AI News & Launches

Google has released PaliGemma, an open-source multimodal vision language model (VLM) capable of understanding and generating content for images and texts. It outperforms other VLMs with its object detection and segmentation capabilities. PaliGemma is designed for fine-tuning on custom datasets, allowing users to optimize its performance for specific tasks.

Tarsier, created by Reworkd and hosted on GitHub, offers vision utilities for web interaction agents. The goal is to assist LLMs in automating web interactions by providing a system that visually tags interactable elements on web pages, enabling actions such as 'CLICK'. It leverages OCR to create a whitespace-structured string representation of webpage content, which can be understood by LLMs. The utility is compatible with LLMs like GPT-4 and is accessible through pip installation.

Silo AI in cooperation with TurkuNLP and HPLT has released Viking 7B, the first open large language model (LLM) tailored for Nordic languages. Viking 7B is designed to understand and generate Danish, Finnish, Norwegian, Icelandic, and Swedish languages, as well as English and programming languages. Trained on the LUMI supercomputer, this open-source model emphasizes local values and aims to improve multilingual AI applications.

Productboard's AI for PM improves decision-making & boosts productivity by integrating AI into product management, enabling faster shipping of products with AI-aided tasks, such as feedback categorization, trend monitoring, and feature specification.

The article highlights the concern around AI's propensity for 'hallucination' or outputting incorrect information, with recent events revealing the errors made by AI systems like Google's Gemini, Microsoft's Copilot, and OpenAI's ChatGPT. The issue is inherent to AI language models, and despite efforts at mitigation, poses significant risks to dependability.

Gary Marcus offers a critical perspective on OpenAI's latest release, GPT-4o. Acknowledging the improvement in speech synthesis, he notes its similarities with its predecessors and points out the lack of significant advancement beyond GPT-4. The presence of quirky errors and hallucinations in GPT-4o's output suggests that the AI has not notably evolved. Marcus implies that this stagnation in significant capability advances might signal a broader trend of diminishing returns in AI development.

TapScanner is an AI-powered app that transforms your device into a versatile scanner for objects, documents, and more. It offers features like nutritional information scanning, plant identification, math problem solving, object counting, and AI home decor tips.

Receipt-AI is a cutting-edge tool that helps busy traveling teams manage receipts effortlessly via AI and SMS, enabling uploads to Xero or QuickBooks in seconds without Wi-Fi, saving 97% of their time.

Jovu is an AI-powered tool by Amplication designed to generate production-ready code, helping developers accelerate their transition from concept to deployment with adherence to best practices and standards.

Warmy.io's AI product offers state-of-the-art machine learning solutions to enhance business operations by leveraging advanced algorithms and data analysis. Aimed at driving efficiency and innovation.

Glitter AI is a tool currently in beta that enables quick creation of step-by-step guides. It is designed to help with training, support, or any instructional purpose, allowing guides to be generated simply by clicking and speaking.

Google's Gemini update introduces the 1.5 Flash model, designed for speed and efficiency, with a revolutionary long context window. This AI assistant innovation is part of the Gemini model family and enhances AI capabilities in summarization, chat, and multimodal reasoning. The update hints at the future of AI with projects like Project Astra.

The study 'The Platonic Representation Hypothesis' explores the trend of converging representations in AI, particularly in deep networks. It highlights evidence of convergence across models and data modalities, proposing that AI is moving towards a universal model of reality reminiscent of Plato's ideal forms. The paper discusses potential pressures driving this convergence and its broader implications.

A breakthrough in memory for language models has been achieved by GPT-4o as demonstrated by the Needle in a Needlestack benchmark, which challenges models on paying attention to context in a prompt filled with limericks. GPT-4o surpassed other models like GPT-4 Turbo and Claude-3 Sonnet, showing near-perfect performance.

Discussion centers around a feeling of disillusionment after advancements in AI, conveying concerns over tech democratization and impact on the ability to build and maintain competitive technologies.

MICRO LLM is an AI personal assistant app designed to help organize tasks, schedule appointments, and answer questions, enhancing productivity for users. Available on Apple devices including iPhone, iPad, and Mac, it boasts a 4.3-star rating.

Google introduced 'AI Overviews' to refine search results, with AI-powered features such as video-based searches using Google Lens, travel planning tools, and automatic results categorization using the new Gemini AI model.

PaliGemma is an open-source Vision-Language Model inspired by PaLI-3, using images and text to answer questions about pictures. It offers customizable research and general-purpose models with immediate exploration capabilities.

Google AI's Model Explorer is an advanced visualization tool that provides insights into machine learning models, assisting in analysis and accelerating deployment on target devices. It helps with successful inference, quantization identification, and performance optimization.

Canny is a comprehensive customer feedback management platform that enables companies to centralize product feedback, uncover insights, and make informed product decisions. The platform facilitates collecting, analyzing, organizing, and prioritizing feedback, as well as creating roadmaps and sharing updates. It caters to various industry leaders and startups, aiming to streamline feedback for product success.

Google unveils Project Astra at their annual I/O conference, promising a new level of AI-powered digital assistance. Astra is designed to perform real-time tasks and respond conversationally, intending to outperform conventional assistants like Siri and Alexa. Additionally, Google introduces several other Gemini updates aimed at improving performance and interaction.

