• unwind ai
  • Posts
  • Google's Gemma beats Llama-2 💪

Google's Gemma beats Llama-2 💪

PLUS: Adobe's AI Assistant, Elon Musk announces Grok V1.5 coming soon

Today’s top AI Highlights:

  1. Google releases its first-ever open models series Gemma

  2. Elon Musk xAI’s potential partnership with Midjourney

  3. Meta’s Model that understands by watching videos

  4. AI, Revenue, and Influence: the future of creator platforms

  5. Adobe Brings Conversational AI to PDFs

& so much more!

Read time: 3 mins

Latest Developments 🌍

Google’s state-of-the-art Open Models

Google has released Gemma, their first family of lightweight, state-of-the-art open models built from the same research and technology used to create the Gemini models.

Here's everything you want to know about Gemma from Google's technical report 👇

  1. Gemma models share infrastructure components with Gemini models.

  2. Gemma 2B and 7B models are trained on 2T and 6T tokens of text, respectively. The data for training is primarily English from web documents, mathematics, and code.

  3. Gemma models are designed for CPU and on-device applications (2B model) and GPU and TPU deployment (7B model). These are pre-trained and fine-tuned for dialogue, instruction-following, helpfulness, and safety.

  4. Gemma models outperform similarly sized open models on 11 out of 18 text-based tasks. Performance is evaluated across language understanding, reasoning, and safety benchmarks.

  5. Google emphasizes safety and reliability in Gemma's pre-trained models and uses automated techniques to filter sensitive data from training sets.

  6. Models are fine-tuned and reinforced with human feedback for responsible behaviors with extensive evaluations being conducted, including manual red-teaming and adversarial testing.

  7. Responsible Generative AI Toolkit is released with Gemma for safe AI application development. It includes safety classification methodology, model debugging tools, and best practice guidance.

  8. Supports multi-framework tools like Keras 3.0, PyTorch, JAX, and Hugging Face Transformers. Also, compatible across laptops, desktops, IoT, mobile, and cloud.

  9. Gemma models are released on Hugging Face along with inference code and the code for fine-tuning the models.

  10. Gemma 7B seems to be a new model that outperforms LLAMA-2 7B and Mistral 7B on several benchmarks, despite having the same number of parameters.

Make Space for Elon Musk 🙋‍♂️

  1. Potential Partnership with Midjourney: In a recent X Spaces session hosted by Katherine Brodsky, Elon Musk disclosed a forthcoming collaboration with Midjourney to introduce text-to-image generation features on X. This partnership may also include the integration of AI-generated image labeling.

  2. Grok V1.5: Elon Musk announced an upcoming upgrade to the xAI's chatbot, Grok, with the launch of V1.5 slated for release in two weeks. This update will feature an Analysis tool designed to summarize threads and responses. Additionally, the enhanced Grok aims to support content creators by offering grammar-checking and draft improvement functionalities.

  3. Neuralink's Breakthrough: A significant milestone for Musk’s Neuralink was achieved in January, with the successful implantation of the first brain-computer interface chip in a human subject. The patient has completely recovered and can now control a computer mouse purely through thought.Nearulink’s device is about the size of a coin and is designed to be embedded in the skull to interact with brain cells. It reads neuron activities in the brain to allow individuals, particularly those with paralysis or other severe neurological conditions, to control devices like smartphones or computers merely through thought.

Video JEPA: A Leap into Contextual Understanding 🎥

Meta AI just unveiled Video Joint Embedding Predictive Architecture (V-JEPA) which marks a significant milestone in the journey towards Advanced Machine Intelligence (AMI). V-JEPA, a model that excels in understanding detailed object interactions within videos, represents an innovative approach to machine intelligence with a grounded comprehension of the physical world. It's designed to learn and predict in an abstract representation space, focusing on the essence of visual content without getting bogged down by irrelevant details.

Key Highlights:

  1. Self-Supervised Learning: V-JEPA leverages a self-supervised learning approach, significantly enhancing training and sample efficiency by focusing on abstract representations rather than pixel details.

  2. Flexible and Efficient: The model's ability to ignore unpredictable information allows for a more efficient learning process, requiring fewer labeled examples and less effort on unlabeled data.

