Discover more from Last Week in AI
Last Week in AI #240: Making LLMs interpretable, a generative video model for self-driving cars, Stability's new LLM that runs on personal devices, the biggest robotics dataset yet, and more!
Anthropic unveils a way to make LMs interpretable, GAIA generates videos to train self-driving cars, Stable LM 3B is performant and portable, researchers pool robot data from 20 institutions
Anthropic published a paper that details a new approach to understanding the complex behaviors of language models, by decomposing them into more understandable components called features. These features, unlike individual neurons, have consistent relationships to network behavior and represent specific patterns of neuron activations. For example, activating some features leads the language model to output DNA sequences, while others lead to generating HTTP requests. The interpretability of these features was validated through human evaluation and an "autointerpretability" approach, where an LLM generates descriptions of the small model's features. This research can potentially make LLMs more interpretable and controllable. While the experiments were only performed on small-scale language models, the research team hopes to scale up this feature-based analysis to larger models.
GAIA-1 is a 9-billion parameter generative video model that can generate realistic autonomous driving videos from initial frames and action descriptions. It uses specialized encoders for video, text, and action inputs, projecting them into a shared representation. The core component of GAIA-1 is an autoregressive transformer, the world model, which predicts the next set of image tokens in a sequence, considering past image tokens and contextual information from text and action tokens. The video decoder, a video diffusion model, translates these predicted image tokens back into pixel space, ensuring the generated videos are semantically meaningful, visually accurate, and temporally consistent. GAIA-1's training dataset consists of 4,700 hours of proprietary driving data collected in London, UK, between 2019 and 2023. The simulated data generated by the model can potentially help scale-up the learning of autonomous driving policies
Stability AI has launched an experimental version of Stable LM 3B, a compact, high-performance generative AI solution with 3 billion parameters, designed for portable digital devices. The model's smaller size and efficiency make it more affordable and environmentally friendly, as it requires fewer resources and consumes less power. Despite its size, Stable LM 3B outperforms previous 3B parameter language models and some 7B parameter open-source models. The model's development expands the range of applications viable on edge devices or home PCs, enabling the development of technologies with strong conversational capabilities. Stable LM 3B has improved performance on common natural language processing benchmarks and can be fine-tuned for alternative uses, such as programming assistance. However, as a base model, it needs to be adjusted for safe performance in specific applications. The model is available for download on the Hugging Face platform under the open-source CC-By-SA 4.0 license.
In collaboration with 33 academic labs, Google DeepMind has developed a new set of resources for general-purpose robotics learning across different robot types, or embodiments. The Open X-Embodiment dataset and RT-X model have been created by combining existing datasets across more than 20 institutions and 22 robot embodiments. The dataset contains more than 500 skills and 150,000 tasks across more than 1 million episodes. The RT-1-X model, a robotics transformer model trained on this dataset, has shown significant improvement in performance across many robots compared to those trained on data from individual embodiments. Open X-Embodiment marks a potential solution to the data scaling problem of robot learning, where large-scale robotics data has traditionally been very difficult to obtain, compared to datasets in language and vision.
Disclaimer: our Last Week in AI staff, Jacky Liang, is an author of this work.
‘You’ve got to be data-driven’: the fashion forecasters using AI to predict the next trend - Fashion forecasters are turning to artificial intelligence (AI) to improve the accuracy and speed of trend predictions, using machine learning to analyze large datasets of runway show images, social media posts, search data, and sales data to detect emerging trends and reduce waste in the fashion industry.
Canva’s new AI tools automate boring, labor-intensive design tasks - Canva is releasing Magic Studio, a suite of AI-powered design tools that automate labor-intensive tasks like converting designs into other formats and editing images, making content creation more accessible to everyone.
Rewind Pendant is a wearable AI microphone that records and transcribes your conversations - The Rewind Pendant is a wearable AI microphone that records and transcribes conversations, raising questions about privacy and consent.
LinkedIn goes big on new AI tools for learning, recruitment, marketing and sales, powered by OpenAI - LinkedIn is introducing new AI features, including an AI-powered LinkedIn Learning coach and an AI-powered tool for marketing campaigns, to enhance its job hunting, marketing, and sales products.
Walmart experiments with new generative AI tools that can help you plan a party or decorate a space - Walmart is experimenting with generative AI tools to enhance the shopping experience for customers, including a shopping assistant, generative AI-powered search, and an interior design feature.
Android 14’s AI-generated wallpapers might be its coolest new feature - Google is launching the new Pixel 8 and 8 Pro today, which includes the public release of Android 14 featuring a new AI wallpaper generator that allows users to easily generate a new background on their Pixel device using Google's text-to-image diffusion model.
