Last Week in AI - A Weekly Unwind
From 9-Jun-2024 to 15-Jun-2024
It was yet another thrilling week in the AI field with advancements that further extend the limits of what can be achieved with AI.
Here are 10 AI breakthroughs that you can’t afford to miss 🧵👇
Apple Intelligence is the New AI 🌟
Apple held its biggest event of the year WWDC 2024. It made sure to highlight the word AI, but it wasn’t Artificial Intelligence, it was “Apple Intelligence.” Integrating generative AI deeply into the new OS, Apple has made its devices and Siri smarter, more personal, and intuitive to use.
Apple Intelligence - It is a personal intelligence system that integrates generative AI at the OS level. It draws from your personal context to assist you in tasks related to language and media, and can also take actions on your behalf. It will be available for free with the new iOS, iPadOS, and MacOS.
Using Apple Intelligence, Siri will now be more natural, helpful, and contextually relevant. Siri can understand the intent behind your request better and let you ask follow-up questions.
Siri will be able to use ChatGPT’s expertise when it finds ChatGPT to be more helpful. This would be helpful for questions related to documents, PDFs, PPTs, and more.
Math Notes in iPad is a whole new experience of using the calculator. Rather than looking for symbols and expressions, you can start writing with the Apple Pencil like you would on a piece of paper, and it automatically calculates as soon as you write “=”.
Apple has also confirmed that they are looking to integrate more AI models in the future, including Google’s Gemini. You might be able to choose the model you want Siri to rely on for assistance, giving you more control over how you interact with your Apple devices.
Future of Prosthetics is Here, It’s Mind-Controlled 🦾
California-based company Atom Limbs is developing a new AI-powered bionic arm that offers amputees unprecedented functionality. The non-invasive arm
uses advanced sensors and machine learning to interpret brain signals,
provides near-human dexterity and individual finger control, and
a basic sense of touch.
It attaches via a sportswear-style vest for comfort and easy weight distribution, and its intuitive AI technology adapts to the user’s movements, reducing the learning curve. Priced at $20,000, it is more affordable than many other bionic options.
Qwen2 Outperforms Llama 3 Across All Benchmarks 🥇
The Qwen team has released Qwen2, the latest version of their LLM suite, featuring improved performance and multilingual capabilities. The suite includes models ranging from 0.5B to 72B parameters, now supporting 27 additional languages.
Qwen2 excels in coding and mathematical tasks, outperforming Llama 3 70B and Mistral 8x22B on benchmarks like MMLU and GSM8K. Additionally, Qwen2 models support longer context lengths up to 128K tokens and have enhanced safety measures to handle multilingual prompts and minimize harmful outputs.
RAG Does Not Work for Enterprises 🏣
Implementing Retrieval-Augmented Generation (RAG) in enterprises faces several significant challenges, particularly in compliance-regulated industries like healthcare and finance. A new paper explores these challenges, emphasizing the need for secure, accurate, and scalable solutions. It also proposes an evaluation framework to validate RAG solutions and outlines the requirements for successful enterprise deployment. Collaboration between researchers and industry partners is crucial to advancing RAG technology in enterprise settings.
Apple Explains How It's Building Apple Intelligence 🧠
Apple has revealed the AI models powering its new Apple Intelligence system in a blog post. The system uses a ~3 billion parameter on-device model for local tasks and a larger server-based model for complex requests, running on Apple silicon servers.
The models, trained with Apple’s AXLearn framework on a mix of licensed and public data, are fine-tuned using adapters for specific tasks. The on-device model outperforms larger models in benchmarks, while the server model rivals GPT-3.5-Turbo. Additionally, these models have lower violation rates in safety evaluations.
Open Dataset of 1,405 Jailbreak Prompts 🔓
A new study reveals the vulnerability of LLMs like GPT-4 to “jailbreak prompts” designed to bypass safety measures and generate harmful content. Researchers analyzed over 1,400 such prompts, highlighting the strategies used to manipulate these AI models.
The study found that LLMs are highly susceptible to these attacks, with some prompts eliciting harmful responses in over 95% of cases. Current safeguards are often insufficient, and the research underscores the urgent need for more robust protections. Additionally, the researchers have made a dataset of over 15,000 prompts publicly available to aid in improving LLM safety.
OpenAI’s Sora Competitor Available to Use 🌅
Luma Labs has released Dream Machine, an AI video model that creates realistic videos from text descriptions and images, available for use unlike OpenAI’s Sora. Trained directly on video, Dream Machine generates physically accurate and action-packed videos, producing 120 frames in 120 seconds. The model excels in creating high-quality, 5-second videos with smooth motion, consistent characters, and cinematic camera moves, allowing for rapid iteration and experimentation with different styles.
Stability AI’s Most Sophisticated Image Generation Model 🪄
Stability AI has released Stable Diffusion 3 Medium, the latest version of their text-to-image AI model, featuring significant improvements in image quality, particularly in rendering hands and faces, and better understanding of complex prompts.
The model is smaller, allowing it to run on consumer-grade hardware without sacrificing performance. Stability AI used synthetic and filtered publicly available data for training, including 1 billion pre-training images and 30 million high-quality aesthetic images for fine-tuning.
Samsung Unveils Roadmap for the Future of AI 💡
Samsung recently held its annual Samsung Foundry Forum (SFF), focusing on the company’s vision for the future of AI and showcasing its latest advancements in semiconductor technology. Samsung announced two new process nodes – SF2Z and SF4U – designed to improve performance and power efficiency for AI chips. The company also revealed its Samsung AI Solutions platform, which integrates its Foundry, Memory, and Advanced Package (AVP) businesses to offer comprehensive AI solutions.
Humanoid Robot Takes the Driving Wheel 🚗
Autonomous driving by humanoid robots is an active field of research, but current robots lack the flexibility and human-like proportions needed to naturally interact with vehicles designed for humans. However, a new project using a musculoskeletal humanoid robot named Musashi is making progress.
Musashi mimics the human body and is built with a structure such that it can perform tasks like sitting on a car seat and operating a steering wheel with both arms, unlike previous humanoid robots.
Which of the above AI development you are most excited about and why?
Tell us in the comments below ⬇️
That’s all for today 👋
Stay tuned for another week of innovation and discovery as AI continues to evolve at a staggering pace. Don’t miss out on the developments – join us next week for more insights into the AI revolution!
Click on the subscribe button and be part of the future, today!
📣 Spread the Word: Think your friends and colleagues should be in the know? Click the ‘Share’ button and let them join this exciting adventure into the world of AI. Sharing knowledge is the first step towards innovation!
🔗 Stay Connected: Follow us for AI updates, sneak peeks, and more. Your journey into the future of AI starts here!
Shubham Saboo - Twitter | LinkedIn ⎸ Unwind AI - Twitter | LinkedIn