Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
The LLaVa model is an open-source model that combines both language and vision. This new version allows the instruction-tuned model to use tools for image editing, generation, and more.
AI town is an amazing experiment where hundreds of agents live out their daily lives as a prompt state in a language model. The RWKV model is a linear language model that requires fewer resources than standard Transformers. This repo uses this cheaper model to run AI town on your local machine.
Personalized AI content for brands.
Poe has updated its AI platform to support creator monetization through subscription revenue shares and per-message fees, aiming to empower bot developers and small AI companies. Monetization is live in the US, with global expansion on the horizon.
Samsung has mostly been quiet in the AI race, but it recently showed off a live AI translation feature for phone calls. The feature is expected to launch in 2024.
Adept is building a useful agent system for enterprises. It is now opening up small mini demos for people to interact with like delegating knowledge working tasks, using new tools, or working with unstructured data.
OpenAI introduces customizable ChatGPT (GPTs), enabling users to tailor the AI for specific tasks without coding. The GPT Store, launching soon, will allow the sharing and monetizing of these custom GPTs, with a focus on privacy, safety, and expanding utility in the real world. Enterprises can deploy internal-only GPTs, enhancing business operations.
Google ran a MaxText + Jax training job on an unbelievable amount of chips concurrently while maintaining 66+ utilization, creating a strong contender for the best possible MFU. The project was mostly built on proprietary technology, but it showcases some interesting tidbits for other large-scale distributed computing groups.
MLPerf is a standard machine learning performance benchmark that accurately shows how platforms and models perform in real-world settings. MLPerf Training v3.1 just landed and it showcases amazing H100 performance in absolute terms, but when comparing wall clock performance per dollar spent, Intel still holds its own.
This study introduces a better way to clean up the data used to train LLMs, ensuring they are tested fairly and work more reliably.
Sponsored
QA Wolf [11] gets web apps to 80% automated test coverage in just 4 months. And working with them is more cost-effective than growing your QA team. After giving QA Wolf a product tour and access to an environment, their QA engineers will build you a comprehensive end-to-end test suite in Playwright. There's no vendor lock-in [11] and it's easy to export your tests. QA Wolf also provides 24-hour maintenance and triage. You receive human-verified bug reports [11] — no false positives. Plus, you get unlimited parallel test runs on their infra at no additional cost. Schedule a demo [11] to learn more.
LongQLoRA is a new tool that lets LLMs understand longer sequences.
Researchers developed a smarter method for self-driving cars to recognize roads by using lidar technology, which requires less manual work but still keeps the system accurate.
Researchers discovered that large language models can be powerful tools for visual tasks even though they were designed for text. The researchers achieved improved results in various visual tasks by using parts of these text-trained models to process images and videos directly.
Nvidia became the AI kingmaker in 2023. Thousands of companies and countries begged to purchase its GPUs. Will the insatiable demand persist? This piece explores the Nvidia run-up, the resulting shortage, and the future of GPU demand.
OpenAI wants to build partnerships for open and private dataset creation for AI training. This could signal it is running out of tokens for its next model training runs.
Current software is still limited in its ability to understand and assist with diverse tasks. Bill Gates predicts that AI agents will revolutionize software interaction, offering personalized assistance across various tasks without needing specific apps, marking the biggest shift in computing since the move from command-line to graphical interfaces.
Browse the web with GPT-4V and Vimium.
Supercharge your growth with AI-powered cold emails.
Lindy introduces the first platform for building teams of AI employees to perform any task without requiring coding skills. Users can create these 'Lindies' in 30 seconds and set them up in AI teams that can support advanced workflows. Lindy enlarges on-scale productivity with these 24/7 AI employees.
Sponsored
Join a new webinar series diving into the future of synthetic data and the Gretel [5] platform starting this Tuesday, November 14, 2023, 9am PST/12pm EST [5].
* Previews and live demos, including the new Model Playground and
* Operationalizing synthetic data across the enterprise with Gretel workflows * A live Q&A session [5] to answer your most pressing questions about generative AI and synthetic data
Google has expanded its partnership with AI safety startup Anthropic, leveraging Google's Cloud TPU v5e chips for AI, enhancing AI security and safety standards, and utilizing Google's technological infrastructure for large-scale data handling and analysis.
