Google’s DeepMind and YouTube have introduced Lyria, an AI music generation model. Lyria, along with two experimental tools called Dream Track and Music AI, aims to revolutionize artistic expression. The collaboration allows creators to generate AI-generated soundtracks for YouTube Shorts. Concerns about the authenticity and sustainability of AI-generated compositions have also been raised. DeepMind and…
Researchers from Carnegie Mellon University, University of Pennsylvania, and Stanford University have proposed a new method called FACTORIZED CONTRASTIVE LEARNING (FACTORCL) to learn multimodal representations beyond multi-view redundancy. FACTORCL explicitly factorizes shared and unique information and maximizes lower bounds on mutual information to capture task-relevant information. It achieves state-of-the-art performance in various sentiment, emotion, and…
Researchers from KAIST AI and Samsung Research have introduced KTRL+F, a knowledge-augmented in-document search task that focuses on real-time identification of semantic targets within a document. The proposed Knowledge-Augmented Phrase Retrieval model balances speed and performance by incorporating external knowledge embedding in phrase embedding, enhancing contextual knowledge for accurate and comprehensive search and retrieval. KTRL+F…
This text discusses the 5 pillars of trustworthy large language model (LLM) testing: hallucination, bias, reasoning, generation quality, and model mechanics. It highlights the importance of understanding LLM behaviors and testing them in different scenarios. The text also emphasizes the ongoing challenge of developing a one-model-for-all LLM that excels in all 5 pillars. Overall, the…
Tencent AI Lab researchers have developed a solution called Chain-of-Noting (CON) to address reliability issues in retrieval-augmented language models (RALMs). CON enhances RALM performance by generating sequential reading notes for retrieved documents, allowing for better evaluation of relevance. The approach improves model understanding, resulting in more accurate and contextually relevant responses. CON outperforms standard RALMs,…
Stanford University researchers have developed a new algorithm called FlashFFTConv to optimize Fast Fourier Transform (FFT) convolutions for long sequences in machine learning. By employing a Monarch decomposition method, FlashFFTConv accelerates the FFT convolution, resulting in better efficiency, improved quality, and longer sequence models. The algorithm allows for kernel fusion at greater sequence lengths, reduces…
OpenAI, the popular AI company, experienced a tumultuous weekend with the firing of CEO Sam Altman. Following the announcement, several senior researchers also quit, prompting chaos within the organization. Altman and another top executive were subsequently hired by Microsoft, while OpenAI hired a former Twitch CEO as its CEO. Over 500 OpenAI employees threatened to…
Microsoft’s recent acquisition of top talent from OpenAI, including Sam Altman and Greg Brockman, suggests that the tech giant is positioning itself as a dominant force in the AI industry. With the possibility of 550 OpenAI employees joining Microsoft, the company aims to integrate OpenAI’s AGI agenda into its own operations, leveraging its corporate structure,…
More than 500 employees of OpenAI are threatening to resign en masse following the controversial removal of CEO Sam Altman and co-founder Greg Brockman from the company’s board. The employees, including chief scientist Ilya Sutskever, have signed an open letter challenging the board’s decisions, stating that they have jeopardized the company’s work and mission. The…
MIT researchers have developed StableRep, a system that uses synthetic images to train machine learning models, surpassing the results obtained from traditional “real-image” training methods. By using a strategy called “multi-positive contrastive learning,” StableRep considers multiple images generated from the same text prompt as positive pairs, enhancing the model’s understanding of high-level concepts. The approach…
OpenAI’s CEO Sam Altman and President Greg Brockman have been dismissed and removed from the board due to lack of transparency with the board. The decision has raised questions, particularly as it follows the release of new features. OpenAI’s unique board operates with a mission to create safe artificial general intelligence. Microsoft has hired Altman…
Microsoft has hired former OpenAI CEO Sam Altman and co-founder Greg Brockman to lead a new advanced AI research team. This move comes after OpenAI’s board lost confidence in Altman’s leadership. Microsoft CEO Satya Nadella expressed enthusiasm for the team’s success and highlighted the company’s experience integrating independent identities. OpenAI appointed Emmett Shear as its…
France, Germany, and Italy have reached a stricter agreement on regulating AI than the proposed EU AI Act. The focus is on regulating the application of AI rather than the technology itself. The agreement calls for AI companies to provide a “model card” for their models. In contrast, the UK has chosen not to impose…
AI article: Enhance your user experience skills with up to 7 comprehensive training courses at the upcoming conference from February 10-16, 2024. This event is designed to equip UX professionals with long-lasting skills necessary for successful design. View the full schedule and pricing details.
Research by the UK Treasury’s Productivity Programme has identified opportunities to reduce administrative work, harness AI, and improve public services. The Home Office will publish recommendations on utilizing AI for routine tasks, potentially saving teaching and police hours. AI is already being used in education and healthcare, with positive outcomes such as improved stroke treatment.…
Google Research, Google DeepMind, and the University of Waterloo have introduced SWIM-IR, a synthetic retrieval training dataset for multilingual retrieval models. Using the SAP method, the dataset allows for fine-tuning of dense retrieval models without human supervision. SWIM-X models trained on SWIM-IR show competitive performance on various benchmarks. The research highlights the potential of synthetic…
AI’s effectiveness heavily relies on data availability for training purposes. However, a study by University of Toronto Engineering researchers suggests that deep learning models may not always require a lot of training data. The researchers found that smaller subsets of data can be used to train models without compromising accuracy. The study emphasizes the significance…
Microsoft has officially announced its in-house designed chips, the Azure Maia 100 AI accelerator and Azure Cobalt CPU, at the Ignite conference. These chips demonstrate Microsoft’s commitment to innovation and self-sufficiency across hardware and software. They are set to power Azure’s AI workloads and will be integrated into specially designed server motherboards and racks. Microsoft…
GOAT is a universal navigation system developed by researchers from various universities and organizations. It operates autonomously in home and warehouse environments, using category labels, target images, and language descriptions to interpret goals. GOAT creates a 3D semantic voxel map for accurate object detection and memory storage, and it has demonstrated superior performance in reaching…
Researchers from MIT and IAIFI have developed a framework called Feature Fields for Robotic Manipulation (F3RM), which addresses the challenge of enabling robots to manipulate objects in cluttered environments. F3RM leverages distilled feature fields to combine 3D geometry with semantic information from 2D models, bridging the gap between 2D image features and 3D geometry. The…