Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
The enhanced version of Bard promises to deliver more complex and better responses compared to the current version, which runs on Gemini Pro and competes with GPT 3.5.
OpenAI has announced its first partnership with a higher education institution, Arizona State University, which will gain full access to ChatGPT Enterprise for coursework, tutoring, research, and more, including building a personalized AI tutor and creative AI avatars for students.
Mark Zuckerberg, Meta's CEO, is entering the race to develop artificial general intelligence (AGI), aiming to enhance Meta's apps and user experience. He emphasizes the need for AI talent and computing power and is considering an open-source approach to AI development, contrasting with the more closed methods of other companies.
This massive document outlines essentially everything one would need to know about the state of the vector retrieval field. It is a valuable resource that will take some time to dig through.
Apple has released a great paper on pretraining image models with an autoregressive loss. It measures scaling behavior and finds that like in language, these large unsupervised image models have very predictable behavior.
This paper improves how images are edited using natural language instructions by tackling challenges like unclear instructions and identifying specific parts of the image to edit.
Sponsored
Developing against OpenAI APIs can be finicky and expensive. MockGPT by WireMock [10] is a readymade mock API for ChatGPT, GPT-3.5, etc. Use it to write and test your code without fees and frustrating wait times — and switch one line of code to go live with the real API.
Compose and Conquer introduces a conditional diffusion model that excels in placing objects in three-dimensional space within generated images.
Vlogger is an advanced AI system capable of generating minute-long video blogs (vlogs) from user descriptions. Vlogger operates in four stages - Script, Actor, ShowMaker, and Voicer, each representing a specific aspect of vlog creation.
This project evaluates the generalization capabilities of deep learning models for instance segmentation to novel objects.
A quantization-aware attention alternative that speeds up inference substantially over some state-of-the-art implementations. Its creators also claim 700 TFLOPs on MI-300X. The details are sparse, but it would be quite an achievement if so.
Dolphin is widely considered to be one of the strongest uncensored open models available. It is mainly built on Mixtral. This 120B model was made by merging two smaller Dolphin models together. It has been found, with Goliath 120B, that this merging process can improve overall model performance.
TACO is a new benchmark for assessing a system’s ability to generate code. It is substantially larger than existing datasets and contains more challenging problems. On easy problems, GPT-4 gets 30% correct, while it only achieves 2% on the hardest subset of questions.
Microsoft has launched Reading Coach, an AI-powered tool designed to provide personalized reading practice available at no cost for anyone with a Microsoft account.
A handy trick to quickly lower your AI cost.
Twitter’s former CTO and interim CEO Parag Agrawal has launched his own AI startups. These startups have received funding from some big-name VCs like Khosla Ventures.
Sponsored
Harness the power of Ray, the world's leading open-source distributed computing framework, tailored for LLMs, conversational AI, computer vision, and enterprise AI applications on Anyscale. Enable your developers to experiment quickly with different open source LLMs. Build RAG for enhanced performance, easily incorporating proprietary data and metadata for specific use cases.
AlphaGeometry solved 25 Olympiad problems in the allotted time limit, improving on the previous state-of-the-art, which solved 10. Importantly, it was trained on 100 million synthetic problems, which eliminates an existing data bottleneck. DeepMind has open-sourced the code and weights.
A Georgia judge recently ruled against OpenAI's motion to dismiss a defamation lawsuit filed by radio host Mark Walters, who alleges that ChatGPT's “hallucinations” produced false and potentially career-damaging claims about him. This ruling allows for the exploration and decision on the merits of the case, challenging OpenAI's liability for ChatGPT's outputs and raising questions about the responsibility for AI-generated content.
LMSYS has released two new innovations for language model inference. The first is a backend change which improves overall tokens per second performance. The second is an embedded domain-specific language for complex prompting techniques which allows for prompting parallelism.
This paper explores the challenge of developing Vision Foundation Models (VFMs) specifically for autonomous driving. By analyzing over 250 research papers, it provides insights into data preparation, pre-training, and task adaptation in AI, highlighting cutting-edge techniques like NeRF and 3D Gaussian Splatting.