  3. Innovative Masking Methodology: Employing a strategic masking strategy in both space and time, V-JEPA fosters a deeper understanding of complex world dynamics, setting a new standard for video-based learning models.

Driving Everywhere with LLaDA 🚗💡

Researchers have unveiled LLaDA, a tool designed to assist autonomous driving by enabling vehicles and human drivers to adapt their behavior to conform with traffic rules and customs of any locale worldwide. LLaDA leverages the zero-shot generalization capabilities of large language models (LLMs), interprets local traffic regulations directly from driver handbooks, ensuring safer and more reliable navigation across diverse environments.

Key Highlights:

  1. Adaptive Driving Across Geographies: LLaDA empowers both autonomous vehicles and human drivers to seamlessly adapt to new traffic laws and norms, enhancing safety and efficiency in unfamiliar territories.

  2. LLM-Powered Traffic Rule Interpretation: Utilizing LLMs, LLaDA accurately interprets and applies local traffic rules from natural language descriptions found in driver handbooks, enabling precise behavior adjustment.

  3. Improved Planning and Safety: Through extensive user studies and real-world dataset applications, LLaDA has demonstrated superior capability in adapting motion planning policies, significantly outperforming traditional planning methods in terms of compliance with local traffic regulations and overall safety metrics.

Connect, Create, Earn 👬

An AI-generated model and influencer with 200k+ followers and earns $10k a month. Surprised? Don't be! Meet Aitana Lopez, an AI model that makes more than $10k a month, something that even an entire ad agency won't be able to make.

But do you know the secret behind the success? Here's a platform that is transforming how creators maximize their creative potential, interact with fans, and earn while saving time. Introducing Fanvue, a subscription-based social platform designed for creators to share content, connect with fans, and earn revenue. It offers several amazing features like:

  • AI Messaging: Offers 24/7 personalized AI messaging to keep your audience engaged.

  • Smart Messaging: Use custom list filters, automated messages and mass messaging to increase your earnings.

  • AI Voice Notes: Reply to fans with automated voice notes in your own voice, adding a personal touch without the time investment.

  • Analytics & Insights: Get detailed insights into your content performance, helping you to refine your strategy and grow your audience further.

Ready to elevate your influence? Then you wouldn't want to miss out on this! Go check it out: Fanvue

Tools of the Trade ⚒️

  1. Adobe's AI Assistant: Adobe’s new AI Assistant in beta, integrated into Reader and Acrobat, offers a suite of features including instant summaries and insights from lengthy documents, answers queries, and formats information for easy sharing. With capabilities like intelligent citations, easy navigation, and formatted output, users can swiftly extract key information and transform it into professional-looking content for various purposes.

  1. Retell AI: A conversational voice API to improve the interaction between speech-to-text, language models, and text-to-speech technologies for a more seamless and efficient conversation experience. It offers features like ultra-realistic voices with emotions, interruption handling, and a low latency of approximately 800ms for more natural conversations.

  2. Fusion: Your digital companion that helps you understand changes in your behavior and improve your quality of life. It provides personalized plans and recommendations to help you meet your desired goals. Fusion allows you to reflect on your daily activities and understand their impact on your health and performance.

😍 Enjoying so far, TWEET NOW to share with your friends!

Hot Takes 🔥

  1. Google finally released a small open-source model. At 7B, it can't be used in real-world applications and, more interestingly, seems identical to Llama2's architecture. However, they claim it's the same research as Gemini. Does that imply that Gemini is, in turn, based on Llama-2? 🤔🤔 ~ Bindu Reddy

  2. After trying Gemma for a few hours I can say it won’t replace my mistral 7B models. It’s better than llama 2 but surprisingly not better than mistral. The mistral team really cooked up a model even google can’t top ~ anton

  3. sama is right about one-person billion dollar companies being right around the corner. lexica just shipped a 4k resolution image generation model with a team of just 5 people. do you understand how insane this is? ~ rahul

Meme of the Day 🤡

Society if Nvidia missed earnings

Image

That’s all for today!

See you tomorrow with more such AI-filled content. Don’t forget to subscribe and give your feedback below 👇

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!

Reply

or to participate.