Top News
OpenAI unveils GPT-4o mini, a small AI model powering ChatGPT
OpenAI has launched GPT-4o mini, a smaller, faster, and more cost-effective AI model than its predecessors. The model, which outperforms other small AI models in text and vision reasoning tasks, is being made available to developers and consumers via the ChatGPT web and mobile app, with enterprise users gaining access next week. GPT-4o mini scored 82% on the MMLU reasoning benchmark and 87% on the MGSM math reasoning benchmark, outperforming other models like Gemini 1.5 Flash and Claude 3 Haiku. The model is also significantly cheaper to run than previous models, being over 60% cheaper than GPT-3.5 Turbo. OpenAI also announced new tools for enterprise customers to help businesses in regulated industries comply with logging and audit requirements.
Meta releases the biggest and best open-source AI model yet
Meta has released Llama 3.1, the largest open-source AI model, claiming it outperforms top private models like GPT-4o and Claude 3.5 Sonnet. With 405 billion parameters, Llama 3.1 was developed using over 16,000 Nvidia H100 GPUs, costing Meta hundreds of millions of dollars. Despite the high cost, Meta offers Llama 3.1 with an open-source license, requiring approval only from large companies, aiming to emulate the success of open-source software like Linux. Meta collaborates with major tech companies to help developers deploy customized versions of Llama 3.1, which is said to cost half as much to run as OpenAI’s GPT-4o. The model uses synthetic data to enhance its performance and has been rigorously tested for cybersecurity and biochemical applications. Meta’s AI assistant, integrated into WhatsApp, Instagram, and Facebook, will initially use Llama 3.1, switching to a smaller model after a usage threshold. CEO Mark Zuckerberg predicts that open-source AI, spearheaded by Llama 3.1, will dominate the industry, similar to previous open-source projects.
Hugging Face Releases SmoLLM, a Series of Small Language Models, Beats Qwen2 and Phi 1.5Â
Hugging Face has introduced SmoLLM, a new series of compact language models available in three sizes: 130M, 350M, and 1.7B parameters. These models are designed for use on local devices, reducing the need for cloud-based resources and energy consumption. The SmoLLM models, trained on FineWeb-Edu and Cosmopedia v2 datasets, outperform existing models in their size categories, including MobileLM-125M, Qwen2-500M, Phi 1.5, and MobileLM-1.5B. This release underscores the growing interest in small language models for local operation, offering benefits in data privacy and cost reduction, and Hugging Face’s commitment to transparency and open-source resources.
Apple, Anthropic and other companies used YouTube videos to train AI
A massive dataset containing subtitles from over 170,000 YouTube videos was used to train AI systems for major tech companies like Apple, Anthropic, Nvidia, and Salesforce without permission, as revealed by Proof News and Wired. The dataset, derived from more than 48,000 channels, includes subtitles from popular creators and major news outlets but not the actual video imagery. This practice raises significant ethical and legal questions, especially since using such data appears to violate YouTube’s terms of service. Despite repeated inquiries, companies like OpenAI have been vague about the specifics of their data sources. The dataset, part of the larger open-source collection by EleutherAI called The Pile, highlights ongoing transparency issues in AI training data usage.
Original source: Apple, Nvidia, Anthropic Used Thousands of Swiped YouTube Videos to Train AI
Other News
Tools
OpenAI just dropped 2 new Sora videos — and they look very impressive – OpenAI has released two new Sora videos, showcasing the model’s impressive capabilities, but its public availability remains uncertain as other AI models are catching up in rendering quality and motion accuracy.
Mistral AI and NVIDIA Unveil Mistral NeMo 12B, a Cutting-Edge Enterprise AI Model – Mistral AI and NVIDIA have unveiled Mistral NeMo 12B, a cutting-edge enterprise AI model that offers unprecedented accuracy, flexibility, and efficiency for diverse applications, and can be easily customized and deployed for enterprise use.
Anthropic releases Claude app for Android – Anthropic releases Claude Android app to expand AI chatbot accessibility, offering features like real-time language translation and image analysis, aiming to compete with ChatGPT.
Mistral releases Codestral Mamba for faster, longer code generation – Mistral releases Codestral Mamba, a new code generating model based on the Mamba architecture, which aims to improve efficiency and speed for programmers and developers.