DoraemonGPT, a new AI system based on large language models, takes a step forward in understanding dynamic real-world scenarios by focusing on video tasks. It converts videos into a symbolic memory for efficient spatial-temporal querying and incorporates specialized tools and a novel planner for complex task-solving.
Sponsored
Kolena's latest research [13] on hallucination detection for Large Language Models (LLMs) has wide-ranging implications for evaluating model performance. Get hands-on perspectives [13] on LLM testing from Kolena’s ML engineers, and see how different models compare in the
Nathan Lambert recently released Interconnect Tools, which allows you to automate his multimodal blogging pipeline using text-to-speech, image synthesis, and his entire workflow. It allows creators to focus on writing while the tools format things into easily consumable media for their audience.
This project proposes Contrastive Preference Optimization (CPO) to improve the performance of moderate-sized language models in machine translation. This approach, applied to the 13B parameter ALMA model, addresses the limitations of supervised fine-tuning and results in improvements.
AlphaCodium introduces a novel approach to enhance code generation by LLMs. This multi-stage, test-based iterative process significantly improves the accuracy of models like GPT-4 in solving complex programming problems, as demonstrated on the CodeContests dataset.
This article explores the challenges of evaluating LLMs, comparing them to human employee assessments. It discusses the difficulty in measuring the intelligence and practicality of LLMs, highlighting the limitations of current evaluation methods and the need for more effective approaches.
This study introduces UV-SAM, an adaptation of the vision foundation model and the Segment Anything Model that can accurately identify the boundaries of urban villages using satellite images. Offering an efficient alternative to traditional field surveys, UV-SAM combines different image representations for precise detection.
Recent studies suggest methods to control AI honesty, such as identifying and manipulating honesty-related vectors within the AI or using unrelated questions to detect lying tendencies based on the AI's response consistency.
Sponsored
The world’s best AI companies recruit with Gem. Hire top talent quickly with an all-in-one recruiting platform that combines sourcing,
A privacy-first coding assistant.
Samsung's Galaxy S24 and S24 Plus, set to be released on January 31st, feature enhanced AI capabilities.
Market analysis done in minutes.
Sakana.ai is building a world-class research lab in Tokyo to build smaller and more efficient foundation models. The company was founded by ex-Google researchers David Ha and Llion Jones, who are responsible for breakthroughs such as Transformers, World Models, and LoRA. It has raised a 30m seed round from Brandon Reeves at Lux Capital and Jon Chu at Khosla Ventures to pioneer this effort and put Tokyo on the AI map.
A strong 3B model that outperforms the 7B CodeLlama model and is small enough to run locally on MacBooks.
OpenAI is forming a new Collective Alignment team to integrate public input into its AI models.
This blog explores the current dominant paradigm for taking a pretrained language model and adding multimodal capabilities. Namely, aligning the embeddings between an encoder and the text model.
Hard problems have been thought to require hard data, but that data is hard to label and noisy. This work shows that fine-tuning models on related, but easy, data dramatically improves their generation capabilities on hard problems. Another data point to support fine-tuning as eliciting knowledge instead of adding knowledge.
Researchers have developed a novel approach called Mutual Distillation Learning For Person Re-identification (MDPR) that enhances person re-identification by combining two different techniques.
Sponsored
Microsoft is marketing that its Direct Lake architecture on Fabric is the ideal way to run direct queries on cloud storage in Power BI. Does the data support their claim? This webinar [12] will dive deep into AtScale’s TPC-DS benchmark report [12] that tests the scalability, cost effectiveness and manageability of Microsoft Fabric Architecture.
One of the most viral uses of generative image models is personal stylized image generation. Historically, this required training a LoRA or using DreamBooth. Now, with the use of ID embeddings, you can dramatically improve quality while reducing computational cost, with just a single image.
Following the recent trend of repurposing diffusion models for various other vision tasks, this work improves the stability of training for super resolution and upscaling.
FILCO is a new technique designed to improve AI systems, especially in areas like answering questions and checking facts, by giving them better context. It works by choosing the most helpful information and training models to filter out less relevant details.