Noam Shazeer rejoins OpenAI. Co-author of the Transformer paper returns to OpenAI shortly after rejoining Google, marking a major talent move in the frontier-model race. Source

Anthropic Claude Fable 5 launch and abrupt suspension. Anthropic's Mythos-class model launched and was quickly pulled, setting a notable safety/policy precedent. Source

Anthropic Claude Opus 4.8 released. Frontier increment release extending the Claude line. Source

Baseten reportedly raising $1.5B. AI inference startup in talks for another mega-round just months after its last raise. Source

Amazon pushes to challenge Nvidia with its own AI chips. Amazon plans to sell in-house accelerators more directly against Nvidia. Source

cuTile Rust + Grout inference engine. Rust ownership/borrow checking extended to GPU kernels; Grout on Qwen3 reportedly competitive with vLLM/SGLang. Source

TesterArmy (YC P26) launches. AI agents that autonomously test web and mobile apps. Source

Adobe adds AI assistants to Photoshop and Premiere. New assistants integrated into Adobe's creative suite. Source

Adobe Firefly AI studio redesigned with persistent memory. The AI studio now remembers prior creations across projects. Source

Scalable 2D Vision-Language Models for Radiology. Training pipeline for spatially grounded VLMs in radiology at scale. Source

UltraQuant: 4-bit KV caching for context-heavy agents. Cuts memory for long-context LLM agents via aggressive KV quantization. Source

ImageWAM: Do world action models need video, or just image editing?. Investigates whether WAMs require full video generation or can rely on image editing. Source

Calibrated MoE under distribution shift. Methods to keep mixture-of-experts routers well-calibrated when deployment drifts from training. Source

AI #173: AI Pauses (Zvi Mowshowitz). Weekly roundup covering the week's pauses, releases, and policy moves. Source

Open-weight open-source model claims #1 on AA Intelligence Index. Viral thread walking through benchmark results. Source