AI Report by Explainx
Posts
🔥 Google enters the Open Source Race

🔥 Google enters the Open Source Race

AI is taking us all on a spin again, and this time, it’s both the giants and open source. This week newsletter will cover the latest release of Google in open source Gemma, World’s fastest chip - 10x

February 22, 2024

AI is taking us all on a spin again, and this time, it’s both the giants and open source. This week newsletter will cover the latest release of Google in open source Gemma, World’s fastest chip - 10x faster than GPT-3.5, and Reka Flash model for video and audio, we then cover Products and News as usual.

Let’s get rolling.

Stuff you should know

Google’s Gemma

Gemma AI is Google's latest contribution to the world of “open source” generative AI, offering state-of-the-art open models for developers and researchers. This is the first open source model by Google. Built on the same technology as the Gemini models, Gemma aims to democratize AI with two variants, Gemma 2B and Gemma 7B, and a Responsible Generative AI Toolkit for safer applications.

Facts

Model Variants: Gemma 2B and Gemma 7B, with pre-trained and instruction-tuned versions.
Compatibility: Works across JAX, PyTorch, TensorFlow, and is optimized for NVIDIA GPUs and Google Cloud TPUs.
Accessibility: Available on platforms like Kaggle, Colab, and integrated with tools like Hugging Face and NVIDIA NeMo.
Commercial Use: Permitted for responsible commercial usage and distribution for all organizations.
Safety: Developed with a focus on safety, including data filtering and reinforcement learning from human feedback (RLHF).

How to run it?

There are a bunch of ways, the easiest way is shown here.

The World’s fastest AI Chip

Groq, an AI startup, has made headlines with its groundbreaking $20,000 LPU (Language Processing Unit) chip, setting new performance benchmarks in the AI industry. This chip is designed to rival the traditional GPU-led sector, offering unprecedented speed and efficiency for processing large language models (LLMs). Groq's LPU Inference Engine has outperformed competitors in key areas such as latency vs. throughput and total response time, particularly with Meta AI's Llama 2-70b model. Its on-die memory bandwidth of up to 80 TB/s enables it to handle immense data loads, significantly reducing latency and increasing data transfer rates within the chip. This innovation marks a significant shift towards specialized processing units tailored for specific AI and machine learning workloads, potentially catalyzing a new wave of innovation in AI hardware.

Stats and Bullets

Performance: Groq's LPU Inference Engine set new records in processing efficiency for LLMs, outperforming eight other participants in a benchmark conducted by ArtificialAnalysis.ai.
Memory Bandwidth: Offers an on-die memory bandwidth of up to 80 TB/s, far surpassing traditional setups that rely on external RAM.
Speed: Demonstrations have shown Groq's LPU to process text sequences much faster, with capabilities of hitting almost 500 tokens per second, compared to 30-50 tokens handled by GPT-3.5.
Market Impact: Groq's LPU chip is challenging major players like Nvidia, AMD, and Intel in the high-performance AI chip market, with a market valuation exceeding $1 billion as of October 2021.

Business Use Cases

AI-Driven Platforms: Enhances the speed of AI-driven platforms like chatbots and consumer electronics, enabling fluid end-user experiences across various applications.
Generative AI Products: Ideal for instant responses required by generative AI products, significantly improving the performance of applications that rely on large language models.
Specialized AI Workloads: The LPU's design caters specifically to the computationally intensive and sequential nature of language processing tasks, making it a valuable asset for businesses focusing on natural language processing and related AI applications.

Why it matters

Groq's LPU chip marks a big step forward in AI hardware. Unlike CPUs and GPUs, it's built specifically for handling large language models, which boosts speed and efficiency. This is crucial as AI becomes more vital for various tasks. Groq's innovation not only shakes up AI hardware but also paves the way for new advancements in AI and machine learning.

We tried it out, here’s a video demo.

Reka Flash

Reka Flash, developed by Reka AI, is a groundbreaking multimodal and multilingual language model. It integrates visual and auditory sensors for fast, efficient processing of diverse tasks, from document analysis to coding. Here are some key points:

Model Parameters: With 21 billion parameters, Reka Flash is high-performance yet relatively compact compared to its contemporaries.
Multimodal Capabilities: Excelling in language, vision, and video processing, Reka Flash can handle complex questions, visual data reasoning, and video captioning with finesse.
Multilingual Support: Pretrained in over 32 languages, it demonstrates superior performance in commonsense reasoning and multilingual question answering.
Benchmark Performance: Reka Flash outperforms many larger models on various benchmarks, including MMLU, GSM8K, HumanEval, and GPQA, showcasing its efficiency and capability.
Business Use Cases: Ideal for applications requiring rapid processing without compromising output quality, such as content creation, communication platforms, and on-device applications.

I personally love the audio and video processing capabilities, full walkthrough available here.

Handpicked video

This video is about Boximator, Tiktok’s own AI Video Generator (the motions look similar to SORA) - can this be the open source alternative to SORA?

This week in AI

The AI Revolution Is Coming to Samsung's Galaxy Smartwatches: AI features are coming soon to deliver ground-breaking digital health experiences. The CEO of Samsung MX, Dr. TM Roh, emphasizes collaborations and privacy while giving consumers control over their data. Future wearables with enhanced health features, such as fall detection, should be anticipated.

Meet DeepGO-SE - an AI tool revolutionizing protein function prediction. Using advanced language models, it excels in accuracy and efficiency. A game-changer for drug discovery and biotech exploration.

AI Model Predicts Drug Interactions : Using a machine-learning algorithm, researchers can predict interactions that could interfere with a drug’s effectiveness.
BharatGPT Hanooman: Learn about the core of BharatGPT Hanooman, an AI model developed by Reliance Industries in association with nine esteemed Indian Institutes of Technology, headed by IIT Bombay.
Pichai's AI Push at Google: 20 New Products, Including Chatbot Google CEO Sundar Pichai's recent collaboration with the founders, dormant since 2019, aims to launch 20 new AI products. Notably, Google plans a ChatGPT-style chatbot within Search, triggering discussions and Meta'sskepticism. This move signifies a pivotal shift in Google's AI landscape.
Reddit has potentially made a AI content licensing deal with Google, not verified however.

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