Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
A new research approach introduces object identifiers in Large Language Models to improve understanding and answer questions about 3D scenes. This method, which focuses on identifying and relating objects in a scene, has shown promising results in making AI more adept at interpreting complex spatial relationships.
SwitchHead is a breakthrough in making AI models more efficient. It reduces the memory and computing needs of Transformers without losing performance.
Sponsored
New Relic is pioneering AI observability with New Relic AI Monitoring (AIM) [13]. AIM makes it easier to troubleshoot and optimize AI applications for performance, quality, cost, and responsible use of AI. With these capabilities, engineers have unprecedented visibility and insights across the AI stack. Learn more about New Relic AI
A work-in-progress repository to help get up to speed on the latest innovations. It is mostly geared towards beginners and includes resources for learning Python, deep learning, and language modeling.
The study introduces LLF-LUT, a technique that enhances the quality of HDR photos by focusing on important image details.
Capybara is a new base model for language and Obsidian is the multimodal extension. It has 3B parameters and can run on edge devices like phones.
Q* is unlikely to be the crucial breakthrough that leads to artificial general intelligence, but it might be an important step toward an AI with general reasoning abilities.
From the Cerebral Valley AI Summit in November, this podcast combines two interviews with tech titans, gathering their views on the future of AI and its impact on society.
Samsung unveiled its generative AI model, Samsung Gauss, which is being used for employee productivity enhancement and will be expanded to product applications. The model includes modules for language processing, coding assistance, and image editing. Alongside this, Samsung is focusing on data privacy and secure AI usage.
Review job applications 10x faster.
Open source AI grants from a16z have provided money and compute to groups working at the frontier of open-source AI. The second batch of grants went out to projects like Axolotl and Lmsys.
Google has launched Imagen 2, an enhanced AI image model available to approved Google Cloud Vertex AI users that offers new capabilities like text rendering, logos, and multilingual translations.
Cerebras makes AI training chips. It has released a framework for training huge models on its systems. The framework is simple and it doesn't require splitting up large models across multiple GPUs and nodes.
Gemini does better than many open models in a variety of vision tasks. It also seems very competitive with GPT-V on the same tasks.
This research introduces a method for compressing large language models using reduced order modeling, significantly reducing memory and time constraints without needing high-end hardware.
FIND introduces a versatile interface for AI models, enabling them to understand images and datasets better without altering the core model.
Sponsored
Building automated test coverage in-house takes years to scale. But manual QA, while more affordable, wastes time and causes bugs to sneak into production. Consider QA Wolf - they get web apps to 80% coverage in just 4 months [11]. They build and maintain your test suite in open-source Playwright and offer unlimited parallel test runs on their
Microsoft's newest model is available for research purposes on HuggingFace.
The HyperRouter method improves the efficiency of large language model training by dynamically adjusting router parameters.
This project delves into the application of the CLIP pre-training paradigm in medical imaging, discussing its adaptation, applications, and limitations. It aims to provide a comprehensive understanding of how CLIP can enhance medical image analysis.
Horizontal software, software designed to solve a user's pain regardless of the type of user that feels it, has defined the last 10 years of SaaS. Vertical SaaS, by contrast, is designed for a single or narrow user set, enabling the solution to be tailored to their needs. In this article, a Greylock partner forecasts that moving forward, users’ expectations around tailored services will govern how we build software.
The EU's AI Act, after a lengthy negotiation process and last-minute debates over facial recognition and foundational AI models, reached a provisional deal that could set a blueprint for global AI regulation. The Act, which categorizes AI systems based on risk, includes a two-tier system for general-purpose AI Systems and provisions for systems posing an "unacceptable" risk. It faces the challenge of adapting to rapidly evolving AI technologies.
Why did Google, arguably the pioneer of all things generative AI, feel the need to juice its marketing around the launch of Gemini? It boggles the mind and shows just how much is at stake here in the game of winner-takes-all in AI.