Helm.ai launches VidGen-1 generative video model for autonomous vehicles, robots – Helm.ai launches VidGen-1, a generative AI model for autonomous vehicles and robots, which allows cost-effective training on thousands of hours of driving footage and could be applied to various domains.
Exclusive: meet Haiper 1.5, the new AI video generation model challenging Sora, Runway – Haiper 1.5, a new AI video generation model, challenges competitors with longer video clips, an upscaler capability, and plans for image generation, aiming to create true-to-life content and compete with industry leaders.
Google brings AI agent platform Project Oscar open source – Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Project Oscar, announced during Google I/O Bangalore, is an open-source platform that can help software product teams monitor issues or bugs.
Salesforce debuts Einstein Service Agent, a new AI Agent for customer self-service – Salesforce introduces Einstein Service Agent, a new AI-powered self-service experience for customers, designed to provide conversational AI interface and take actions such as product returns or refunds, with seamless integration into Salesforce’s existing customer data and workflows.
Microsoft’s Designer app arrives on iOS and Android with AI editing and creation – Microsoft’s AI-powered Designer app, now available on iOS and Android, allows users to create custom images, stickers, greeting cards, and invitations using templates and AI editing capabilities.
Spotify launches a new voice and language for its AI DJ – Spotify introduces a Spanish-language version of its “AI DJ” feature, allowing users to switch between different voices for personalized music recommendations and radio-like commentary.
Google Vids is available to test out Gemini AI-created video presentations – Google launches Vids, an AI-powered productivity app in Workspace Labs, allowing users to create presentation videos by dropping docs, slides, voiceovers, and video recordings into a timeline.
Introducing Llama-3-Groq-Tool-Use Models – Maximize AI performance by combining specialized tool use models with general-purpose language models, implementing a routing system to analyze queries and direct them to the most suitable model.
Meet Phenomenal AI, the first made in India text to video generator platform – An Indian startup has unveiled Phenomenal AI, the first text-to-video generator in India, to address the increasing demand for video content and make video generation tools more accessible.
Business
TSMC second-quarter revenue jumps on AI boost, handily beats market forecasts – TSMC’s second-quarter revenue significantly surpassed market forecasts, driven by the increasing demand for AI applications, leading to a 32% year-on-year growth.
Google, Microsoft offer Nvidia chips to Chinese companies, the Information reports – Google and Microsoft are providing Chinese companies access to Nvidia’s AI chips through data center services located outside of China, despite the Biden administration’s efforts to restrict the use of US technology for AI in China.
OpenAI reportedly holding talks with Broadcom and others to develop new AI server chip – OpenAI is reportedly talking with chip designers, including Broadcom Inc., about developing a new artificial intelligence chip for servers.
After Tesla and OpenAI, Andrej Karpathy’s startup aims to apply AI assistants to education – Andrej Karpathy, former head of AI at Tesla and researcher at OpenAI, is launching Eureka Labs, an “AI native†education platform that aims to leverage recent progress in generative AI to create AI teaching assistants that can guide students through course materials.
Fujitsu partners with Cohere to build LLMs for Japanese enterprises – Fujitsu partners with Cohere to develop secure generative AI solutions for Japanese enterprises, focusing on building large language models tailored for the Japanese language.
Menlo Ventures and Anthropic team up on a $100M AI fund – Menlo Ventures and Anthropic have teamed up to create a $100 million fund, the Anthology Fund, to invest in pre-seed and Series A AI startups, leveraging their close relationship with Anthropic to identify and support promising companies.
Tech industry teams up to set AI security standards – Top tech companies form coalition to develop cybersecurity and safety standards for AI, aiming to ensure rigorous security practices and keep malicious hackers at bay.
Disney Music Group Teams With AudioShake to Separate Stems of Classic Songs Using AI – Disney Music Group partners with AudioShake to use AI for stem separation and lyric transcription, aiming to enhance fan engagement and create new listening experiences for their classic catalog.
SoftBank acquires UK AI chipmaker Graphcore – SoftBank has acquired UK AI chipmaker Graphcore, with the terms of the deal undisclosed, but Graphcore’s CEO expressing positivity about the acquisition and confirming that there will be no layoffs.
Samsung to launch upgraded voice assistant Bixby this year with its own AI – Samsung is set to launch an upgraded version of its voice assistant Bixby this year, incorporating its own artificial intelligence models and aiming to bring more AI capabilities to its suite of devices.