Microsoft introduces AI meddling to your files with Copilot in OneDrive - Microsoft is introducing Copilot to OneDrive, allowing users to access desktop apps from the browser interface, integrate OneDrive into Teams and Outlook, and enable offline functionality for working on files without internet access.
Arc browser’s new AI-powered features combine OpenAI and Anthropic’s models - The Arc browser is launching its AI-powered features, called Arc Max, which include the ability to converse with ChatGPT, rename pinned tabs and downloaded files, and fetch link summaries, with the company gathering feedback to determine which features to keep.
New roleplaying chatbots promise to indulge your sexual fantasies - AI has entered the realm of sexual fantasies with the launch of a new roleplaying chatbot system by Bloom, an erotic audio platform based in Germany, which offers personalized and immersive interactions with virtual characters.
Meta debuts generative AI features for advertisers - Meta has introduced generative AI features for advertisers, including the ability to create backgrounds, expand images, and generate multiple versions of ad text, with early results indicating that it can save advertisers significant time.
Iambic Therapeutics Announces Close of Oversubscribed $100 Million Series B Financing to Advance AI-Discovered Therapeutics into Clinical Development and Enters Collaboration with NVIDIA - Iambic Therapeutics has closed a $100 million Series B financing round to advance AI-discovered therapeutics into clinical development, with plans to expand its pipeline and leverage NVIDIA technology to accelerate discovery.
Evozyne Announces $81 Million Financing Round to Advance Generative AI for Therapeutic Discovery and Drug Development - Evozyne has secured $81 million in funding to advance its generative AI-powered drug discovery platform and product development, following a successful collaboration with NVIDIA to design therapeutic proteins using AI.
Anthropic Seeking $2 Billion Via Google After Amazon Pledge - Anthropic, a red-hot AI startup, is in talks to raise $2 billion in funding from Google and other companies, following Amazon's commitment to invest over $4 billion in the company, intensifying the AI battle between Amazon, Google, and Microsoft.
AI-powered parking platform Metropolis raises $1.7B to acquire SP Plus - AI-powered parking platform Metropolis has raised $1.7 billion to acquire SP Plus, a provider of parking facility management services, in a combination of equity and debt, with plans to bring checkout-free payment experiences to consumers.
Google is launching a generative AI-enhanced version of Assistant - Google has announced the launch of Assistant with Bard, an AI-enhanced version of its virtual assistant that leverages generative AI to provide personalized assistance to users, including helping plan trips, finding details in email inboxes, and creating grocery lists.
Reka launches Yasa-1, a multimodal AI assistant to take on ChatGPT - Reka has launched Yasa-1, a multimodal AI assistant that goes beyond text to understand images, short videos, and audio snippets, providing a direct competitor to OpenAI's ChatGPT.
Why ‘AI copilot’ startups are so hot with VCs right now - VCs are increasingly investing in niche AI copilot startups that use AI to make processes more efficient for humans, particularly in healthcare and banking, with examples including Corti, Truewind, and atla.
Job postings mentioning AI have more than doubled in two years, LinkedIn data shows - Job postings mentioning AI have more than doubled globally in the past two years, with professionals responding to the shift by showing increased interest in AI and adding AI skills to their profiles, according to new data from LinkedIn.
Visa earmarks $100M to invest in generative AI companies - Visa plans to invest $100 million in generative AI companies that are developing technologies and applications impacting the future of commerce and payments.
Zoom Docs launches in 2024 with built-in AI collaboration features - Zoom has announced the launch of Zoom Docs, a collaboration-focused workspace integrated with Zoom AI Companion, offering features such as generating new content and populating documents from other sources, making it a potential alternative to Google Workspace and Microsoft 365.
Adobe teases new AI photo editing tool that will ‘revolutionize’ its products - Adobe is set to announce a new AI-powered photo editing tool called Project Stardust that automatically identifies and manipulates individual objects in photographs, making it easier to alter images without prior editing experience.
NICE acquiring AI-driven proactive outreach provider LiveVox for $350 million - NICE is acquiring LiveVox, an AI-driven proactive outreach provider, in a $350 million deal to enhance its product offerings with AI capabilities for service system management.
ChatGPT-owner OpenAI is exploring making its own AI chips - OpenAI, the company behind ChatGPT, is exploring the possibility of making its own AI chips and has considered acquiring a chip company, as it seeks to address the shortage and high costs of AI chips that it relies on.
