Stay up to date with what is new in this rapidly changing AI world. Dive into updates on spotting fish from space and Uber’s LLM power-up.
- Open Source and In-House: How Uber Optimizes LLM Training
- Uber optimizes its large language model (LLM) training for applications like Uber Eats recommendations, chatbots, and code development by leveraging both open-source models (e.g., Meta’s Llama 2, Mistral AI’s Mixtral) and proprietary models from OpenAI and Google. Uber fine-tunes these models using domain-specific data through continuous pre-training and instruction tuning to improve performance at scale. The company's infrastructure, combining NVIDIA GPUs, Kubernetes, Ray, and tools like Hugging Face and DeepSpeed, allows rapid and efficient LLM training. Uber also enhances training throughput through memory optimizations and explores fine-tuning techniques like LoRA and QLoRA to improve scalability, enabling Uber to maintain state-of-the-art AI-driven services.
- This AI Paper Explores If Human Visual Perception can Help Computer Vision Models Outperform in Generalized Tasks
- Researchers from MIT and UC Berkeley investigated whether aligning computer vision models with human visual perception improves performance in generalized tasks. Their paper, "When Does Perceptual Alignment Benefit Vision Representations?" shows that models fine-tuned with human similarity judgments performed better in tasks like semantic segmentation and object counting. Using the synthetic NIGHTS dataset, they focused on spatial representations and local features, enhancing models with methods like Low-Rank Adaptation (LoRA). However, they found these models prone to overfitting and bias, underscoring the importance of diverse, high-quality human annotations for visual intelligence advancements.
- On the question of cheating and dishonesty in education in the age of AI
- In his article, Enrique Dans argues that concerns over AI-driven cheating in education overlook a deeper problem: the flawed metrics used to measure student performance. He contends that academic dishonesty is a symptom of a system overly focused on grades, which encourages students to prioritize high scores over genuine learning. He highlights Goodhart's Law, which states that when a metric becomes the goal, it loses its value. Rather than punishing students for using AI, Dans suggests that institutions should embrace it, teaching students how to use AI effectively to enhance their learning. He calls for a rethinking of education's ethics and methodologies to reflect modern technological realities.
- Combining next-token prediction and video diffusion in computer vision and robotics
- MIT researchers have developed a new method, "Diffusion Forcing," which combines next-token prediction and full-sequence diffusion models to improve AI capabilities in video generation and robotics. This technique helps robots handle tasks with noisy, corrupted data by predicting the next steps and generating high-quality video sequences. It merges the strengths of both models, allowing for flexible, long-horizon planning and precise decision-making. In experiments, Diffusion Forcing successfully guided a robotic arm to manipulate objects despite visual distractions and outperformed traditional models in video generation and AI planning tasks. The researchers aim to scale this method for broader AI applications.
- Remote sensing and computer vision for marine aquaculture
- Aquaculture, the cultivation of aquatic plants and animals, has grown rapidly since the 1990s, but sparse, self-reported, and aggregated production data limit the effective understanding and monitoring of the industry’s trends and potential risks. Building on a manual survey of aquaculture production from remote sensing imagery, we train a computer vision model to identify marine aquaculture cages from aerial and satellite imagery and generate a spatially explicit dataset of finfish production locations in the French Mediterranean from 2000 to 2021 including 4010 cages (average cage area, 69 square meters). We demonstrate the value of our method as an easily adaptable, cost-effective approach that can improve the speed and reliability of aquaculture surveys and enables downstream analyses relevant to researchers and regulators. We illustrate its use to compute independent estimates of production and develop a flexible framework to quantify uncertainty in these estimates. Overall, our study presents an efficient, scalable, and adaptable method for monitoring aquaculture production from remote sensing imagery.
- SAP doubles down on AI to transform enterprise operations
- SAP is leveraging AI to transform enterprise operations by integrating generative AI capabilities into its suite of business applications, including SAP S/4HANA Cloud. Their AI copilot, Joule, helps businesses automate workflows, enhance efficiency, and uncover intelligent insights across various domains such as finance, supply chain, and asset management. Joule allows users to interact with SAP systems through natural language, making complex processes more accessible. SAP is also using AI to streamline cloud migration and enhance developer productivity with tools like SAP Build. As AI-driven solutions evolve, SAP continues to prioritize data quality and governance, unlocking new insights for enterprise customers.
- FDA approves AI startup’s new software for echo assessments
- iCardio.ai
, an AI startup co-founded by cardiologist Aakriti Gupta, MD, and entrepreneur Joseph Sokol, has received FDA approval for its EchoMeasure software, designed to automate echocardiography assessments. This is iCardio.ai
's first FDA clearance, laying the foundation for future algorithms aimed at detecting structural heart diseases, such as valvular conditions. The company has also secured partnerships with key players like Viz.ai
, Abbott, and SARC MedIQ, and joined the Cedars-Sinai Accelerator program, enhancing its momentum in AI-driven cardiovascular diagnostics.
About Plainsight Technologies
Plainsight Technologies is the enterprise vision data company that makes cameras count. Through our pre-built Vision Intelligence Filters, containerized applications that solve business problems with accurate data from visual sources, we empower organizations to scale from concept to industry adoption while prioritizing privacy, security, and rapid innovation. Our mission is to "make your cameras count," extracting valuable insights from visual data to optimize processes in agriculture, marine biology, manufacturing, food service, and more. Headquartered in Kirkland, Washington, Plainsight Technologies operates as a distributed team, delivering cutting-edge solutions worldwide. To learn more, visit plainsight.ai
.