The latest Cohere model is an upgrade for search applications that engage with real-world, noisy data and for retrieval-augmentation generation (RAG) systems. Its multilingual models now support 100+ languages and can be used to search within a language (e.g., search with a French query on French documents) and across languages (e.g., search with a Japanese query on Finnish documents).
Microsoft's GitHub Copilot software saw a 40% rise in paid users in Q3, affirming CEO Satya Nadella's focus on AI integration across products. GitHub’s AI offering is fast gaining traction with "over 1 million paid copilot users in more than 37,000 organizations". These AI advancements contributed to a positive impact on Microsoft's Azure revenue.
Using a clever interleaving of left and right channels during fine tuning, these state of the art text to music models are brought from mono to stereo which doesn't improve benchmark scores but does improve the overall experience.
Researchers have developed JaSPICE, a metric for evaluating Japanese image captions, addressing the gap where tools like SPICE did not exist for languages other than English.
The RobustMat approach aims to improve how autonomous vehicles match landmarks under diverse conditions by using spatial information and neural differential equations.
Sponsored
Ring 一 Acquired by Amazon for $1.2B. Nest 一 Acquired by Google for $3.2B.
is the tech firm poised to dominate the Smart Shades market and you can still invest at $1.25/share [12].
Roboflow is a computer vision toolkit that makes building applications easier with composable pieces of code. Now that GPT-4 Vision is in the OpenAI API, the Roboflow team has put together some demos of using it in conjunction with segment anything and DINO.
Punica enables running multiple LoRA fine-tuned models at the cost of running one.
The new I2VGen-XL model tackles the challenges of video synthesis, such as maintaining semantic accuracy and clear, continuous imagery, by using a two-step process that separates semantic coherence from video quality enhancement. The first stage focuses on preserving content from static images, while the second stage refines details and resolution.
The article analyzes 2 million Hacker News comments to compare the public sentiment towards AI and crypto. The study reveals trends in topic popularity and sentiment over time, showing AI's consistent prominence and a complex, fluctuating public opinion on both AI and crypto.
Fear is driving regulatory approaches towards AI, with the US and UK initiating measures due to concerns about the long-term risks of AI. While some regulation is necessary, fear-driven decisions may lead to hasty and potentially detrimental regulations. Actions should not be dictated solely by fear.
MyCompanions.AI aims to blend human influencers with AI, providing users an engaging and personalized experience. The platform plans to address predicted dominance of AI influencers and the flaws in existing platforms like Patreon. Promising features include custom role plays, strong contextual and conversational memory, and multimodal automation.
Sponsored
STYLO [19] is an AI assistant that lives inside every Zendesk ticket to solve your 10-MINUTE TICKETS IN 10 SECONDS. Customers get (correct) answers fast; agents do less gruntwork 🙌
AI startup Hugging Face's H4 team, formed in January, is focused on developing open source large language models and alignment techniques for AI-powered chatbots inspired by ChatGPT.
A compilation of reasons from the biggest AI companies on why they shouldn’t have to pay for copyrighted content.
CopyRemix uses AI to turn your user testimonials into full case study content.
Sponsored
Need more compute power (and choices) to fine-tune your generative AI? Intel® Developer Cloud [5] accelerates your AI development with Intel®-optimized software on the latest Intel® Xeon® processors and GPU compute.
OpenAI CEO Sam Altman showcased new features and improvements to the company’s AI models at its first developer conference, emphasizing a future where AI integration will play a central role in consumer technology. The event highlighted OpenAI's shift towards a product-centric approach, with new tools being made immediately available, and hinted at potential forays into hardware to reduce friction in AI accessibility.
Nvidia revealed the testing of a large language model named ChipNeMo to enhance chip designers' productivity during a keynote at the IEEE/ACM International Conference on Computer-Aided Design. Although not yet fully proven, ChipNeMo aids in scripting for design tools, summarizing bug reports, and serving as an experienced chatbot for designers.
YouTube is testing new generative AI features as part of its premium subscription service. These features include a conversational tool for asking questions about content and getting recommendations and another for summarizing comment topics.
PixArt is a new text to image model that uses T5 text encodings, cross attention, and a diffusion transformer to great results at a fraction of the compute cost of comparable models.
The DARE method streamlines the enhancement of language models like BERT, allowing the integration of new functions into a unified model, increasing efficiency on various linguistic tasks.
Cross-View Training (SCT) elevates small language models, enabling them to produce sentence embeddings previously only achievable by larger models, optimizing both performance and computational efficiency.