Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
HelixML was able to better fine-tune Mistral-7B by implementing a suite of qapair prompts that extracted content from all sorts of different perspectives and generating a content-addressed hash for each document.
Google's newly released Gemini Advanced, a GPT-4 class AI model, showcases capabilities equivalent to OpenAI's GPT-4. It shows strengths in explanations and integrating images and search.
This point of view, from an investor, helps to bring a little context to the layers of value within AI right now by breaking them down into the infrastructure layer, like cloud providers and chip makers, modeling and core, like OpenAI and Anthropic, and AI-enhanced products, like all of you that use AI to improve your products.
Shortwave released a mobile email assistant for iOS & Android that can create instant summaries during its AI Launch Week last week.
A benchmark dataset that contains facts about various anime characters and quotes to assess language model performance.
Singapore-based Brilliant Labs has introduced Frame, lightweight AR glasses featuring a multimodal AI assistant, Noa, that performs visual processing, image generation, and more with integrated AI models like GPT-4 and Stable Diffusion.
Microsoft announced a major update to its Copilot AI, including a new design, AI image creation, editing functionalities, and a new AI model named Deucalion.
A new open-source language model named "Smaug-72B," developed by Abacus AI and derived from "Qwen-72B" by Alibaba Group's Qwen team, now leads Hugging Face's natural language processing leaderboard. Outperforming established models like GPT-3.5 and Mistral Medium in various benchmarks, Smaug-72B signifies a milestone for open-source AI by scoring an average of over 80 across major evaluations, hinting at the potential to rival proprietary AI technologies.
The European Union's AI Act, aimed at regulating artificial intelligence applications based on risk, has passed a crucial vote by Member State representatives, who have confirmed the final text of the draft law.
These text-to-speech models, trained by Stability AI, can be guided by precise natural language instructions. As there is no large dataset with proper textual descriptions of audio for generation, its creators synthetically annotated a large corpus of speech for training. This is another example of a broader trend of annotation, up-captioning, and training for generative modeling.
The MusicLM team at Google used 300k pieces of feedback, along with other reward signals, to run an RL process on their music generation models. They found it outperforms the base model in human preference studies, but it is unclear which RL method yields the highest fidelity output.
This paper revisits the classical Gaussian Discriminant Analysis (GDA) algorithm to improve the performance of CLIP in image classification tasks without additional training or resources.
Sponsored
They build and maintain your automated test suite, plus provide unlimited parallel test runs. (24-hour maintenance included).
Object detection is the process of identifying objects and their bounding boxes. This can usually only be done for a fixed set of objects chosen before training. This work introduces a real-time method that can do Open Vocabulary object detection, which means it can detect bounding boxes for any run-time-specified combination of objects.
Google proposed a novel prompting technique that allows language models to use a set of reasoning primitives to discover a larger framework for problem-specific reasoning. This means the models can select different modules and combine them to better solve complex problems. This repository is an unofficial implementation of these ideas.
MobileVLM V2 is a range of advanced vision-language models for mobile devices that showcase notable performance improvements through innovative architecture.
Recent shifts in AI research focus on efficiency and depth over mere accuracy and breadth. NVIDIA's H100 sales and AI's growing energy demand highlight the industry's scale. Investments demand profitability, shifting research towards smaller, more efficient models like Phi 2 and emphasizing sustainable economics from model architecture to deployment. Innovations in training, fine-tuning, and design promise to improve AI's energy and computational efficiency. On-device capabilities reflect a broader trend towards more sustainable and practical AI applications.
A new study reveals a weakness in multi-modal large language models (MLLMs) like GPT-4V: they struggle with specific types of image-text inputs, leading to errors. CorrelationQA is a benchmark designed to evaluate MLLMs' performance in scenarios where images may mislead or contradict text.
The Rabbit R1 is a bright orange walkie-talkie for AI. It features one big button that users push to talk, a chunky scroll wheel to browse the screen, and a rotating camera with its own privacy door. At $199, the device can be programmed to control apps and websites. This article tells the story of how Rabbit and Teenage Engineering collaborated to create what may be the most successful AI hardware launch to date.
