AI Daily — July 2, 2026
2026-07-02
anthropic.com
Anthropic launches Claude Science AI Workbench
Anthropic announced the Claude Science AI Workbench, a dedicated environment aimed at accelerating scientific research workflows. Details from the announcement page are sparse, but the product appears to target researchers needing structured, reproducible AI-assisted analysis. This is part of Anthropic's continued expansion of Claude into domain-specific professional tooling.
anthropic.com
Anthropic redeploying Fable 5
Anthropic published an announcement regarding the redeployment of Fable 5, though full details require visiting the page directly. This appears to relate to Anthropic's ongoing investment in AI safety research via the Fable program, which uses interactive narrative and simulation environments to study model behavior. The redeployment likely signals updated methodology or expanded scope for that research initiative.
engineering.fb.com
Meta's AI Storage Blueprint at Scale
Meta's engineering blog details the storage infrastructure underpinning its AI training and inference operations, as model sizes and dataset volumes have grown exponentially and frontier model release cadences have compressed from months to weeks. The post outlines architectural decisions for reliable, high-throughput storage access, positioning fast storage as a direct lever on both training speed and computational cost. This is a practical reference for teams building large-scale ML infrastructure.
huggingface.co
Hugging Face and Cerebras bring Gemma 4 to real-time voice AI
Hugging Face and Cerebras have integrated Google's Gemma 4 model with Cerebras' high-throughput inference hardware to enable real-time voice AI applications. The combination targets latency-sensitive use cases where Cerebras' wafer-scale chip architecture can deliver the token throughput required for interactive speech. This represents a practical demonstration of specialized silicon enabling multimodal inference at conversational speeds.
arxiv.org
GRPO, Dr. GRPO, and DAPO Are Three Operations on One Number: The Group-Standard-Deviation Identity
This paper shows that three widely-used RL-based reasoning training algorithms — GRPO, Dr. GRPO, and DAPO — are all variations of a single operation on the standard deviation of per-prompt answer correctness distributions. GRPO divides by this standard deviation, Dr. GRPO removes that division, and DAPO filters out prompts where it is zero, unifying what appeared to be distinct algorithmic choices under one identity. The result simplifies the design space for reasoning-focused RLHF and provides a principled basis for choosing among these methods.