Meta dropped Llama-3 just a few weeks ago and it has taken everyone by surprise. People are coming up with wild use cases every day, pushing the model to its limits in incredible ways. 

Here are 10 impressive examples of what it can do.

Llama-3 8B with a context length of over 1M

Developed by Gradient and sponsored by compute from Crusoe Energy, this model, called Llama-3 8B Gradient Instruct 1048k, extends LLama-3 8B’s context length from 8k to over 1048K. This model shows that SOTA LLMs can efficiently manage long contexts with minimal training by appropriately adjusting RoPE theta.

The model was trained progressively on increasing context lengths, drawing on techniques like NTK-aware interpolation and Ring Attention for efficient scaling. This approach allowed for a massive increase in training speed, making the model both powerful and efficient in handling extensive data.

RAG App with Llama-3 running locally 

You can build a RAG app with Llama-3 running locally on your computer (it’s 100% free and doesn’t require an internet connection). 

The instructions include simple steps like installing the necessary Python Libraries, setting up the Streamlit App, creating Ollama embeddings and a vector store using Chroma, and setting up the RAG chain among other things. 

Agri Vertical Dhenu 1.0 model fine-tuned on Llama3-8B

KissanAI’s Agri Vertical Dhenu1.0 model has been fine tuned on Llama3 8B for 150K instructions. It is India-focused and available for anyone to download, tinker and provide feedback.

Tool Calling Champion

Llama-3 70b on GroqInc is a tool-calling champion. The 70b model passed the task when given a query, was very fast, and had the best pricing. It’s also performing great at benchmarks and tests. 

Lightning-fast Copilot in VSCode

You can connect @GroqInc with VSCode, unlocking the full potential of Llama-3 as your Copilot. 

Just create your account on the Groq console, head to the ‘API Key’ menu and generate yours, download the CodeGPT extension from the VSCode marketplace. After this, open CodeGPT and select Groq as the provider, click on ‘Edit Connection’, paste your Groq API Key, and then click ‘Connect’. 

That’s how you can connect Groq to VSCode and access all the models offered by this service.

Llama-3 Function Calling

Llama-3 function calling works pretty well. Nous Research announced Hermes 2 Pro, which comes with Function Calling and Structured Output capabilities. The Llama-3 version now uses dedicated tokens for tool call parsing tags to make streaming function calls easier.

The model surpasses Llama-3 8B Instruct on AGIEval, GPT4All Suite, TruthfulQA, and BigBench.

TherapistAI, powered by Llama3-70B

TherapistAI.com now runs on Llama3-70B, which, according to the benchmarks, is almost as good as GPT-4. The Llama3-70B model significantly enhanced the app’s conversational capabilities, enabling a back-and-forth, ping-pong style interaction. The responses have become concise, direct, and highly focused on problem-solving. 

With Llama-3, Therapist AI now actively engages by asking questions, which helps it understand better and address specific user needs. It also exhibits an impressive memory, allowing it to maintain context over longer conversations, thereby enhancing its ability to deliver relevant and actionable answers. 

You can also use Llama-3 to build such applications. It ensures great performance and is less expensive than using ChatGPT 4, which is around $20 per month. 

AI Coding assistant with Llama 3

It’s time to give your productivity a boost by building an AI Coding assistant with Llama3. 

To develop an AI coding assistant using Llama3, start by downloading Llama3 via Ollama, and then integrate a system message to enable it as a Python coding assistant. Next, install the Continue VSCode extension, connect it with my-python-assistant, and activate the tab-autocomplete feature to enhance coding efficiency.

Superfast Research Assistant using Llama 3 

You can build a research assistant powered by Llama-3 models running on Groq. You can then take any complex topic, search the web for information about it, package it up, and send it to Llama-3 running on Groq. It will send back a proper research report.  

Building RAG Capabilities for Accessing Private Data

Subtl.ai is building in-house RAG capabilities for accessing private data. Founded with the goal of democratizing access to private data for specific professional needs, the platform significantly improves efficiency by offering a 5x faster access to information. It does all this while maintaining data security through an AI that securely processes and recalls your data, allowing AI-enhanced access as well as data protection.

The company will be releasing their AI bot built on Llama-3 soon.