Top News
Introducing Devin, the first AI software engineer
Cognition AI unveiled Devin, an AI that aims to automate tasks done by software engineers. Devin is capable of planning and executing complex engineering tasks, learning over time, and fixing mistakes. It can also use common developer tools and actively collaborate with users. Devin's abilities include learning unfamiliar technologies, building and deploying apps end-to-end, autonomously finding and fixing bugs in codebases, training and fine-tuning its own AI models, and contributing to mature production repositories. Devin was evaluated on the SWE-bench benchmark, resolving 13.86% of real-world GitHub issues, far exceeding the previous state-of-the-art of 1.96%. Cognition is currently offering early access to Devin for engineering work.
A generalist AI agent for 3D virtual environments
Google DeepMind has introduced a new generalist AI agent, the Scalable Instructable Multiworld Agent (SIMA), capable of following natural-language instructions to perform tasks in various video game environments. SIMA, which doesn't require access to a game's source code or APIs, uses pre-trained vision models and a main model that includes a memory and outputs keyboard and mouse actions. The agent was trained and tested on nine different video games in collaboration with eight game studios, demonstrating its ability to understand a broad range of gaming worlds. The research aims to translate the capabilities of advanced AI models into useful, real-world actions through a language interface, with the ultimate goal of creating AI systems that can understand and safely carry out a wide range of tasks in a way that is helpful to people online and in the real world.
Reddit’s Sale of User Data for AI Training Draws FTC Inquiry
Reddit, ahead of its IPO, has disclosed that it could generate $203 million in revenue over the next few years by licensing user posts to Google and others for AI projects. However, this new business line has drawn the attention of the US Federal Trade Commission (FTC), which has sent a letter to Reddit inquiring about the sale, licensing, or sharing of user-generated content with third parties for AI training. The FTC, which has the power to sanction companies for unfair or deceptive trade practices, is interested in the privacy risks, fairness, and copyright issues surrounding this practice. Reddit is not the only company involved in data licensing for AI, with Stack Overflow, the Associated Press, and Tumblr owner Automattic also engaging in similar deals. The use of online data for AI training has raised questions about content ownership, fairness of licensing without compensating creators, potential data leaks, and the risk of further empowering dominant companies.
Open Release of Grok-1
xAI has announced the open release of Grok-1, a large language model with a 314 billion parameter Mixture-of-Experts model. The model, which was trained from scratch, is not fine-tuned for any specific application, such as dialogue. The base model weights and network architecture are being released under the Apache 2.0 license. Grok-1 was trained using a custom training stack on top of JAX and Rust, with 25% of the weights active on a given token. Instructions for using the model can be found on xAI's GitHub page.
Other News
Tools
Pika Labs just added sound effects to its generative AI videos — here’s how it sounds - Pika Labs has added the ability to create sound effects from a text prompt for its generative artificial intelligence videos, allowing users to add realistic sounds like bacon sizzling or lions roaring to their videos.
AI tool predicts kidney failure six times faster than human expert analysts - AI tool predicts kidney failure six times faster than human analysts, providing accurate and super-fast analysis of total kidney volume, potentially revolutionizing kidney clinics worldwide.
Inflection-2.5: meet the world's best personal AI - Inflection-2.5, the upgraded personal AI model, offers competitive performance with leading LLMs like GPT-4, incorporating enhanced IQ and EQ features, resulting in significant user engagement and retention.
New Adobe Express Mobile App Brings Firefly Generative AI Models Directly into Mobile Workflows - Adobe has released a new mobile app, Adobe Express, which integrates Firefly generative AI models to revolutionize content creation on-the-go for everyone, with features like Text to Image, Generative Fill, and quick video editing.
Salesforce announces new AI tools for doctors - Salesforce introduces AI tools for doctors to streamline administrative tasks and unify health data from various sources, aiming to alleviate physician burnout and improve patient care.
