Grok 4 Just Beat Every AI Model!

Grok 4: Overview, API Usage, and Agent Creation

Key Concepts:

• Grok 4: XAI's latest large language model (LLM) with improved intelligence, coding capabilities, and context window.
• Context Window: The amount of text or data the model can consider at once (256,000 tokens for Grok 4).
• Function Calling: The ability of the model to use external tools or APIs.
• Structured Outputs: The ability of the model to generate data in a specific format (e.g., JSON).
• Reasoning: The model's ability to think through problems and provide logical solutions.
• XAI SDK: Software Development Kit provided by XAI for interacting with their models.
• Praise AI Agents: A library for creating AI agents that can use LLMs.
• Agentic Behavior: The ability of the model to act autonomously and use tools to achieve a goal.
• Benchmarks: Standardized tests used to evaluate the performance of LLMs (e.g., GPQA, VendingBench).
• Token: A unit of text used for pricing and model input/output.

Grok 4 Overview

• Capabilities: Grok 4 is presented as a highly intelligent model, surpassing even Claude 3 Pro in some areas. It excels in coding tasks and demonstrates strong reasoning abilities.
• Performance:
  • Top performer in GPQA benchmark AIME25 LCB HMMT and USA.
  • Number one for coding.
  • Number two for output speed.
  • Tops the math index.
  • Grok 4 heavy is the top performing humanity lost exam.
  • Grok 4 is the top beating Claude Opus 4 in vending bench.
• Variations: Three versions exist: Grok 4 (standard), Grok 4 Heavy (top-performing), and Grok 4 without tools.
• Pricing: $3 for input and $15 for output per million tokens. Cached input is $0.75 per million tokens.
• Access: Available through a paid subscription on grok.com ($300/year or $30/month for Grok 4, $300/month for Grok 4 Heavy).

Using the Grok 4 API with XAI SDK

• Installation:

  1. Open terminal.
  2. pip install xai-sdk
  3. export XAI_API_KEY=<your_api_key> (API key obtained from XAI website).

• Code Example (app.py):

  from xai import client
  from xai.client import Chat
  
  client = Client(api_key="YOUR_API_KEY")
  
  chat = client.chat.create(
      model="grok",
      messages=[
          {"role": "system", "content": "You are a PhD level mathematician."},
          {"role": "user", "content": "Solve this equation: 2x + 3 = 7"}
      ]
  )
  
  print(chat.response)
  

• Explanation: The code initializes the XAI client, creates a chat session with the "grok" model, provides a system instruction (role-playing as a mathematician), asks a question, and prints the response.

• Running the Code: python app.py in the terminal.

Creating AI Agents with Praise AI Agents

• Installation: pip install praise-ai-agents llm

• Code Example:

  from praise_ai_agents import Agent
  from llm import Instruction
  
  agent = Agent(
      instruction=Instruction("You are a helpful assistant."),
      llm="grok"
  )
  
  response = agent.ask("What is the capital of France?")
  print(response)
  

• Explanation: This code imports the necessary libraries, creates an agent with a specific instruction and LLM ("grok"), and then asks a question.

• Enhancements:

  • Multiple agents can be added for collaboration.
  • MCP tools can be integrated with a single line of code to enhance agent capabilities.

• OpenAI SDK Integration: Possible by providing the Grok API base URL and API key.

Grok 4 Testing and Examples

• Misguided Attention Test: Grok 4 correctly identified the ethical dilemma in a modified trolley problem.
• Python Expert Level Challenge (Vel Q): The model provided a response, but the testing system indicated a failure.
• Simplify Josephus Task: The model successfully solved the task.
• Far Sequence Task: The model generated code, but it initially produced a syntax error due to the testing system's Python version. Grok 4 was able to identify the Python version issue and correct the code.
• Safety Test (Breaking into a Car): The model refused to provide instructions for illegal activities and suggested legitimate alternatives (e.g., calling a locksmith).

Observations and Insights

• Agentic Behavior: The Grok interface exhibits agentic behavior, utilizing tools and internet search to answer questions.
• Python Version Awareness: Grok 4 can identify and adapt to different Python environments.
• Safety Measures: The model incorporates safety measures to prevent misuse.
• Limitations: The testing system used in the video may have limitations (e.g., outdated Python version) that can affect the results.

Conclusion

Grok 4 is a powerful LLM with strong capabilities in coding, reasoning, and general knowledge. The XAI SDK and Praise AI Agents library provide convenient ways to access and utilize Grok 4 in various applications. While some limitations were observed during testing, the model demonstrated impressive performance and adaptability. The presenter encourages viewers to try Grok 4 and share their experiences.