Free AI model better than ChatGPT 4 👀

Gooood afternoon!
This week, we're diving into a world where artificial intelligence isn't just a tool—it's your new classmate, dance partner, and even a life-saving prophet. Meta has just flipped the script on what AI can do, and we're here to break it all down. From AI that can groove to the latest beats, to advanced systems predicting the next big storm, this week’s newsletter is your all-access pass. Plus, we're rolling out the welcome mat in our AI-classroom to explore the vibrant realm of open-source AI. Ready to get smart with us? Let's get this intellectual party started!

Table of contents 📚

• Meta has changed the AI game 🤯

• Template: AI note taker for meetings ✍️

• The weekly Rundown

  • This AI robot can teach you some moves 🕺

  • AI predicting natural disasters 🌋

• AI-classroom: What is open source AI? 🔗

Meta has changed the AI game, here’s how you can use it:

Meta has unveiled two new additions to its Llama 3 series of large language models: the Llama 3 8B and the Llama 3 70B. These models showcase substantial enhancements in performance compared to the earlier Llama 2 series, with the Llama 3 70B model especially rivaling some of the industry-leading generative AI models.

The Llama 3 8B model features 8 billion parameters, whereas the Llama 3 70B boasts a staggering 70 billion parameters. Meta asserts that these models rank among the top generative AI models based on their impressive results across renowned AI benchmarks such as MMLU, ARC, and DROP. In terms of performance, the Llama 3 8B model surpasses other open-source models like Mistral’s Mistral 7B and Google’s Gemma 7B in various benchmarks. Meanwhile, the more robust Llama 3 70B model competes closely with Google’s Gemini 1.5 Pro and holds its ground against Anthropic’s top-tier model, Claude 3 Opus as shown in the official Meta comparison.

1. MMLU (Massive Multitask Language Understanding): This benchmark is designed to test language models' understanding across a wide variety of subjects. It includes tasks related to multiple-choice questions from a range of academic and professional domains, assessing the model's ability to understand and process information from diverse fields of knowledge.

2. GPQA (Generative Physics Question Answering): GPQA is a specialized benchmark that evaluates a model's ability to generate answers to physics-related questions. It tests the model's understanding of physical concepts and its ability to apply this knowledge to solve problems.

3. HumanEval: HumanEval is a benchmark comprising a set of programming problems designed to assess the coding capabilities of AI models. It challenges models to generate correct and efficient code snippets in response to problem statements, measuring the model’s programming proficiency and its ability to understand and implement logical and algorithmic solutions.

4. GSM-8k (Grade School Math - 8k problems): This benchmark consists of approximately 8,000 math problems typically encountered in a grade school curriculum. It's used to evaluate an AI model's ability to understand and solve basic arithmetic, algebra, and geometry problems, reflecting its capacity to perform mathematical reasoning and problem-solving.

Llama 3 70B also beat versions of GPT 4 as shown below in the public rankings.

"Why is this a big deal?" you might wonder. Even though the Llama 3 70B, with its 70 billion parameters, is outperforming giants like GPT4, Claude, and Gemini Pro—which boast far larger parameter counts, with GPT4 rumored to have 1.7 trillion—it's not just about size. The real kicker is how these 'fewer' parameters allow users like us to run the models on our own computers, privately and without sending data to big companies like OpenAI. Yes, you can operate Llama 3 entirely offline, enjoying performance comparable to those hefty, online-only models. This kind of local access was already available with models like Mistral 7B or the older Llama 2, but they didn't quite match up to the likes of GPT 4, Claude, or Gemini Pro—until now.

In combination with Croq servers (accessible via API or their website), is Llama 3 one of, if not the fastest AI model out there. We recommend you testing it out, it is seriously a game changer. You can acces it here.

Meta credits the enhanced performance of the Llama 3 models to a significantly larger training dataset, which is seven times the size of that used for Llama 2, enriched with more code and multilingual content. Additionally, synthetic data was employed to generate longer documents for training purposes. This means that the training of these models are becoming more efficient, which in turn, give companies that make open source

🔗 Template: AI note taker for meetings

Streamline your meetings with our make.com AI note-taker template! Just set it up once, and it automatically transcribes discussions, highlights key points, and sends out summaries to all participants.