Sponsored
80% of enterprise data lives in difficult-to-use formats (HTML, PDF, CSV, PNG, PPTX…). Unstructured’s software makes this data available for vector databases and LLMs. Start with open-source or
AI infrastructure company Anyscale has introduced JSON mode and function calling capabilities to its Anyscale Endpoints service, significantly enhancing the usability of open models. Developers can now ensure LLM output consistency with JSON format and tailor them to specific schema requirements, as well as empower models to use APIs while choosing the right functions and arguments to call.
OpenAI has reached an agreement with Axel Springer, a Berlin-based owner of publications including Business Insider and Politico, to train its generative AI models on the publisher’s content and add recent Axel Springer-published articles to ChatGPT.
Simplify business travel with an AI assistant.
Sponsored
As the preferred synthetic data platform for enterprises [5], Gretel delivers tools to anonymize sensitive info, augment limited dataset, and create data from scratch using the company’s homegrown, tabular
→ Product demos including the new Model Playground and Tabular LLM
In this recorded demo [5], you’ll hear from Gretel Co-Founder and Chief Product Officer, Alex Watson, and the rest of the product leadership team. Watch now to learn:
Phi 2 is the next generation in the series of tiny language models from Azure trained primarily on synthetic data. The 2.7B parameter model matches the performance of 13B parameter models. The challenge here is detecting and mitigating “test set rephrasing”, but in either case, the model is quite powerful.
A group of senior leaders at OpenAI raising concerns to the board about Sam Altman's allegedly psychologically abusive behavior, which included creating chaos and pitting employees against each other, along with accusations of Altman's dishonesty in board communications, led to his brief firing. Altman's reinstatement followed overwhelming employee support and threats of mass resignations, but the incident has left uncertainties within the company and questions about its future direction.
Mistral now has an API with new models, an embedding, and instruction-tuned models. It has an unreleased medium-sized model that outperforms GPT-3.5.
Researchers have introduced Compound Text-Guided Prompt Tuning (TGP-T), a new method for Vision-Language Models like CLIP. TGP-T reduces GPU memory usage significantly while improving performance. It achieves this by optimizing prompts with text supervision, allowing more flexibility and efficiency in handling a variety of categories.
CorresNeRF introduces a novel approach to improve Neural Radiance Fields (NeRFs) performance, especially in scenarios with limited input views. By utilizing image correspondence priors for supervision, this method refines NeRF training, resulting in enhanced photometric and geometric accuracy.
DiAD leverages the power of diffusion models to detect anomalies. It uniquely combines a pixel-space autoencoder, a Semantic-Guided (SG) network, and a feature-space extractor to accurately detect and localize anomalies in multi-class settings.
AGAP is a new method that simplifies 3D editing. By using a 2D image, called a canonical image, AGAP allows users to easily edit 3D scenes without re-optimizing for each change.
Sherpa3D is a groundbreaking framework that improves 3D content creation from text prompts. It combines the strengths of 2D and 3D diffusion models, using coarse 3D knowledge to guide the creation process. This results in high-quality, diverse, and geometrically consistent 3D assets, overcoming the limitations of existing methods.
EasyVolcap is a PyTorch library for accelerating neural volumetric video research, particularly in areas of volumetric video capturing, reconstruction, and rendering.
Coordinating communication between language models and users is key given how fast model capabilities are advancing. The Interoperable Authorization Protocol (IAP) aims to address this by establishing secure, adaptable communication channels and a consent management system. This open-source approach encourages collaboration in the AI community to align AI operations with diverse human values and goals.
A committee of MIT leaders and scholars has released policy briefs providing a framework for U.S. AI governance, advocating the extension of existing regulatory and liability approaches to oversee AI effectively. Their proposal aims to enhance U.S. leadership in AI, minimize potential harms, and encourage beneficial AI applications, emphasizing the need for a clear purpose and intent definition in AI tools and suggesting possible new oversight structures.
In the 2022 TCV Engage Summit, the conversation highlighted that while AI represents a technological shift, it lacks a new distribution channel, which is crucial for creating significant consumer opportunities. Current AI innovations must rely on traditional distribution methods, favoring established companies or innovative startups, but the potential for new distribution platforms remains uncertain.
Google's "Project Ellmann" proposes using AI to analyze mobile data like photographs and searches, creating a detailed "bird's-eye" narrative of users' lives, although its integration into existing products like Google Photos remains uncertain.