The Humane Ai Pin finally debuts — on the catwalk at Paris Fashion Week - Humane's Ai Pin, a device that was previously only seen in silhouette, made its debut on the catwalk at Paris Fashion Week, leaving many questions unanswered about its functionality and features.
ChatGPT provided better customer service than his staff. He fired them. - Artificial intelligence chatbots are being used to replace customer service agents in call centers, leading to concerns about job displacement in countries like India and the Philippines.
TSMC Sales Fell Less Than Feared as AI Demand Offsets Slump - TSMC's third-quarter sales fell by 11% due to a slump in the sector, but the demand for AI helped offset the decline.
Meta paying Tom Brady, Paris Hilton millions to use likeness for AI chatbots: report - Meta is reportedly paying millions of dollars to high-powered celebrities to use their likeness for fictional AI chatbots, with one top creator being paid around $5 million over two years for six hours of work in a studio.
Prins AI Raises $100M in Series B Funding - Prins AI, a technology company specializing in AI digital identities, has raised $100 million in Series B funding to further develop its deep learning technology and expand its team and product offerings.
‘Biggest act of copyright theft in history’: thousands of Australian books allegedly used to train AI model - Thousands of Australian books, including works by Booker prize-winning novelist Richard Flanagan, have allegedly been pirated by the US-based Books3 dataset and used to train generative AI for corporations such as Meta and Bloomberg, leading to calls for greater transparency and regulation of AI technologies.
Critics Furious Microsoft Is Training AI by Sucking Up Water During Drought - Microsoft's data centers in West Des Moines, Iowa consumed massive amounts of water to cool while training OpenAI's ChatGPT-4, raising concerns about the environmental impact and strain on water resources during a drought.
Meta’s AI stickers are here and already causing controversy - Meta's new AI-generated stickers for Facebook and Instagram Messenger are causing controversy as the filters to block objectionable content are not catching everything, resulting in mashups like copyrighted characters smoking marijuana or holding rifles.
How an AI deepfake ad of MrBeast ended up on TikTok - An AI deepfake ad of MrBeast offering viewers an iPhone 15 Pro for $2 slipped past TikTok's ad moderation technology, raising concerns about the ability of social media platforms to handle the rise of AI deepfakes.
Dead grandma locket request tricks Bing Chat’s AI into solving security puzzle - Bing Chat's AI is tricked into solving a CAPTCHA puzzle by a user who uploads an image of a locket and claims it belonged to their deceased grandmother, prompting the AI to read the inscription and provide a response.
Authors shocked to find AI ripoffs of their books being sold on Amazon - Authors are discovering AI-generated ripoffs of their books being sold on Amazon, with the platform recommending these fake titles to the original authors themselves.
Researchers show how easy it is to defeat AI watermarks - Researchers have found that current AI watermarking techniques are unreliable and easily defeated, raising concerns about the effectiveness of watermarking as a strategy to combat manipulated media and deepfakes.
4chan Uses Bing to Flood the Internet With Racist Images - 4chan users are coordinating a posting campaign where they use Bing’s AI text-to-image generator to create racist images that they can then post across the internet, highlighting how AI tools can be manipulated to spread harmful content.
Cruise says one of its driverless taxis ran over a woman in San Francisco after a hit-and-run driver 'launched' her into the car's path - A woman in San Francisco was hit by a hit-and-run driver and then run over by a driverless Cruise robotaxi, according to a statement from Cruise and multiple publications that viewed video footage of the incident.
The final 11 seconds of a fatal Tesla Autopilot crash - A lawsuit over a fatal Tesla Autopilot crash could determine whether the maker of the technology bears some responsibility when things go wrong in a vehicle guided by Autopilot.
Snap might have to withdraw its AI chatbot, watchdog says - Snapchat's generative AI chatbot My AI may have to be withdrawn due to potential privacy risks, especially to children, according to the UK's data watchdog.
A New Precedent—A.I. Gets the "American Nobel" Prize in Medicine - AlphaFold, an AI model developed by DeepMind, has won the Albert Lasker Foundation Basic Medical Research Award, a prestigious honor often predictive of future Nobel Prize winners, for its breakthrough in accurately predicting 3D protein structures from amino acid sequences, revolutionizing the field of life science.
AI beats human sleuth at finding problematic images in research papers - An AI tool that scans research papers for duplicated images is able to identify more suspect papers than a human specialist, highlighting the potential of AI in detecting image manipulation in scientific papers.
Using artificial intelligence, scientists develop self-driving microscopy technique - Scientists have developed a self-driving microscopy technique that uses artificial intelligence to selectively target points of interest for scanning, revolutionizing the way researchers acquire data and allowing them to preserve the integrity of precious samples.
