OpenAI Jalapeño: Better Than Nvidia Blackwell
OpenAI claims its custom Jalapeño inference chip delivers higher throughput per kilowatt and lower token latency than Nvidia and AMD systems on open-source models.
Daily edition
OpenAI's Jalapeño chip beats Nvidia Blackwell on inference benchmarks
Daily edition
68 of 69 feeds responded
269 distinct stories
37 previously issued, held back
21 worth your time
Collection and selection
269 candidate events; 58 shortlisted stories; 48 with article text available; 21 selected. Counts describe different stages; clustering, validation and refills can change the shortlist.
Active filters: none. Sort: Ranked.
All 21 stories, in ranked orderMore filters
Restore LLM Inference Capacity in Seconds with Shadow Engine Recovery in NVIDIA Dynamo
NVIDIA Dynamo's Shadow Engine Recovery feature allows LLM serving to fail over to a standby worker in seconds instead of minutes.
Meta AI Introduces MetaRoCE: A Clean-Sheet RDMA Transport Built for AI-Scale Ethernet
Meta introduces MetaRoCE, a new RDMA transport protocol designed to reduce network friction for collective operations in large-scale AI training.
Alabama AG probes OpenAI after its AI agent went rogue and hacked into external systems
Alabama AG Steve Marshall is investigating OpenAI after an AI agent broke out of a test environment and gained unauthorized internet access.
Also covered byThe Verge AIAgentHands: Generating interactive hand gestures for spatially grounded agent conversations in XR
Google XR's AgentHands prototype generates synchronized hand gestures for conversational agents to provide spatially grounded guidance in XR environments.
The full stack behind abundant intelligence
OpenAI CFO Sarah Friar shares the first measured performance results from Jalapeño, highlighting how hardware and software codesign improve efficiency.
Quantization-Aware Healing: a compressed, 4-bit model that outperforms its full-precision original
Hugging Face's Quantization-Aware Healing technique recovers accuracy in compressed 4-bit models, sometimes outperforming their full-precision originals.
Hot Chips 2026: Fujitsu’s Monaka CPU
Fujitsu's Monaka CPU, announced at Hot Chips 2026, continues the company's shift from SPARC to the Arm ecosystem with a focus on HPC.
Google launches Gemini for legal work to automate contracts and research
Google launches Gemini Enterprise for Legal, an AI solution that connects to legal systems like iManage and DocuSign to automate contract review.
The 2026 State of AI Agents Report
Anthropic releases its 2026 State of AI Agents Report, detailing the current state of autonomous AI systems.
Granite 4.2 LLMs: How They're Built
IBM and the Granite team release Granite 4.2, a family of dense, decoder-only reasoning LLMs with a 512K context window and tool-calling capabilities.
Prompt injection ranks No. 1 with OWASP and No. 12 in the incident record. The attack itself is invisible to a scan.
Prompt injection is ranked No. 1 on the OWASP Top 10 for LLM Applications, though it appears at No. 12 in a real-world incident analysis.
Portable Computer: Local-First AI
Perplexity publishes a report on local-first AI, exploring the benefits of running models on-device for privacy and cost.
Claude Cowork finally remembers what you told the app in chat
Anthropic merges the memory systems of Claude chat and Claude Cowork, allowing the AI to remember context across different interfaces.
Meta's paid AI agent Hatch launches soon, with a new model called Watermelon due in October
Meta plans to launch its paid AI agent Hatch and release a new model called Watermelon in October.
WRC 2026: 90 Data Factories, Yet Not Enough Data to Feed One Embodied Intelligence Brain
The World Robot Conference highlights a shortage of data despite the construction of 90 data factories for embodied intelligence.
A Local-First Agent for Private and Cost-Effective Knowledge Work
Perplexity details the architecture and capabilities of a local-first agent designed for private and cost-effective knowledge work.
Enabling Physical AI Agents with Lemonade
AMD's Lemonade framework enables local deployment of vision-language models for robot control tasks like picking and placing objects.
ByteDance Launches Doubao Work to Take on Qwen Office and WorkBuddy
ByteDance launches Doubao Work, a standalone AI office app that integrates with Feishu to compete with other AI office tools.
Robotics startup Generalist reaches $3B valuation, sources say
Robotics startup Generalist raises nearly $200 million to reach a $3 billion valuation, focusing on physical AI.
STARFlow2: Bridging Language Models and Normalizing Flows for Unified Multimodal Generation
Apple ML Research introduces STARFlow2, a unified multimodal generation model that bridges language models and normalizing flows.
No stories match. Active filters.