Llama 3: The best OpenSource LLM

Today’s top AI Highlights:

1. Meta releases the most anticipated LLM Llama-3 in 8B and 70B variant

2. Meta rolls out its AI chatbot Meta AI to compete with ChatGPT and Claude

3. ChatGPT-on-the-go with the Nothing AI wearable Earbuds

4. Hugging Face releases benchmark to evaluate LLMs in healthcare

5. Hugging Chat comes to iOS, chat with famous opensource models on your phone

& so much more!

Read time: 3 mins

Exciting Opportunity: Share how you use AI and get featured in Unwind AI! Details below.

Latest Developments 🌍

The Most Capable Opensource Models Yet 💪

Meta had successfully sparked anticipation among the AI community for its Llama 3 models. The company has finally unveiled its next-generation of opensource models family Llama 3, starting with two models: 8B and 70B. These models will soon be available on Hugging Face and other platforms but you can chat with Llama 3 for free on Meta AI’s website!

Key Highlights:

1. Architecture: Llama 3 models are decoder-only transformer models. The models use a new tokenizer with a vocabulary of 128K tokens that encodes language much more efficiently and improved model performance significantly.

2. Context Window: Their context window is 8k tokens, double that of Llama 2.

3. Training Data: Models are trained on over 15T tokens, 7x larger than that used for Llama 2, and it includes 4x more code, including a significant component of non-English data covering more than 30 languages. Advanced data-filtering pipelines were employed to ensure that the training data is of high quality.

4. Performance:

   1. The models are the best opensource models so far for their respective sizes. Llama 3 8B outperforms all other similar-sized opensource models like Gemma 7B and Mistral 7B Instruct across all benchmarks with a huge margin.

   2. Llama 3 70B scored 79.5 in MMLU, which is the highest MMLU score for any opensource model yet, also outperforming Gemini 1.5 Pro and Claude-3 Sonnet, along with other benchmarks.

5. Upcoming Releases: Llama 3 family’s largest model is still in training, will have over 400B parameters, larger context window, multimodal, and already is at par with GPT-4 and Claude-3 Opus.

Meta is Giving a Tough Competition to Gemini and ChatGPT

Meta is rolling out its conversational AI “Meta AI” across all Meta apps including Facebook, WhatsApp and Instagram. It is also available here for free. Beyond just text, Meta AI can surf the web and provide updated information, generate images from text prompt, and even lets you edit and animate the generated images.

Key Highlights:

1. You can use Meta AI in search, feed, and messages in Facebook, Instagram, WhatsApp and Messenger app so you can take help instantly without switching apps. Suppose you’re planning a dinner with your friends on an Instgram message group, you can directly ask Meta AI using “@” within the chat which dinner places to go to.

2. You can take Meta AI’s help directly in your Facebook feed when you’re scrolling. If you want to get more information on a post, directly tap Meta AI and put your prompt to know about it.

3. Meta AI can also search for you within Facebook and Instagram. Suppose you’re looking for beautiful reels of sunset on a beach, you can give this prompt to Meta AI within the app and it’ll search and give you the relevant reels.

The Kind of “AI Wearable” We Actually Need Right Now 🎧

A lot of companies are jumping into the AI wearable space to deeply integrate AI into our daily lives, in a way that is not obtrusive. But instead of creating an entirely new device with a steep learning curve, the most straightforward initial approach could have been to embed AI assistants into commonly used non-smartphone devices. Nothing, a London-based consumer electronics company, has successfully mastered this basic concept.

The company is integrating ChatGPT into its wireless earbuds Ear and the Ear (a). You can now talk to the AI chatbot directly through your earbuds, just like you would with Siri or Google Assistant. This works by pinching the earbud stem and asking your questions, provided you have the ChatGPT app on your Nothing phone.

It is like ChatGPT-on-the-go. Whether you’re walking, running, commuting, just ask whatever you want and get the answer!

How Well can AI Models Handle Medical Data

Generative AI is steadily making its mark in the healthcare sector, where it can process and analyze vast amounts of data and support more informed medical decisions. However, the high stakes associated with healthcare—where errors can significantly impact patient outcomes. To address this, Hugging Face has introduced a new benchmark, the Open Medical-LLM Leaderboard, to rigorously assess the performance of LLMs GPT-4, Med-PaLM 2, Mistral 7B, Gemma 7B in medical contexts.

Key Highlights:

1. Multiple Datasets: The Open Medical-LLM Leaderboard evaluates models using a variety of medical datasets, including MedQA (USMLE), PubMedQA, MedMCQA, and subsets of MMLU related to medicine and biology.

2. Focus: Accuracy (percentage of correct answers provided by a model) is the primary metric. This ensures that only the most precise and dependable models are recommended for practical use.

3. Insights: 

   1. Commercial models like GPT-4-base and Med-PaLM-2 consistently demonstrate high accuracy across various datasets.

   2. Open-source models, such as Starling-LM-7B, Gemma-7b, and Mistral-7B show competitive performance on certain datasets and tasks, despite having smaller sizes.

   3. Google’s Gemini Pro excels in procedural knowledge tasks but shows lesser proficiency in critical medical areas such as Anatomy and Cardiology.

   4. This detailed analysis helps identify specific strengths and areas for improvement within each model, guiding further development and specialization in medical applications.

4. Open Participation and Continuous Improvement: The leaderboard also serves as a platform for ongoing improvement. Developers can submit their models for evaluation, engage in discussions, and suggest additional datasets for a more comprehensive testing environment.

Tools of the Trade ⚒️

1. HuggingChat: Chat with multiple opensource LLMs from companies like Mistral AI, Cohere, Meta, and Google. You can generate creative ideas, get expert advice, enhance learning, boost productivity, and receive quick answers to everyday questions. You can even explore the AI community’s best chats and learn from their prompts, and customize the chatbot’s outputs to your preference.

1. Recraft: Create and edit digital illustrations, vector art, icons, and 3D graphics while maintaining a consistent brand style. It offers intuitive tools to customize design elements, manage color schemes, and generate visuals from text or visual inputs quickly.

2. Lingolette: Master a language faster by chatting with AI. It leverages AI to simulate conversations to enhance your language proficiency and pronunciation. It customizes your learning experience based on your proficiency level and interests, making language practice both intensive and interactive.

3. Pentest Copilot: The ultimate ethical hacking copilot that automates processes like payload generation, command execution, and analysis of security vulnerabilities. It assists in every phase of penetration testing, from initial reconnaissance to maintaining access, to efficiently uncover and exploit security weaknesses without needing extensive manual setup.

Text within this block will maintain its original spacing when published🌟 Spotlight on You: Share Your AI Use Case and Get Featured!

Hot Takes 🔥

1. The tech for AGI is already here; it's just a matter of details now. ~Bindu Reddy

2. Based on shows of hands at my recent talks, if you are using a GPT-4 class model, you are still very early. ~Ethan Mollick

Meme of the Day 🤡

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

Real-time AI Updates 🚨

⚡️ Follow me on Twitter @Saboo_Shubham for lightning-fast AI updates and never miss what’s trending!

PS: I curate this AI newsletter every day for FREE, your support is what keeps me going. If you find value in what you read, share it with your friends by clicking the share button below!