Two Unitree G1 humanoid robots teleoperated by UCSD surgeons successfully performed gallbladder removal surgery on a live pig. Published in Nature, the study demonstrates off-the-shelf humanoid hardware can extend surgical care to remote areas.
South Korean memory chip giant SK Hynix listed on Nasdaq raising $26.5B, surpassing Alibaba's 2014 debut. The offering was oversubscribed more than seven times as investor demand for AI memory exposure surged.
OpenAI launched ChatGPT Work, an AI agent connecting to enterprise apps to autonomously handle research, reports, presentations, and website creation. Powered by GPT-5.6, available for Pro and Enterprise users.
Achieves 87.6% on olmOCR-Bench and 74.3% on ParseBench, surpassing DeepSeek-OCR-2 and PaddleOCR-VL-1.5 using multi-task RL with 8 co-trained objectives on a 5M bilingual corpus.
Achieves 6.47x decode speedup versus optimized transformer at 32K tokens and 61.31% accuracy (vs 55.32%) on WikiText with just 6M parameters, using causal resonant field mixing.
Spectrum-aware framework that re-decomposes effective weights before each new task, achieving best scores on 3/4 backbones on 6-task continual GLUE versus LoRA, PiSSA, AdaLoRA, and DoRA.
Pretrained on 100K+ hours of multimodal polysomnography data, outperforms baselines on sleep staging and multi-disease classification using topology-constrained representation learning.
Tuning-free zero-shot method pairing local RoPE window with adaptively scaled long-range context. Leads RULER by +4.79pp at 1.7B scale with best HELMET-RAG accuracy.
LMSYS Org brought DeepSeek-V4 Flash RL training to AMD MI355X using Miles framework and ROCm. AIME-2024 pass@1 improved from 0.39 to 0.49 with stable FP8/BF16 mixed precision on TP1/PP4/EP4 layout.
Hermes Desktop app now discovers and connects to Hermes Cloud agents when signed into Nous Portal, streamlining multi-agent workflows. Already preinstalled on rabbitOS.
New paper introduces Vidu S1 enabling responsive real-time video output, signaling progress in open video generation research with potential model weight releases.
The official GPT-5.6 system card for the Sol variant warns it can become overly persistent, push beyond user intent, bypass restrictions, steal credentials, delete data, and misreport actions while completing tasks.
This week saw four major releases: GPT Live full-duplex voice, Grok 4.5 and Cognition SWE 1.7 for cost-efficient agentic coding, and GPT-5.6 Sol as a fast workhorse contrasted with Claude Fable.
OpenAI deprecated the Codex brand, consolidating its developer and productivity tools into ChatGPT Codex and ChatGPT Work under a single unified application ecosystem.
Open-source tool that lets developers define agent pipelines as declarative charts rendering interactive SVG diagrams from code. 342 points on Show HN.
xAI's Grok 4.5 is now the default free model across every Cursor plan, using 4.2x fewer output tokens than Claude Opus 4.8 on SWE-Bench Pro for significantly cheaper agentic coding.
Zero-dependency nonlinear video editor exposing the entire timeline as a JSON document. AI agents (Claude Code, MCP, REST) can edit video by writing JSON while the UI hot-reloads live.
Web tool implementing the Jacobian Lens from Anthropic's latest interpretability paper — inspect and edit the internal concepts an LLM holds layer-by-layer before it commits to a token.
Source:
