-drop interface. Connect pre-built nodes for tasks like data ingestion from various sources (AWS S3, Sharepoint, Snowflake), LLM integration, and output formatting. This enables rapid prototyping and iteration for RAG pipelines and other AI agents.

  2. Modular Nodes Ecosystem - Choose from a wide range of pre-built nodes including LLM connectors (OpenAI, Anthropic, Azure), knowledge bases with vector search, and integrations with common tools like Notion and Gmail. Create custom nodes to integrate your own API endpoints and services.

  3. LLM Memory Management - Build stateful AI agents that remember previous interactions with Stack AI's memory management features. Use the sliding window approach for simple chat applications or leverage VectorDB memory to store and retrieve contextually relevant information from long conversations, enhancing the agent's ability to handle complex tasks.

  4. Deployment Options - Deploy your AI applications as chatbots with customizable UIs, REST APIs for backend integration, or through platforms like Slack and WhatsApp. Built-in analytics and logging help you monitor usage patterns and performance metrics.

ReAG (Reasoning Augmented Generation) takes a new approach to knowledge retrieval by letting language models work with raw documents directly, instead of relying on traditional RAG's two-step process of semantic search and generation.

ReAG feeds complete documents to the model and lets it decide what's relevant through reasoning, eliminating the need for document chunking, vector databases, and embedding pipelines. It integrates with AI SDK by Vercel and LiteLLM, ReAG supports any model, offers reasoning traces, metadata filtering, and works in both TypeScript and Python. It’s completely opensourced under MIT license.

Key Highlights:

  1. No More Vector DBs or Complex Pipelines - ReAG bypasses the need for document chunking, embedding, indexing, and vector databases. This drastically simplifies your application architecture, making it easier to deploy and maintain knowledge-intensive AI solutions. Think less infrastructure, more focus on core functionality.

  2. Contextual Understanding - ReAG lets the LLM analyze full documents and find connections based on reasoning, not just keyword matching. This solves the "lost in the middle" problem and allows you to surface insights from complex data that traditional RAG might miss, leading to more reliable and accurate outputs.

  3. Dynamic Data - Forget the pain of re-embedding and re-indexing. ReAG processes documents on-the-fly, making it ideal for applications dealing with frequently updated information like news, research, or live data feeds. This keeps your AI applications current without constant maintenance.

  4. Multi-Model - ReAG works with any LLM (DeepSeek, Llama, etc.) and offers SDKs in both TypeScript and Python, giving you the flexibility to choose the best tools for your specific project requirements. It is built to directly analyse charts, diagrams, and spreadsheets alongside text.

Quick Bites

AI2 has released Tülu 3 405B, an open-source post-training model that outperforms DeepSeek-V3 and achieves results comparable to GPT-4o across standard benchmarks. The model, built on the Llama-405B base model, implements AI2's novel Reinforcement Learning from Verifiable Rewards (RLVR) framework, resulting in particularly strong improvements in MATH performance at the 405B scale.

The complete training recipe, RLVR code, and model weights are now available on GitHub and Hugging Face, and the model is accessible through AI2's Playground for testing.

Anthropic has launched a public challenge with rewards up to $15,000 for anyone who can successfully jailbreak their new Constitutional Classifiers system by getting it to answer a set of forbidden queries. The challenge follows their newly released research paper showing the system reduced jailbreak success rates on Claude 3.5 Sonnet from 86% to 4.4%, with minimal impact on legitimate queries. During initial testing, 183 participants spent over 3,000 hours attempting to break the system's defenses, but none succeeded in achieving a universal jailbreak across all test queries.

Peking University researchers have adapted DeepSeek R1's "aha moments" approach to vision tasks, creating R1-V - a model that crushes visual counting benchmarks with remarkable efficiency. Their 2B-parameter model outperformed a 72B counterpart using Reinforcement Learning with Verifiable Rewards (RLVR), needing just 100 training steps and $2.62 in compute. The team is releasing everything open-source soon.

Tools of the Trade

4 free and open source alternatives to OpenAI’s $200 and Google’s $20 Deep Research agents:
  1. DeepResearch: By Jina AI, this is a Node.js-based implementation using Gemini, Brave/DuckDuckGo, and Jina Reader, mimicking OpenAI's agentic search, read, and reasoning process. It iteratively searches and extracts information from web pages, returning a comprehensive answer within a defined token budget. Can be accessed via a simple API or run locally.

  2. Open Deep Research: By Firecrawl's team, this opensource clone of Deep Research uses Firecrawl for data extraction and the AI SDK for reasoning, instead of using a fine-tuned version of o3. The demo Next.js application features real-time data extraction, structured data retrieval, and Vercel integration.

  3. Deep-Research: This open-source implementation performs iterative deep research by generating search queries, scraping websites with Firecrawl, and utilizing a mini O3 model (via OpenAI) for reasoning. The tool is built on Node.js, allows for adjustable breadth and depth, and compiles comprehensive markdown reports based on its findings.

  4. OpenDeepResearcher: By Matt Shumer, this is an asynchronous AI agent that performs comprehensive research by iteratively searching with SERPAPI, extracting content with Jina, and using OpenRouter for LLM processing. It produces a final report after refining its search queries until it gathers sufficient information for a comprehensive analysis.

  5. Awesome LLM Apps: Build awesome LLM apps with RAG, AI agents, and more to interact with data sources like GitHub, Gmail, PDFs, and YouTube videos, and automate complex work.

Hot Takes

  1. OpenAI’s deep research is very good. Unlike Google’s version, which is a summarizer of many sources, OpenAI is more like engaging an opinionated (often almost PhD-level!) researcher who follows lead. ~
    Ethan Mollick


  2. Somehow OAI keeps releasing thinking models but they aren’t better than sonnet at code, they think too long and score well on benchmarks but have no taste and can’t produce a good frontend ~
    anton

That’s all for today! See you tomorrow with more such AI-filled content.

Don’t forget to share this newsletter on your social channels and tag Unwind AI to support us!

Unwind AI - X | LinkedIn | Threads | Facebook

PS: We curate this AI newsletter every day for FREE, your support is what keeps us going. If you find value in what you read, share it with at least one, two (or 20) of your friends 😉 

Reply

or to participate.