Microsoft’s neural voice tool for people with speech disabilities arrives later this year - Microsoft is introducing new accessibility features, including a neural voice tool for people with speech disabilities, as well as updates to its AI-powered assistive products.
Amazon’s generative AI bot Rufus makes online shopping easier (for the most part) - Amazon's new generative AI chatbot Rufus is designed to make online shopping quicker and easier by providing answers to questions about product categories and individual items, although it still has some limitations.
Microsoft begins blocking some terms that caused its AI tool to create violent, sexual images - Microsoft's Copilot AI tool is being modified to block requests for generating violent, sexual, and controversial images following concerns raised by a staff AI engineer.
DoorDash’s new AI-powered ‘SafeChat+’ tool automatically detects verbal abuse - DoorDash introduces AI-powered 'SafeChat+' to automatically detect and address verbal abuse in customer and delivery interactions, aiming to reduce safety incidents on its platform.
The AI dolls to tackle loneliness of South Korea’s elderly (and watch them) - AI companion doll for seniors, showcased at the Mobile World Congress, aims to tackle loneliness and provide emotional support to the elderly, with features such as medication reminders and constant monitoring.
Oracle adds generative AI features to finance, supply chain software - Oracle is adding generative AI features to its corporate software lineup, aiming to save time for businesses by automating tasks such as report generation and data summarization.
Amazon Adds Another Generative AI Feature to Automate Product Listings - Amazon introduces a generative AI tool that automates product listings by pulling information from a URL, potentially increasing the presence of AI-generated listings on the platform.
Tencent’s latest AI tool animates static images with simple prompts - Tencent has collaborated with academic partners to introduce an image-to-video AI model called Follow-Your-Click, which allows users to animate still images with simple text prompts, addressing issues faced by other models in the market.
Croissant: a metadata format for ML-ready datasets - Introducing Croissant, a new metadata format for ML-ready datasets, developed to reduce the data development burden and enable a richer ecosystem of ML research and development, with support from major tools and repositories.
Business
Building Meta’s GenAI Infrastructure - Meta is investing in AI infrastructure, sharing details on two 24,576-GPU data center scale clusters, aiming to continue growing their infrastructure build-out to include 350,000 NVIDIA H100 GPUs by 2024, and emphasizing their commitment to open innovation in AI software and hardware.
Apple Buys Canadian AI Startup as It Races to Add Features - Apple acquires Canadian AI startup DarwinAI to bolster its artificial intelligence division and prepare for a major push into generative AI in 2024.
How the A.I. That Drives ChatGPT Will Move Into the Physical World - A start-up founded by former OpenAI researchers is using chatbot technology to develop A.I. that can navigate the physical world, enabling robots to understand their surroundings and interact with humans.
Optimus who? Figure humanoid robot’s new talk power topples Tesla fame - OpenAI collaborates with Figure AI to enhance humanoid robot capabilities and develop next-generation AI models for swift, precise robot actions.
AI-Generated Marilyn Monroe Chatbot Can Hold an Extended Conversation With ‘Realistic Emotions’ and Expressions, Company Claims - AI technology firm Soul Machines has created an AI-generated digital avatar of Marilyn Monroe that can hold extended conversations with realistic emotions and expressions, aiming to connect brands and consumers through immersive interactions.
AI Poll: 68% Use artificial intelligence at work, nearly half for productivity - AI has captured the corporate consciousness, with 68% using it at work, primarily for productivity, and a majority having aggressive or balanced deployment strategies.
Research
Mathematicians use AI to identify emerging COVID-19 variants - AI framework developed by mathematicians at The Universities of Manchester and Oxford can quickly identify and track new and concerning COVID-19 variants, potentially enabling a more proactive response and tailored vaccine development.
SaulLM-7B: A pioneering Large Language Model for Law - Introducing SaulLM-7B, a pioneering large language model tailored for the legal domain, with 7 billion parameters and state-of-the-art proficiency in understanding and processing legal documents.