OpenAI has established a Child Safety team to explore ways to prevent misuse of its AI tools by children.
Create your own free ChatGPT widget for your website.
A Chrome extension that allows you to use a local language model to filter social media posts based on criteria you choose.
OpenAI is implementing watermarks in images generated by DALL-E 3, following standards from the Coalition for Content Provenance and Authenticity (C2PA), to include invisible metadata and a visible CR symbol for provenance verification.
Microsoft's AI-driven Copilot for Sales and Service is now generally available. It integrates with CRM systems like Salesforce to enhance productivity for sales and service professionals. The tools automate repetitive tasks and provide insights directly within Microsoft 365 apps, aiming to improve customer interactions and streamline business workflows. Early adopters like Avanade report significant time savings and better customer engagement using these AI capabilities.
The removal of large-scale datasets such as LAION, along with valid copyright concerns, makes training large-scale image models challenging. This work shows that you can train a powerful CLIP model with 30 million synthetically generated images.
This research explores how to make powerful language models more suitable for mobile devices by focusing on tiny models with fewer parameters.
Addressing the critical issue of hallucinations in Multimodal Large Language Models (MLLMs), researchers have developed MHaluBench, a new benchmark for evaluating hallucination detection methods.
Bria AI has open sourced (non-commercial) its background removal model and tools, including training data mixture and performance metrics. It is a significant improvement over alternative open models.
InteractiveVideo introduces a novel way of making videos where users can dynamically interact with the creation process. Unlike traditional methods, this user-friendly framework allows for real-time modifications using text, images, painting, and even drag-and-drop.
DeepSeekMath 7B is a language model specifically designed to tackle complex mathematical problems. It's been trained on a massive amount of math-focused content.
The new Meta Voice model is a small and powerful text-to-speech model that supports generation and voice cloning.
Advanced AIs like GPT-5 could have significant impacts across markets, real estate, labor, energy, national economies, geopolitics, and the economy's inflationary dynamics, leading to a reevaluation of value accrual, global power shifts, and changes in employment and economic structures.
As AI tooling and business use cases become more sophisticated, we’ll start to see more and more real-world applications of the tech. One lens to look at the opportunity here is how AI will make processes cheaper and more efficient whilst growing margins substantially. This might unlock a whole new way of doing things that previously didn’t make economic sense because margins were too thin. This article explores a few of these examples.
These models have been specifically tuned for Basque on a corpus of around 4B tokens, starting with Llama 2.
Create personalized AI workflows without needing to know how to code.
An open-source framework for augmenting humans using AI.
From 0.5B to 72B parameters. These are an impressive suite of multilingual models. Interestingly, the smallest model is the first serious sub 1B parameter language model.
In May 2023, Colin Megill met with OpenAI, including co-founder Wojciech Zaremba, to discuss integrating his public deliberation platform, Polis, which used AI to enhance democratic engagement. The collaboration aimed to align AI with human values through broad public input. It sparked OpenAI's "Democratic Inputs to AI" initiative to explore AI governance with a $1 million grant program.
Video synthesis is notoriously hard to control. This work proposes using boxes and arrows to guide the generation over time, which improves human preference judgment but still doesn’t give perfect guidance to the user.
Kahneman-Tversky Optimization (KTO) is a new approach for training AI models to align better with human thinking. By utilizing concepts from Kahneman & Tversky's prospect theory, KTO focuses on maximizing utility rather than just preference likelihood.
This paper sheds light on why large vision-language models (LVLMs) sometimes incorrectly describe images, a phenomenon known as multimodal hallucination. Semantic shift bias, particularly around paragraph breaks, is a key factor.
Sponsored
Fly to Hawaii for $149, Paris for $289, or anywhere else. Exclusive for TLDR readers: 93% off Dollar Flight Club’s lifetime membership until midnight — just $129 (normally $1,690) with a
A group is working to build an open-source version of the Rabbit hardware including the language modeling portion.
NaturalSQL by ChatDB is a series of models with state-of-the-art performance on Text-to-SQL instructions.
CapHuman is a framework that can create various images of a person using just one reference photo. It cleverly maintains the person's identity while changing head positions and expressions.