MIT AILLMs factor in unrelated information when recommending medical treatments
LLMs show sensitivity to nonclinical text variations like typos and gender markers, leading to inconsistent and potentially harmful medical treatment recommendations, especially for female patients.
SalesforceHow Salesforce Engineering Operationalized AI Productivity at Scale
Salesforce Engineering successfully operationalized AI productivity at scale by integrating AI tools like Cursor, CodeGenie, GitHub Copilot, and Gemini into daily workflows, achieving over 90% adoption across multiple teams and driving measurable improvements in development speed, testing, and knowledge sharing.
EleutherAIResearch Update: Applications of Local Volume Measurement
An evaluation of local volume measurement using the Tyche library reveals its limited effectiveness for detecting neural network misalignment and anomalous datapoints compared to activation perturbation methods like POSER.
MIT AIResearchers present bold ideas for AI at MIT Generative AI Impact Consortium kickoff event
MIT researchers presented innovative generative AI projects at the MGAIC kickoff event, showcasing cross-disciplinary collaboration and impactful applications across education, urban sustainability, and the arts.
SalesforceInside Data Cloud’s Open Lakehouse: 4M Tables and 50PB – Enabled by Apache Iceberg
Salesforce’s Data Cloud Lakehouse leverages Apache Iceberg to enable scalable, real-time, AI-ready data architecture supporting 4 million tables and 50PB of data with advanced features like incremental processing, event-driven optimization, and robust governance.
Lambda LabsApriel 5B: ServiceNow’s Enterprise AI Trained and Deployed on Lambda
Apriel 5B is a compact 4.8B parameter transformer LLM by ServiceNow, optimized for efficient enterprise AI deployment on Lambda's infrastructure, delivering high throughput, low latency, and versatile NLP and code generation capabilities.
Fly.ioPhoenix.new – The Remote AI Runtime for Phoenix
Phoenix.new is a fully online, root-access AI coding agent tailored for Elixir and Phoenix that builds, tests, and deploys real-time collaborative applications directly from a browser using an ephemeral virtual machine and headless browser integration.
Modular AIModular: How is Modular Democratizing AI Compute? (Democratizing AI Compute, Part 11)
Modular is revolutionizing AI compute by building a unified, high-performance, and scalable software stack that democratizes access across diverse hardware and developer communities, enabling seamless portability, programmability, and orchestration for the GenAI era.
OpenAIPreparing for future AI risks in biology
Exploring strategies to anticipate and mitigate potential AI-driven risks in the field of biology.
OpenAIToward understanding and preventing misalignment generalization
Exploring strategies to understand and prevent misalignment in AI generalization.
Modular AIModular: Modular 25.4: One Container, AMD and NVIDIA GPUs, No Lock-In
Modular 25.4 introduces seamless support for AMD GPUs alongside NVIDIA, enhanced AI model compatibility, open-source GPU programming advancements, and improved developer experiences, enabling flexible, high-performance, and vendor-neutral AI deployments.
MIT AIA sounding board for strengthening the student experience
MIT's Undergraduate Advisory Group in the Schwarzman College of Computing empowers students to enhance the educational experience through direct dialogue with leadership and innovative projects like the NerdXing course visualization tool.