Stable Diffusion 3 License Revamped Amid Blowback, Promising Better Model – Stability AI revamps its licensing terms for Stable Diffusion 3, allowing free use for research and limited commercial purposes, but still prohibiting the creation of new foundational models using SD3-generated work.
Small language models rising as Arcee AI lands $24M Series A – Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More The trend toward small language models is accelerating as Arcee AI announced its $24M Series A funding only 6 months after announcing its $5.5M seed round in January 2024.
Research
DeepMind’s PEER scales language models with millions of tiny experts – DeepMind introduces PEER, a novel architecture that scales MoE models to millions of experts, improving the performance-compute tradeoff of large language models by efficiently routing input data and using tiny experts with a single neuron in the hidden layer.
Qwen2 Technical Report – The Qwen2 Technical Report introduces the latest large language and multimodal models, showcasing their impressive performance across diverse benchmarks and their robust multilingual capabilities.
OpenAI Is Secretly Working on a New Reasoning Technology Codenamed Project Strawberry – OpenAI is developing a new reasoning technology called Project Strawberry, which aims to enable AI models to conduct autonomous research and improve their ability to answer difficult user queries.
Toto: Time Series Optimized Transformer for Observability – A new state-of-the-art foundation model for time series forecasting, Toto, has been developed by Datadog, outperforming existing models on observability data and achieving state-of-the-art zero-shot performance on multiple open benchmark datasets.
SpreadsheetLLM: Encoding Spreadsheets for Large Language Models – Efficient encoding method, SheetCompressor, significantly improves performance in spreadsheet table detection task and achieves state-of-the-art F1 score, demonstrating effectiveness across a variety of spreadsheet tasks.
MambaVision: A Hybrid Mamba-Transformer Vision Backbone – A novel hybrid Mamba-Transformer vision backbone, MambaVision, is proposed and shown to achieve state-of-the-art performance in image classification and outperform comparably-sized backbones in downstream tasks.
Husky: A Unified, Open-Source Language Agent for Multi-Step Reasoning – Husky is an open-source language agent that outperforms existing models in addressing complex reasoning problems by using a unified action space and expert models.
LMMs-Eval: Reality Check on the Evaluation of Large Multimodal Models – Introducing LMMS-EVAL, a unified multimodal benchmark framework with over 50 tasks and 10 models, addressing the challenges of low cost and zero contamination in evaluating large multi-modal models.
Transformer Layers as Painters – Understanding the impact of removing or reorganizing information throughout the layers of a pretrained transformer can yield better usage of existing models and make architectural improvements to produce new variants, as shown by a series of empirical studies on frozen models.
Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing – Magpie presents a method for synthesizing high-quality instruction data at scale by extracting it directly from aligned large language models, demonstrating its effectiveness in comparison to other public instruction datasets.
GraphFM: A Scalable Framework for Multi-Graph Pretraining – GraphFM is a scalable framework for multi-graph pretraining, but the specific details of the article are not available.
Concerns
Microsoft faces UK competition investigation over hiring of AI startup’s founder and key staff – Microsoft’s hiring of an AI startup’s key staff is under investigation by British regulators over concerns of potential competition issues in the AI market.
AI’s ‘Oppenheimer moment’: autonomous weapons enter the battlefield – A squad of soldiers is under attack and pinned down by rockets in the close quarters of urban combat.
Inside the Face Fraud Factory – Fraudsters are using stock videos and photos of ordinary people to bypass verification checks on cryptocurrency exchanges and other online services, with some individuals selling their faces for as little as $5 to be used in fraudulent activities.
Policy
Trump allies draft AI order to launch ‘Manhattan Projects’ for defense – Trump allies are drafting a sweeping AI executive order to launch “Manhattan Projects” for defense, signaling a potential second Trump administration’s pursuit of AI policies favorable to Silicon Valley investors and companies, including repealing the Biden AI executive order and spurring AI research and development in the United States.
Meta Follows in Apple’s Footsteps by Restricting AI Releases in EU Countries – Meta and Apple are restricting the release of their upcoming AI models in EU countries due to the bloc’s strict regulations, which have already caused potential penalties and threaten their ongoing activity within the region.
Source: Read MoreÂ
