Claude Now Controls Your Apps: Slack, GitHub & More

Today’s top AI Highlights:

1. Cursor's new AI agent now writes code and controls your Terminal

2. Anthropic’s Model Context Protocol lets you connect Claude to GitHub, Slack, Google Drive and more

3. Run inference on vision language models with HF’s transformers library

4. DeepSeek’s new model brings o1-level performance and transparent though process

5. Run Ollama models locally on your mobile

& so much more!

Read time: 3 mins

AI Tutorials

A ChatGPT-like assistant that runs entirely offline and recalls past conversations—an AI that learns from each chat and personalizes its responses, all without any internet dependency. Giving this kind of control to users is a powerful way to make AI both secure and adaptable for private use cases.

In this tutorial, we’ll build a local ChatGPT clone using Llama 3.1 8B with a memory feature, making it capable of recalling past conversations. All components, from the language model to memory and vector storage, will run on your local machine

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 levelling 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

Cursor Responds to ChatGPT AI Agent 🤖

The Cursor code editor just dropped a major update (v0.43) featuring an AI Agent that's like having a coding buddy in your IDE. This Agent can independently tackle tasks, navigate your terminal, and even pick its own context for operations. Beyond the Agent, the update brings a slicker Composer UI in the sidebar with inline diffs, smarter file recommendations, and improved semantic search within your projects.

Key Highlights:

1. Agent-Driven - The new Agent feature in the Composer can execute entire coding tasks based on your instructions, using the terminal and intelligently selecting its own context, automating significant portions of your workflow.

2. Code Understanding - The AI now provides file suggestions directly in the chat and Composer, plus it uses semantic search to better understand your codebase and find relevant information. This should translate to quicker navigation.

3. Composer Workflow - The redesigned Composer UI lives in the sidebar and incorporates inline diffs, simplifying code review and making changes more transparent. Improved image drag-and-drop functionality also enhances workflows involving UI design and code generation.

4. Bug Detection (Beta) - Version 0.43 offers a sneak peek at an upcoming bug finder. While still in beta, this feature hints at future automated bug detection capabilities directly within Cursor.

Instant Setup, Instant Results: Hire a Synthflow AI Agent Today

Your Next Best Hire: A Synthflow AI Voice Agent. With human-like interaction, it manages calls, qualifies leads, and more, 24/7. Cost-effective plans starting at $29/month, and integrates with top CRMs. Start your free trial and welcome your new team member!

Hire AI Voice Agent

Anthropic’s Protocol to Connect AI with Apps, Data & Tools ⛓️

Anthropic has released Model Context Protocol (MCP) that lets you connect your AI applications directly to your data—no more custom integrations for every data source. Think of it as a universal adapter for your AI. With MCP, you define servers that give your AI access to specific data and tools, enabling secure and controlled interactions.

Here's the core concept: you create servers, using provided SDKs and examples, that can access specific resources (databases, files, APIs, etc.). These servers communicate with AI applications using MCP, a standardized protocol for AI-to-data interaction. This allows AI applications to securely query and interact with resources through the intermediary server.

Key Highlights:

1. One Protocol, Many Sources - MCP provides a single, standardized way to connect your AI to diverse resources – local files, databases, APIs, cloud services – without writing custom code for each integration.

2. Security First - MCP servers control what the AI can access. They act as gatekeepers, ensuring your AI only interacts with authorized data and tools within your environment. No more sharing sensitive API keys directly with your LLM provider.

3. Streamlined Workflow - Setting up an MCP server is straightforward. Anthropic provides SDKs and examples to help you quickly build servers for common resources. Imagine connecting Claude to your internal systems in minutes.

4. Agentic Potential - MCP doesn't just allow data access; it lets your AI act through tools exposed by servers. This opens up powerful automation possibilities – think of AI running scripts, updating databases, or interacting with web services based on your data.

5. Quick Start - Install Claude Desktop and run
   npx @modelcontextprotocol/create-server 
   to bootstrap your first MCP server. Pre-built servers for GitHub, Slack, and PostgreSQL can be installed directly through the Claude Desktop app.

Quick Bites

DeepSeek has released DeepSeek-R1-lite-preview, a new language model with OpenAI o1-like reasoning architecture and transparent thought-process in real-time. It strongly competes against o1 model on benchmarks like AIME and MATH, sometimes even surpassing it. It will soon be available open-source with an API. You can try it out here (50 messages cap daily).

Hugging Face has expanded its Transformers library's pipeline abstraction to support vision language models. The pipeline abstraction handles all the messy details of tokenization and model loading for you. This means you can describe images with just a few lines of code, or even use the Inference API if you don't want to run models locally.

Rabbit just rolled out "teach mode" for its R1 device, letting you create custom AI agents and teach them tasks like saving songs or drafting social posts. You can access this experimental feature through the Rabbithole web interface, record yourself performing a task, and then have the AI repeat it across various websites.

Cerebras achieved a record-breaking 969 tokens/s with Llama 3.1 405B, making it 12x faster than GPT-4o and and 18x faster than Claude 3.5 Sonnet. This also includes the fastest time-to-first-token at 240ms and best-in-class performance at 128K context length. Expect general availability in Q1 2025 at $6/M input tokens and $12/M output tokens.

PyTorch's LLM fine-tuning library, Torchtune, has released v0.4.0 with several updates. The new version adds support for activation offloading to reduce memory usage, enables fine-tuning of the Llama 3.2 Vision 90B model with QLoRA, and integrates the Qwen 2.5 model family. Updated documentation is also available to guide users on customizing recipes, configurations, and components.

Tools of the Trade

1. MyOllama: Opensource mobile client that enables interaction with Ollama-based LLMs from your iOS/Android device. It lets you connects to a computer running Ollama software, allowing interacting with various LLMs directly on your iOS or Android device.

2. Prompt Fuzzer: Opensource tool that tests and improves the security of your AI application's system prompt by simulating various attacks. It helps you identify and fix vulnerabilities to make your AI more robust and secure.

3. LiveAPI: Automatically generates API documentation directly from code repositories like GitHub and GitLab. It has a user-friendly interface, API testing capabilities, code generation, and more.

4. SimpleMind: A Python library that provides an easy-to-use interface for interacting with various AI APIs like OpenAI, Anthropic, and xAI to simplify AI integration. It supports text generation, structured data responses, conversational AI, and allows for custom plugins to extend functionality.

5. Awesome LLM Apps: Build awesome LLM apps using RAG to interact with data sources like GitHub, Gmail, PDFs, and YouTube videos with simple text prompts. These apps will let you retrieve information, engage in chat, and extract insights directly from content on these platforms.

Hot Takes

1. Every serious AI researcher I've met believes we need something beyond LLMs to reach AGI. The field splits on whether missing ideas are small (LLM++) or fundamental (Deep Learning + Program Synthesis). In public debates, LLM++ gets collapsed into just LLM. ~
   Mike Knoop
   
   

2. A data scientist isn't a scientist, a prompt engineer isn't an engineer, machine learning doesn't learn, artificial intelligence isn't intelligence, and agents don't have agency. This industry is so full of shit. ~
   Andriy Burkov

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

Awesome LLM Apps | Sponsor Us

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 😉