ShortGPT: Layers in Large Language Models are More Redundant Than You Expect - Large Language Models (LLMs) contain redundant layers, and a new pruning approach called ShortGPT significantly outperforms previous methods in model pruning.
Learning to Decode Collaboratively with Multiple Language Models - Teaching multiple large language models to collaborate by interleaving their generations at the token level improves performance in cross-domain settings and specific tasks, without direct supervision.
PALO: A Polyglot Large Multimodal Model for 5B People - Introducing PALO, a Large Multilingual Multimodal Model with visual reasoning capabilities in 10 major languages, covering 5B people, and incorporating diverse instruction sets to boost performance across underrepresented languages.
Enhancing Vision-Language Pre-training with Rich Supervisions - Using web screenshots for Strongly Supervised pre-training with ScreenShots (S4) significantly enhances the performance of image-to-text models in various downstream tasks, demonstrating improvements of up to 76.1% on Table Detection and at least 1% on Widget Captioning.
Rainbow Teaming: Open-Ended Generation of Diverse Adversarial Prompts - A novel approach called Rainbow Teaming uses open-ended search to generate a diverse collection of adversarial prompts, improving the safety of large language models without compromising their general capabilities.
AtomoVideo: High Fidelity Image-to-Video Generation - AI has led to significant advancements in video generation, and AtomoVideo presents a high fidelity framework for image-to-video generation, achieving superior results compared to popular methods.
Data Interpreter: An LLM Agent For Data Science - Introducing the Data Interpreter, a solution designed to enhance problem-solving in data science through dynamic planning, tool integration, and logical inconsistency identification, demonstrating superior performance in various tasks.
US Army tests AI chatbots as battle planners in a war game simulation - US Army Research Laboratory tests AI chatbots for battle planning in war game simulations, but experts caution against using AI in high-stakes situations.
Stealing Part of a Production Language Model - Attackers can extract precise information from black-box production language models, confirming hidden dimensions and recovering projection matrices for a low cost.
Concerns
Top AI researchers say OpenAI, Meta and more hinder independent evaluations - Top AI researchers are calling on generative AI companies to allow independent access to their systems, arguing that strict protocols are hindering safety-testing and independent evaluations.
Florida Middle Schoolers Arrested for Allegedly Creating Deepfake Nudes of Classmates - Florida middle schoolers arrested for allegedly creating and sharing AI-generated nude images of classmates without consent, leading to the first criminal charges under a state law making it a felony to share altered sexual depictions of a person without their consent.
Political operative and firms behind Biden AI robocall sued for millions - Political operative and firms behind Biden AI robocall sued for millions after a fake robocall using AI to impersonate Joe Biden urged Democrats not to vote, violating federal and state laws and prompting a lawsuit seeking damages and a permanent injunction.
Policy
European lawmakers approve world's first major act to regulate AI - European Union Parliament approves world's first major set of regulatory ground rules to govern artificial intelligence, dividing the technology into categories of risk and aiming for implementation by 2025.
US spearheads first UN resolution on artificial intelligence — aimed at ensuring equal access - US spearheads first UN resolution on artificial intelligence, emphasizing the need for safe, secure, and trustworthy AI systems and equal access for all countries, especially those in the developing world.
Five of this year’s Pulitzer finalists are AI-powered - Five of this year’s Pulitzer finalists for journalism used AI in their submissions, prompting discussions about AI policies and its potential impact on investigative reporting.
Google restricts election-related queries for its Gemini chatbot - Google restricts election-related queries for its Gemini chatbot in response to concerns about misinformation and AI-generated content, with the changes already rolled out in India and the U.S.
India drops plan to require approval for AI model launches - India walks back on AI approval requirement, now advises labeling unreliable models and ensuring lawful content, following criticism from entrepreneurs and investors.
Copyright © 2024 Skynet Today, All rights reserved.