Weekly AI - 28 Oct 2023
Greetings! I'm thrilled to bring you this week's artificial intelligence newsletter. Inside, I've compiled some of the top AI stories in recent news. You'll read quick summaries about major AI product launches, new research milestones, and organizations implementing cutting-edge applications. My aim is to distill these complex topics into clear, engaging overviews accessible to general audiences. Whether you're an AI enthusiast or just looking to stay up-to-date on this transformative technology, I think you'll find these crisp highlights interesting and informative. Let's dive in!
This newsletter is also available in Spanish and Catalan.
IBM's New AI Chip Could Lead to Faster, More Efficient Computing
IBM unveiled its new NorthPole processor chip designed to speed up artificial intelligence. NorthPole brings memory and processing together on one chip, eliminating the need to access external memory. This allows AI tasks like image recognition to run faster while using much less power. According to the researchers, NorthPole's energy efficiency is "mind-blowing" compared to existing architectures. While not large enough for complex AI like ChatGPT, NorthPole shows promise for speed-critical applications like self-driving cars. Its innovative design principles could pave the way for the next generation of energy-efficient, high-performance AI.
Adobe Previews Next-Gen AI Photo Editor
Adobe previewed Project Stardust, a new AI-powered photo editing tool. Powered by Adobe's latest Firefly Model 2, Project Stardust makes it easy for anyone to edit photos by deleting objects, changing backgrounds, and more using AI. While only a sneak preview for now, Project Stardust combines Adobe's advanced AI models to perform complex edits with just simple clicks and drags. Don't miss the following video.
Google Search adds AI image generation and writing tools
Google announced new AI capabilities for its Search Generative Experience (SGE) feature, which allows more natural conversations with its search engine. Users in the U.S. can now generate images from text prompts directly in SGE using Google's Imagen model. They can also customize draft text outputs to be longer/shorter or change the tone. These new features build on SGE's existing abilities to summarize content and define terms. While aimed at enhancing search, Google hopes to introduce the technology responsibly by restricting image generation to adults and implementing filters to block inappropriate/harmful content.
Google rolls out new tools to provide context for images
Google announced new features to give users more information about images they find on the web. The "About this image" tools show an image's history, metadata, and how others have described it. Google says this context can help debunk false claims and misinformation spread through images. The company is also experimenting with using AI to generate descriptions of unfamiliar sites that images come from. These moves aim to curb the spread of misinformation through images, which has become a major issue with the rise of deepfakes and generative AI.
Open Source Takes On Multimodal AI
Earlier this month, researchers released LLaVA-1.5, an open source AI system that can understand images and text. LLaVA-1.5 follows other open source multimodal models like Alibaba's Qwen-VL and models from Google. While not as advanced as OpenAI's GPT-4V, these free models aim to make multimodal AI more accessible. Supporters argue openness drives progress, but critics warn of potential misuse.
Baidu Claims Its New AI Chatbot Matches GPT-4
Chinese tech company Baidu unveiled the latest iteration of its AI chatbot, ERNIE 4.0. During the company's annual flagship event, Baidu CEO Robin Li stated that ERNIE 4.0 "is not inferior in any aspect to GPT-4,". Li demonstrated ERNIE 4.0's abilities live on stage, showing how it could generate creative content like commercials and martial arts novel plots in Mandarin Chinese. He claimed the new bot has significantly improved comprehension, complex response generation, and memory capabilities. While Baidu plans to initially limit access, this announcement suggests rapid advancement in consumer AI chatbots.
Anthropic's AI Assistant Claude Now Available in 95 Countries
Anthropic's natural language processing chatbot, Claude, launched worldwide recently. The conversational AI is now accessible in 95 countries across the globe, greatly expanding from its initial launch in the US earlier this year.
Meta Launches Habitat 3.0 to Train Intelligent Robots in Simulated Environments
Meta launched Habitat 3.0, an advanced simulation system for training intelligent robots. The new version allows robots to collaborate with simulated humans on everyday tasks like cleaning and cooking. According to Meta researchers, Habitat 3.0 reduces robot training time from months to days by providing realistic virtual environments. The system aims to accelerate research into "embodied AI" - robots that can perceive, move, and assist people in the real world.
D-ID launches AI-powered app to create videos from photos
Tel Aviv startup D-ID launched a new mobile app that uses AI to turn still photos into video. Users simply upload a photo, add text for the video script, and choose a voice. D-ID's technology animates the photo in sync with the audio to generate a realistic fake video. This makes video production accessible and affordable for anyone with a smartphone. The app could enable new forms of communication, but may also raise concerns about misuse of deepfake technology.
Nvidia Unveils Eureka Agent to Revolutionize Robot Training
Nvidia announced the development of Eureka, a new AI agent powered by GPT-4 that can train robots to perform complex skills better than human-designed methods. Nvidia stated that Eureka has taught robots pen-spinning and other dexterous tasks by automatically generating reward functions through evolutionary optimization of code written by the large language model. The company claims Eureka outperforms human experts in 83% of tested scenarios across 29 environments, with an average improvement of 52%. This combination of generative AI and GPU simulation could enable major advances in robotic control and animation.
Google Cloud Offers Industry-First Generative AI Indemnification
Google Cloud announced an industry-first two-pronged indemnification policy to protect customers using its generative AI services. The policy covers claims related to both training data and generated output, with Google assuming responsibility for potential legal risks. According to Neal Suggs, VP Legal at Google Cloud, this comprehensive coverage aims to give customers confidence and peace of mind when using Google's latest generative AI offerings.
Amazon Launches New AI Tool to Enhance Product Images for Advertisers
Amazon has introduced a new generative AI tool that allows advertisers to automatically generate customized backgrounds for their product images. By simply typing a text description of the desired scene, advertisers can now create more visually compelling ads that place products in relevant lifestyle contexts. According to Amazon, early testing shows this AI-generated imagery can boost ad performance, with click-through rates increasing by up to 40%. While not perfect, the technology aims to streamline creative production and storytelling for brands marketing on Amazon.
That wraps up this week's AI news! If you found this informative, don't forget to share it with friends and colleagues. And be sure to subscribe to get next week's news straight to your inbox. Thanks for reading!