COMPUTE IN ORBIT
SpaceX and Nvidia Will Send a Vera Rubin NVL72 Into Orbit
A space-optimized Rubin rack is slated to launch in Q4 of next year, with significant scale by 2028.
SpaceX, in partnership with Nvidia, has designed a space-optimized Vera Rubin NVL72 system for launch to orbit in the fourth quarter of next year, with significant deployment planned for 2028. The pairing pushes AI-compute capacity beyond terrestrial gigawatt factories, matching Rubin-class accelerators to the power, thermal, and mass budgets of a spacecraft.
DEVELOPER TOOLS
GPT-5.6 Lands on kiro.dev
OpenAI's newest models now slot into the production workflows developers already use to plan, build, test, and review software.
INFERENCE
Sol Engine Drags H3's 768p Latency to 14.93s
A 4-step low-res H3 draft plus a 3-step LTX refinement with Sol-Attn takes a 10-second 768p clip on a single GB200 from 414s down to 14.93s.
AGENTIC CPU
Vera CPU Powers SpaceX's Agentic AI at Scale
NVIDIA says Vera, its first CPU built for agents, is accelerating orchestration, code execution, and data processing at SpaceX.
PRODUCTION
Groq 3 LPX Enters Full Production
NVIDIA's Groq 3 LPX is now in full production, with Vera Rubin NVL72 positioned as the foundation of every AI factory.
Releasing AI research openly used to be the norm.
Andrew Ng — on the Marin project's open training
RESEARCH SHIFT
Chinese Open-Source LLMs Now Cited in ~40% of AI Papers
A Codex-assisted scan of 500,000 arXiv AI/ML papers since ChatGPT found mentions of Chinese open models rose from roughly 10% in 2024 to nearly 40% today — overtaking American open models, which held near 30%.
Claude Long Replies Now Stream ~4x Smoother
A rebuilt renderer touches only what's still changing, cutting long-answer jank 9x and holding 120fps on a 120Hz MacBook.
Claude's Enterprise MCP Auth Goes GA
Team and Enterprise admins centralize connector authorization through their identity provider — no per-user OAuth.
WAN 3.0 Hits Runway
Generate video and audio from multiple image, video, and audio reference inputs.
Pika Adds WAN 3.0
30-second generations, 20 reference inputs, enhanced audiovisual realism — up to 35% cheaper via Pika API Club.
World Model LDR: 20x Less Extrapolation Error
Latent dynamics reasoning beats video-diffusion baselines on PhyWorld with 26x fewer parameters and 143x faster inference.
Runway Ruby Sends Any Model to Delivery Spec
Seedance 2.5, Gen-4.5, or MiniMax H3 output converts straight to 16-bit EXR or 10/12-bit ProRes and HEVC.
Frontier Labs Shift From APIs to Vertical Agents
A San Francisco founder dinner weighed moats in the AI era — ChatGPT Health, Claude for Legal.
AI Short Open-Sourced, $1M Film Festival
"If You Stop Loving Me, I'll Die" ships with all prompts and assets public.
Mistral Partners With HUMAIN
AI infrastructure, model development, and localized frontier models for Saudi Arabia and the region.
GLiNER2.5 Ships Its Biggest Upgrade
fastino calls it the most significant update to the GLiNER architecture yet.
Thomson Reuters Builds on Qwen
Its CTO says self-built models on Alibaba Qwen reduce reliance on Claude.
Grok Voice Scales at Starlink
Musk: Starlink runs Grok Voice at scale for customer support and sales.
Headlong: Persistent Agents, Open Source
A micro-harness for self-guided agents that think continuously.
sPTC Brings Speculative Tool Calls
Speculate on tool calls during generation and discard if unneeded — like CPU speculative execution.
Grok Bot 0.18.0 Shipped Source Maps On
An analyst recovered the full macOS client source; an unofficial rebuild is now public.
Xiaomi's Mystery 6nm Chip, Under a Microscope
Wafer-on-wafer NPU-DRAM bonding at 1.4µm pitch — an analyst likens it to logic folding.
Andrew Gordon Wilson Joins Perplexity
To lead research on continual learning, synthetic data, and long-horizon RL.
Qwen3.8 27B Gets an NVFP4 Checkpoint
BF16 lm_head plus FP8 KV cache improves accuracy.
ARC-AGI-3 Climbs to 4.58%
Tufa Labs sets the new high and open-sources its approach.
Epoch AI: GDP Undercounts Nvidia
US GDP statistics miss much of Nvidia's economic value.
Jensen: xAI Pulled Off the Impossible in 19 Days
Huang praises xAI's engineering sprint.
Grok Build Adds Browser Use
Grok now drives a real browser to finish tasks.
A Complete RL-for-LLMs Guide
Policy gradients, PPO, GRPO — from first principles to frontier.
Kotoba Does Real-Time Voice-to-Voice
Simultaneous translation keeps the speaker's own voice.
Hugging Face Adds "Buckets"
A new repo type beyond models, datasets, and Spaces.
Sasha Rush Leaves Cursor
The NLP researcher departs, calling it "an incredible place."
Americans Most Skeptical of AI
Korea and India most enthusiastic in a 25-country poll.
Firefly's "Era" Prompt Pack
One photo, restyled across historical eras.
Pika Pits WAN 3.0 Against Seedance 2.5
The same perfume brief yields near-final ad spots from both.
Wan3.0 Reviewers Call It a Step-Change From 2.7
Third-party tests praise character consistency.
WAN 3.0 Holds Up at 30 Seconds, 1080p
A wave clip stays coherent end to end — and cheap.
Wan3.0 Leaned on Aesthetic Experts
Advisors and richer pretraining data sharpened visual quality.