AI demand could triple power needs by 2035—pressuring PJM and Texas; Nvidia’s Alpamayo‑R1 narrates its driving plans; Perplexity adds opt‑in, portable memory across chats.
Google aims to deeply personalize AI responses using user data from services like Gmail and Drive, promising helpfulness but raising privacy concerns as the line between assistant and surveillance blurs.
Apple named Amar Subramanya as its new VP of AI, bringing experience from Microsoft and Google as it tries to accelerate its lagging AI efforts. He replaces John Giannandrea, who will advise until retiring due to delays to Siri upgrades.
Alibaba and ByteDance are training advanced AI models in Southeast Asia using foreign data centers with Nvidia GPUs to navigate U.S. export controls, while deploying the resulting models in China on local chips as training abroad remains legally permitted.
Data center power demand is set to reach 106 GW by 2035 as far larger facilities and a surge in AI-driven workloads fuel rapid expansion, especially in the PJM region and Texas, prompting warnings that grid capacity may struggle to keep up.
Nvidia released Alpamayo-R1, an open-source self-driving model that narrates its reasoning as it interprets road scenes and plans movements, offering greater transparency to help developers diagnose and improve autonomous-vehicle behavior.
DeepSeek-V3.2-Exp debuts DeepSeek Sparse Attention to boost long-context efficiency with minimal output trade-off, matching V3.1-Terminus in benchmarks and resolving prior RoPE indexing issues.
FLUX.2 delivers high-detail, photoreal image generation with multi-reference support, improved text rendering, and stronger world knowledge, powering creative workflows across open and pro model tiers while advancing Black Forest Labs’ open-core approach.
DeepSeekMath-V2 uses a self-verifying approach to train a proof generator and verifier that improve mathematical rigor, enabling strong theorem-proving performance including gold-level IMO 2025 and a 118/120 Putnam 2024 score.
Compact, human-readable, schema-aware JSON for LLM prompts.
Perplexity now uses a secure memory system that stores preferences and past conversations to deliver more precise, personalized answers across all models, offering continuity, context portability, and full user control over what is remembered.
It is available to all Google AI subscribers, now including Pro and Plus.
You can talk, watch answers appear, review earlier messages, and see visuals like images or maps in real time.
New ad related strings appeared in the latest build for Android and will likely be limited to search ads only.
Equally concerning is what the drop implies: that the earlier growth was mostly from paid marketing (organic growth rarely drops that quickly).
This guide explains how prompt caching really works, showing that KV-cache reuse depends on identical prefixes across requests and is enabled by vLLM’s paged-attention, which maps token blocks to reusable hashed memory pages for efficient, scalable inference.
A diff-in-diff study of GitHub projects shows Cursor adoption briefly accelerates development but consistently raises static analysis warnings and code complexity, which later contribute to slowed long-term velocity.
A curated weekly roundup of AI news, tools, and breakthroughs to stay ahead of the curve in a fast-moving industry.
Curated by devs, for devs. Helping 400+ developers stay ahead of the AI curve.