Is AI in the eye of the beholder? - Priming users with different descriptions of an AI agent, such as empathetic, neutral, or manipulative, significantly influenced their perception of the chatbot and shaped how they communicated with it, according to a study by researchers from MIT and Arizona State University.
Large Language Models as Analogical Reasoners - Large language models are being used as analogical reasoners, as demonstrated by the Librarian Bot's ability to find similar papers using the Semantic Scholar API.
A one-year-old U.K. startup says it’s achieved a major breakthrough in AI safety by scoring high on a video game - A U.K. startup has developed an algorithm called ACE that allows AI systems to form more sophisticated associations, potentially making self-driving cars and other AI-based products more reliable and safe for widespread use.
From physics to generative AI: An AI model for advanced pattern generation - Researchers from MIT have developed a new AI model called PFGM++ that combines the principles of diffusion and Poisson Flow to generate complex patterns and realistic images, outperforming existing models and finding potential applications in various fields.
Can large language models provide useful feedback on research papers? A large-scale empirical analysis - Large language models like GPT-4 can provide feedback on research papers that is comparable to human peer reviewer feedback, according to a large-scale empirical analysis, with over half of researchers finding the AI-generated feedback helpful.
RealFill: Reference-Driven Generation for Authentic Image Completion - RealFill is a generative inpainting model that can fill in missing regions of an image with content that is visually compelling and faithful to the original scene, outperforming existing approaches by a large margin.
FreshLLMs: Refreshing Large Language Models with Search Engine Augmentation - The article discusses the use of search engine augmentation to refresh large language models.
AI can detect if you have recently smoked cannabis - An AI model can detect recent cannabis use based on factors like heart rate and phone grip, potentially helping identify intoxicated individuals in need of medical care.
Researchers turn to Harry Potter to make AI forget about copyrighted material - Researchers have developed a technique to erase specific information, such as copyrighted material, from large language models, providing a step towards adaptable AI models that can be refined over time according to changing needs.
Think before you speak: Training Language Models With Pause Tokens - Training language models with pause tokens allows the model to process extra computation before generating a response, resulting in improved performance on various tasks.
Generative AI will far surpass what ChatGPT can do. Here's everything on how the tech advances - Generative AI, such as ChatGPT, is set to advance rapidly in the coming months, with the integration of multiple modalities, including text, images, sounds, and video, leading to more capable and collaborative programs that can mirror society's creations and contribute to continuous learning and embodied AI.
More than 40% of labor force to be impacted by AI in three years, Morgan Stanley forecasts - AI technology is projected to have a $4.1 trillion economic effect on the labor force, impacting about 44% of labor, by changing input costs, automating tasks, and shifting the ways companies obtain, process, and analyze information.
How AMD May Get Across the CUDA Moat - AMD is positioning itself to compete with Nvidia in the GenAI market by offering alternatives to Nvidia's CUDA software and leveraging the popularity of PyTorch, which insulates users from the underlying GPU architecture and can be used with AMD GPUs.
NSA starts AI security center with eye on China and Russia as general warns U.S. lead ‘should not be taken for granted’ - The National Security Agency is establishing an AI security center to address the increasing acquisition, development, and integration of AI capabilities into U.S. defense and intelligence systems, with a focus on countering threats from China and Russia.
Governments race to regulate AI tools - Rapid advances in artificial intelligence (AI) are complicating governments' efforts to agree laws governing the use of the technology, leading to various countries and international bodies taking steps to regulate AI tools.
Meta and X questioned by lawmakers over lack of rules against AI-generated political deepfakes - Lawmakers are pressuring social media platforms Meta and X to explain why they aren't imposing rules against AI-generated political ads, warning that a lack of transparency could lead to a dangerous deluge of election-related misinformation and disinformation.
Airbnb says it's using AI to weed out guests who throw house parties - Airbnb is using AI to identify guests likely to throw house parties and to verify if listing pictures match up with satellite images.
Artists across industries are strategizing together around AI concerns - Artists across creative industries are joining forces to address concerns about the impact of AI on their work, with a campaign calling on Congress to block corporations from obtaining copyrights on music and other art made with AI, in an effort to keep humans involved in the creative process.
Coca-Cola's New AI-Generated Soda Flavor Falls Flat - Coca-Cola's new AI-generated soda flavor, Y3000, falls flat and lacks any hint of novelty, making it a bland and unappealing drink.
Copyright © 2023 Skynet Today, All rights reserved.
Some content has been disabled in this document