MIT's Music Technology and Computation Graduate Program showcased innovative research projects, blending artful engineering with music performances. The event highlighted AI co-improvising agents, sound-art installations, EEG signals for musical tunes, and the program's interdisciplinary approach to shaping the future of expression.
"Beyond Data-Driven Aesthetics" by MIT alumnus Alexandros Haridis explores transforming computing into creative production in architecture. The exhibition translates algorithms and machine-learning systems into physical installations, questioning the intersection of computation and aesthetics.
Efficiently digitize scanned yearbook pages with Amazon Nova 2 Lite and Anthropic’s Claude Sonnet 4.6, producing 3,122 accurate name-to-face associations at a lower cost than single-model alternatives. The two-model pipeline uses native multimodal extraction and spatial reasoning to match names to faces based on page layout, offering a cost-effective solution for large-scale document digitization.
Automated claims processing pipeline using Amazon Bedrock reduces manual errors in healthcare forms, improving accuracy and efficiency. AI-powered services streamline document extraction and validation, creating FHIR resources in AWS HealthLake for faster processing.
PAR Technology Corporation supports 300+ restaurants with data-driven decisions. Their innovative LLM analytics system enforces strict row-level security, ensuring accurate results for each user.
David Autor, renowned labor economist, appointed head of MIT Department of Economics, aims to advance AI in research and teaching. Autor, a recipient of numerous prestigious awards, will lead the department through budget tightening and a shifting political landscape.
MIT researchers have developed a new approach, Masked IRL, to teach robots tasks with minimal human effort, using language models to clarify instructions and reduce demonstration data by nearly five times. The system enables robots to understand ambiguous prompts and safely complete chores in various settings, such as homes, offices, and factories.
Common ways to evaluate a machine learning regression model include MSE, accuracy, and R2. R2 is a key metric but can be negative, with no theoretical limit to how low it can go.
Stripe built a production-grade AI system on AWS that reduced review handling time by 26%. The system achieved over 96% helpfulness ratings with human oversight, scaling compliance operations without compromising quality.
Scientific American highlights the importance of early-career American scientists in driving innovation and prosperity. MIT faculty emphasize the need for continued public investment in curiosity-driven research to ensure future scientific advancements and societal impact.
Chaplin, an open-source solution, uses AI agents to provide self-service health event analytics for AWS users. Teams can ask questions in natural language and receive precise answers without depending on AWS Support.
Training large AI models on Amazon SageMaker AI with NVIDIA Blackwell GPUs removes constraints like limited batch sizes and sequence lengths, allowing for faster iteration cycles and reduced infrastructure costs. Blackwell's expanded memory and precision formats optimize training jobs, enabling longer sequence lengths and larger batch sizes for improved throughput and efficiency.
Researchers from MIT and Microsoft developed Murakkab, an intelligent system that automates the design and optimization of complex agentic workflows, reducing energy usage and costs while improving performance. This new method allows developers to describe tasks in plain language, letting the system choose the best models, tools, and hardware configurations dynamically based on user priorities.
Data teams often struggle with reconciling numbers, leading to slower decision-making and decreased confidence in analytics. Amazon Quick Sight datasets on Snowflake semantic views streamline data interpretation, reducing the risk of AI hallucinations and enabling natural-language queries for more efficient analysis.
Baidu's Unlimited OCR improves efficiency by replacing decoder attention with R-SWA, maintaining constant memory. It surpasses DeepSeek OCR, with a 93.23 score on OmniDocBench v1.5, using a 3B-parameter MoE model.