Tech companies are increasingly engaging with the Middle East, drawn by the region's wealth and investment opportunities. Despite a complex backdrop of political and social issues, the Biden administration is encouraging these ties as a counterbalance to China's influence. Entrepreneurs like Andrew Feldman are exploring partnerships, with his company, Cerebras, receiving substantial funding from the UAE to support projects in the US and the Emirates.

The San Francisco Bay Area has emerged as the leading hub for AI tech and startup funding, accounting for over 50% of global venture funding for AI-related startups in 2023. The trend began with OpenAI’s ChatGPT, which saw rapid user growth. OpenAI, Anthropic, and Inflection AI, all Bay Area-based, each raised over $1 billion. Bay Area companies accounted for 17% of global AI funding deals and surpassed all countries outside the U.S. AI has also propelled the region's leasing market, with companies like OpenAI and Anthropic securing significant office space in San Francisco.

Google's AI Test Kitchen reveals VideoFX, an AI-powered tool that provides illustrative examples of visually vibrant projects, including a colorful crochet elephant. Users are invited to sign in for a sneak peek.

Linum, a YC W23 startup, is seeking a Founding AI Engineer with 2-3 years of experience in PyTorch and Python. The role involves developing text-to-video models in San Francisco with a salary range of $100k-$180k and equity of 0.50%-2.00%.

Model Explorer is an open-source project by Google AI Edge that provides an intuitive and hierarchical visualization of model graphs for debugging. It supports multiple model formats including TFLite, TF, TFJS, MLIR, and PyTorch.

AI-generated spam is inundating Facebook, creating unusual images and causing confusion among users. The images, which range from surreal to exploitatively emotional, have raised concerns for their strange content and potential scams. Research indicates that Facebook's algorithm may be amplifying these images, possibly for ad revenue or audience engagement. As AI-generated content becomes more accessible and prevalent, the line between authenticity and fiction in social media is increasingly blurring.

The Commodore 64, a classic personal computer from 1982, can now run generative AI to create 8x8 pixel sprites with a resolution upscaled to 64x64, inspired by developer Nick Bild's project. Despite taking 20 minutes to complete 90 iterations, the C64's ability to handle AI is a testament to its enduring legacy in computing. The AI-supported sprite generation process demonstrates creativity in repurposing vintage hardware for modern applications.

Fynk offers a comprehensive contract management solution with AI-powered analytics, allowing for automated drafting, electronic signing, and AI contract analysis for optimized contract workflows. The tool simplifies the creation, editing, and management of contracts, catering to various teams like Legal, Finance, HR, and Sales.

The 'podman-desktop-extension-ai-lab' project on GitHub offers an open source extension for Podman Desktop to work with LLMs (Large Language Models) locally, endorsed by its Apache-2.0 license. It provides a recipe catalog for AI use cases, open source models, and a playground for experimentation.

CustomerIQ is an AI-powered platform designed to aggregate and organize customer feedback, providing actionable insights to drive revenue, retention, and customer satisfaction. It features a self-organizing database, integration with various apps, AI-assisted content creation, enterprise-grade security, and tools for team alignment.

Wegic is an AI-powered web design and development platform that assists users in creating their ideal websites with the help of three intelligent assistants. It offers a simple and conversational interface for users to express their website needs, such as building a flower shop website, a personal resume, or a photographer's studio.

Stoic is a journal and mental health app designed to improve users' mental well-being through daily reflections, guided journals, and a variety of mental health tools, including meditation, breathing exercises, and habit tracking. The app has a 4.8-star rating based on 3K reviews.

SofaBrain is an AI-powered tool that allows for the easy redesigning, virtual staging, and rendering of rooms in seconds. With more than 1,376,850 redesigned spaces, it's trusted by over 253,438 homeowners and professionals.

Japanese researchers have released the Fugaku-LLM, a 13-billion-parameter large language model with enhanced Japanese capabilities. Built with Fugaku's supercomputing technology, it is trained on a rich mix of data including Japanese language for improved AI applications in research and business. It outperforms other models in Japan with a high benchmark score, notably in humanities and social sciences.

Sam Altman discusses the announcement of GPT-4o, highlighting the commitment of OpenAI to provide highly capable AI tools for free or affordably. Emphasizing the strategic shift from creating AI for direct use to enabling others to innovate with AI, and mentioning the revolutionary voice and video interface that feels like AI from the movies.

IBM has open-sourced its Large Language Models (LLMs) named Granite AI models, under the Apache 2.0 license, enabling researchers and commercial entities to innovate without restrictions. These models, trained on extensive coding datasets, are geared towards programming tasks, such as automating unit tests, writing documentation, and modernizing COBOL applications, offering significant benefits to developers and businesses alike.

Nectar is a platform designed to help teams collect, tag, and analyze customer feedback to make better business decisions. The insights generated are connected to dollar values, offering a clear picture of the impact on a business's bottom line.

The Pipecat GitHub repository is an open-source framework designed for creating voice and multimodal conversational AI applications, such as personal assistants, customer support bots, and storytelling toys for kids.

UAE's Technology Innovation Institute (TII) has released Falcon 2, a new AI model series that outperforms Meta's Llama 3. The series features the Falcon 2 11B model with 11 billion parameters and is capable of vision-to-language tasks. Falcon 2 11B matches Google's Gemma 7B, with both being open-source and independently verified by Hugging Face. The models offer multilingual support and efficient GPU optimization, later to explore a 'Mixture of Experts' for further enhancements.