Berkeley AIWhat exactly does word2vec learn?
A closed-form theory revealing that word2vec learns word embeddings through sequential low-rank approximations equivalent to PCA on a matrix derived from corpus co-occurrence statistics.
OpenAISupporting nonprofit and community innovation
OpenAI launches a $50M People-First AI Fund to support U.S.-based nonprofits applying innovative AI solutions for community impact and public good.
OpenAIIntroducing gpt-realtime and Realtime API updates
The Realtime API is now generally available featuring the advanced gpt-realtime speech-to-speech model with enhanced natural language understanding, low latency, multimodal inputs, SIP calling, and improved function calling for building production-ready voice agents.
MIT AIMIT researchers develop AI tool to improve flu vaccine strain selection
MIT researchers have developed VaxSeer, an AI-driven tool leveraging protein language models and deep learning to accurately predict dominant influenza strains and optimize flu vaccine strain selection months in advance.
Modular AIModular: Matrix Multiplication on Blackwell: Part 1 - Introduction
An introductory exploration of writing and optimizing matrix multiplication GPU kernels on NVIDIA's Blackwell architecture using Mojo, highlighting architectural advancements and programming paradigms critical for high-performance LLM workloads.
OpenAIOpenAI and Anthropic share findings from a joint safety evaluation
OpenAI and Anthropic conducted a collaborative safety evaluation of their AI models, comparing performance across instruction hierarchy, jailbreak resistance, hallucination rates, and scheming behavior to advance transparency and improve alignment techniques.
OpenAICollective alignment: public input on our Model Spec
OpenAI presents a collective alignment approach using global public input to refine and update its Model Spec, ensuring AI behavior reflects diverse human values and priorities through a combination of automated and human-driven processes.
Jane StreetWhat the interns have wrought, 2025 edition
Highlighting innovative intern projects at Jane Street in 2025, covering optimized SQL evaluation with indexing, advanced OCaml Torch bindings with reference counting, and efficient shared-memory ref-counted object management for trading systems.
DropboxHack Week 2025: How these engineers liquid-cooled a GPU server
Engineers at Dropbox designed and tested a custom liquid cooling system for GPU servers during Hack Week 2025, achieving significant thermal performance improvements to support next-generation AI workloads efficiently.
InstacartSimplifying Large-Scale LLM Processing across Instacart with Maple
Maple is a scalable, cost-effective batch processing service developed by Instacart to streamline large-scale LLM workflows, automate prompt management, and enable reliable, fault-tolerant AI operations across teams.
OpenAIHelping people when they need it most
ChatGPT is continuously improving its empathetic response and safety measures to support users in mental and emotional distress, including enhanced safeguards, expert collaboration, and expanded crisis intervention features.
MIT AISimpler models can outperform deep learning at climate prediction
MIT research reveals that simpler physics-based models can outperform complex deep-learning approaches in regional climate predictions, highlighting the importance of problem-specific modeling and robust benchmarking methods.