THE Best Week for Open Source AI!

2 Major open source models were announced this week, Llama-3 and Mistral's 8X22B MOE Model. A lot of folks built out RAG Chatbots, Co-pilot alternatives and so much more!

Stuff you should know

LLama 3 breaks the Industry!

Meta Llama 3 is a large language model (LLM) developed by Meta AI for commercial and research use in English. It is designed to provide context for controversial queries and offer nuanced responses to potentially contentious topics. The model is available in 8B and 70B parameter sizes and is trained on 15 trillion tokens of data. Meta Llama 3 is released under a permissive commercial and private use license, but not as permissive as traditional open-source options. It aims to answer questions more accurately and handle a wider range of questions, including controversial topics, and is part of Meta's goal to make Llama-powered Meta AI the most useful assistant in the world.

In fact, this is so good, we replaced Github co-pilot with it:

Microsoft’s Talking Avatars are so good!

Microsoft's VASA-1 is an AI model developed by Microsoft Research Asia that has the capability to create deepfake content using just one photo and one audio track. This technology allows for the creation of realistic deepfake videos where a person's face can be manipulated to match the audio track, enabling the generation of convincing fake content. Microsoft's VASA-1 is a significant advancement in AI technology, showcasing the potential for creating sophisticated and deceptive media content with minimal input data.

Google Announces CodeGemma!

CodeGemma is a collection of lightweight models developed by Google for various coding tasks, such as code completion, generation, natural language understanding, mathematical reasoning, and instruction following. These models aim to enhance the accuracy of code generation by making it both syntactically correct and semantically meaningful, reducing errors and debugging time. CodeGemma supports multiple programming languages, including Python, JavaScript, Java, Kotlin, C++, C#, Rust, and Go.

The models are available through the Hugging Face model hub, with pre-trained variants like the 2B and 7B variants optimized for code infilling tasks. The 2B variant offers faster code completion, while the 7B variant is specially tuned for code infilling. Developers can integrate CodeGemma models into their development environment to write less boilerplate code and focus on more interesting and differentiated code. The models support fill-in-the-middle (FIM) tasks, which can be used in autocomplete or coding assistant tooling. Official documentation and quickstart guides are available for more information on using CodeGemma models.

Mistral announces 8x22B Open Source Model

Mixtral 8x22B is an open-source sparse Mixture-of-Experts (SMoE) model developed by Mistral that offers unparalleled cost efficiency and performance. With only 39B active parameters out of 141B, it is fluent in English, French, Italian, German, and Spanish, and has strong mathematics and coding capabilities. The model is natively capable of function calling and has a 64K tokens context window for precise information recall from large documents.

Mixtral 8x22B is released under the permissive Apache 2.0 license, allowing anyone to use the model without restrictions. It is part of a family of highly efficient models that deliver the best performance-to-cost ratio compared to other open models.

In terms of performance, Mixtral 8x22B excels in reasoning and knowledge benchmarks, outperforming other leading open models on tests such as MMLU, HellaSwag, Wino Grande, Arc Challenge, TriviaQA, and NaturalQS. It also demonstrates strong multilingual capabilities, surpassing LLaMA 2 70B on HellaSwag, Arc Challenge, and MMLU benchmarks in French, German, Spanish, and Italian.

Furthermore, Mixtral 8x22B performs best in coding and maths tasks compared to other open models, as evidenced by its scores on HumanEval, MBPP, GSM8K, and Math benchmarks. The instructed version of the model shows even better math performance, with a score of 90.8% on GSM8K maj@8 and 44.6% on Math maj@4.

Handpicked video

In case you want a video coverage, here you go:

Top Products launched this week

• Cascadeur 2.0 -  Animation Unbaking, it is a tool that simplifies the process of turning any baked animation into an editable one.

• Tabula - The only AI-powered data tool your startup needs. Centralize data and automate report generation with all your custom metrics.

• Brave Search - With Brave Search, you get real answers, faster. AI-powered answers to almost any question, in addition to a list of links. Brave delivers results from its own, built-from-scratch index. No secret methods or algorithms.

• Open Agent Studio - Open Agent Studio invents new RPA concepts, enabling building previously impossible business automations in simple english rather than brittle code selectors or computer vision. APA solves a fundamental problem that breaks all current RPA tools.

• AI Flash Cards - The ultimate AI tool to create personalized flashcards from any text or PDF file in one click. Enhance retention, master topics faster, and study smarter, not harder.

• Grimo AI (Alpha)- Grimo is not just another notebook. It is built for crazy learners and creators. It is where you get access to serious and valuable knowledge, without leaving the notebook. Less time on searching and organizing knowledge. More on learning.

• Promomix - Don't know what to say after shooting a UGC video or product demo video? Use Promomix Generate Voice Over For Your Short Video.

• Vidnoz AI 2.8 - With Vidnoz AI, making videos can be 200% faster by using video templates, AI voiceovers, and AI real avatars. No need to film, prepare lights, hire actors, etc. Simply paste your text and set your preference, and then wait for your engaging videos to be AI-generated in minutes.

• Sonnet - Sonnet is your end-to-end AI meeting assistant. Get more out of your conversations with pre-meeting briefs, customizable AI-generated notes, and fully automated CRM—all without a meeting bot.

This week in AI

• Menteebot -  Menteebot is a humanoid robot developed by Mentee Robotics that displays a complete end-to-end cycle from a verbal command to complex task completion including navigation, locomotion, scene understanding, object detection and localization, grasping and natural language understanding.

• Mixtral 8x22B New Bench mark for Open Model  – Mixtral 8x22B is a large-scale language model from Mistral AI that sets new benchmarks in performance and efficiency. It uses a sparse mixture-of-experts architecture, is natively multilingual, and supports native function calling. The instruct variant is fine-tuned for chat usage and aligned without overshooting the mark.

• Rivos raises $250 million to develop RISC-V AI chips-Rivos has raised $250 million to develop RISC-V AI chips. The company is designing a server chip that combines a CPU and AI accelerator component, optimized for large language models and data analytics. Rivos' unique approach involves designing the software necessary to compile computer code and then designing the processor around it. The company plans to use this capital injection to develop its inaugural server chip and expand its operations.

• Google Axion Processors - Google's Axion Processor is a custom Arm-based chip for data centers, offering high performance and energy efficiency for various workloads like web servers, databases, AI, and more.

• How to Stop ChatGPT’s Voice Feature From Interrupting You - The article from Wired discusses the issue of ChatGPT interrupting users during voice-only interactions and suggests a solution for users to tap on the screen to signal that they have finished speaking. The article also discusses the potential for ChatGPT to be used for cheating in academic settings and the need for technical tools to ensure that text is created by human beings.

Hope this was a helpful issue. That’s going to be all.