Opus 5 원샷 생성·vLLM v0.26 최적화·프롬프트 기반 연속작업 동향

Ok Opus 5 is insane. People are already one-shotting 3D games, worlds, and Blender builds with it. 10 wild examples:

9. Claude generated this cinematic sequence in 10 minutes First try. Zero editing.

vLLM v0.26.0 is out! 411 commits from 212 contributors (61 new). Highlights: Attention backend selectable per KV-cache group Sliding window is now an explicit backend capability Tiered KV offloading with an object-store secondary tier DeepSeek-V4 speedups across NVIDIA, ROCm and XPU Inkling family lands with a full support stack Rust frontend adds multimodal video, audio and native vllm-bench Thread

Serving, frontend, and what to know before upgrading: Decode Context Parallel: hybrid attention support, plus DCP + Eagle for Tokenspeed MLA backends PD disaggregation: NIXL pipeline-parallel prefill in push mode, and encoder-cache connectors including CPU offloading Rust frontend: multimodal video and audio, Seed-OSS tool parser, native vllm-bench port OpenAI API: bad_words in /v1/completions, logprob_token_ids on Python endpoints, include_reasoning for non-Harmony models, plus a new endpoint plugins framework Quantization: Humming w[2-7]a[4,8] weight-only inference, nvfp4_per_token online MoE quantization, CuTe-DSL FlashInfer MXFP4, INT2 weight-only on XPU Security: diskcache replaced to remove pickle deserialization, a concurrent sparse-invariant race fixed, bounded prompt lists and a regex-compilation timeout Before you upgrade: • Dependency bumps: Transformers 5.13.0, FlashInfer 0.6.14, NIXL 1.3.1 • Models removed: TeleChat, Persimmon, Fuyu Thanks to all 212 contributors this cycle (61 first-timers). Full release notes →

Sam Altman

@sama

chatgpt work is remarkable, and "work" undersells it. from my phone i sent: "use all my chat history to figure out ideas for a long weekend trip with 8 friends, plan the best three options, make a full-stack site where the 9 of us can coordinate on what we would want to do in each place and decide where to go, and then after we get to group agreement make reservations. draft an email in my gmail i can send out to my friends when the site is ready." it...just worked.

Greg Brockman

@gdb

put chatgpt to work

The Top AI Papers of the Week (July 20 - July 26): - GAMUT - PRO-LONG - Harness Handbook - From Memory to Skills - Progressive Disclosure - Global Workspace in LLMs - Structured Output Collapses Diversity Read on for more:

"Expanding Flow Maps" This paper propose a way to let flow models generate without a fixed canvas. Instead of choosing output size upfront, EFM learns to expand the state with new coordinates or tokens, then transport and denoise them in a few steps. This turns variable length text, graphs, and 3D molecular conformers into one flow framework where size becomes a learned part of generation.