AI Daily — May 15, 2026
2026-05-15
anthropic.com
Anthropic forms $200 million partnership with the Gates Foundation
Anthropic has announced a $200 million partnership with the Bill & Melinda Gates Foundation, marking a significant institutional commitment to applying frontier AI to global health and development challenges. This is one of the larger philanthropic AI partnerships announced by a major lab and signals growing interest in directing frontier model capabilities toward high-stakes scientific and humanitarian domains.
openai.com
Work with Codex from anywhere
OpenAI has extended Codex to the ChatGPT mobile app, enabling engineers to monitor, steer, and approve long-running agentic coding tasks from any device. The update positions Codex as an asynchronous coding agent rather than a synchronous assistant, allowing tasks to run remotely while users interact via mobile.
openai.com
Helping ChatGPT better recognize context in sensitive conversations
OpenAI has shipped safety updates to ChatGPT that improve longitudinal context tracking in sensitive conversations, allowing the model to detect escalating risk signals across a conversation rather than evaluating each message in isolation. The update targets scenarios involving mental health and crisis situations, where single-turn context is insufficient for safe response generation.
huggingface.co
Granite Embedding Multilingual R2: Open Apache 2.0 Multilingual Embeddings with 32K Context — Best Sub-100M Retrieval Quality
IBM released Granite Embedding Multilingual R2 under Apache 2.0, claiming state-of-the-art retrieval quality among models under 100M parameters with a 32K token context window. The model targets enterprise RAG and multilingual search use cases where both permissive licensing and long-context embedding are required.
huggingface.co
Unlocking asynchronicity in continuous batching
Hugging Face published a technical deep-dive on introducing asynchronous scheduling into continuous batching for LLM inference serving, addressing latency bottlenecks caused by synchronous iteration barriers. The approach decouples prefill and decode scheduling, improving throughput and reducing head-of-line blocking in high-concurrency serving environments.
rss.arxiv.org
Mistletoe: Stealthy Acceleration-Collapse Attacks on Speculative Decoding
Researchers identify a novel adversarial attack surface in speculative decoding systems: because draft models only approximate the target distribution, small adversarial perturbations to inputs can sharply reduce draft-token acceptance rates without visibly altering target model outputs, effectively collapsing the throughput gains speculative decoding provides. The attack exploits the structural drafter-target mismatch inherent to all model-based speculative decoding setups, raising practical security concerns for inference pipelines using this technique.