AI Daily — 2026-04-22

English 中文

Introducing ChatGPT Workspace Agents for Shared Workflows · Google Cloud processes 16B+ tokens/mi...

Covering 27 AI news items

🔥 Top Stories

1. Introducing ChatGPT Workspace Agents for Shared Workflows

OpenAI announced ChatGPT workspace agents, shared agents that can handle complex tasks and long-running workflows across tools and teams. The feature aims to streamline cross-tool collaboration and automate multi-step tasks within organizations. Source-x

2. Google Cloud processes 16B+ tokens/min with Gemini Enterprise Platform

Google Cloud reported 16B+ tokens per minute processed via direct API usage, up from 10B last quarter, alongside the Gemini Enterprise Agent Platform and a new mission control for building, scaling, governing, and optimizing agents, plus 8th-generation TPUs. The move underscores Google’s push to scale enterprise AI workloads with end-to-end tooling. Source-x

3. Google Unveils TPU 8t and 8i: Dual-Chip AI Accelerators

Google introduced TPU 8t (training) and TPU 8i (inference), with 8t offering nearly triple the compute per pod versus Ironwood and 8i enabling high-throughput, low-latency inference to scale millions of agents cost-effectively. This aligns hardware with Google’s end-to-end AI stack, from software to models to applications. Source-x

📰 Featured

LLMs & Benchmarks

Qwen3.6-35B Excels in Polyglot Benchmark with Right Agent — Paired with Little-Coder scaffolding, Qwen3.6-35B reaches 78.7% in the Polyglot benchmark, placing in the top tier and rivaling cloud models; scaffold-model alignment issues are a noted factor and suggest avenues for future benchmarks. Source-reddit
Kimi K2.6 Tops OpenRouter Programming Leaderboard — K2.6 leads OpenRouter’s programming leaderboard, highlighting strong coding performance and boosting visibility within the AI developer community. Source-x

AI Safety & Industry

** Anthropic Faces Claude Mythos Leak, OpenAI Gains Ground** — Claude Mythos was unintentionally exposed on Discord, fueling concerns about safety and public perception while OpenAI gains momentum with Image Gen 2, Codex usage growth, and chatter about future releases. Source-x

Multimodal Vision & UI

** Flipbook Demonstrates Live Pixel Streaming From a Model** — A prototype streams every pixel directly from an AI model, bypassing traditional UI pipelines to explore a new AI-driven rendering paradigm for live UI outputs. Source-x
** ChatGPT Images 2.0 Creates QR Code Die Linked to Wikipedia** — A viral multimodal demo where each face of a game die shows a working QR code linking to Wikipedia articles, illustrating playful cross-modal capabilities. Source-x

Tools & Automation / Enterprise

** OpenClaw Codex harness auth fixed; OpenAI model integration improved** — An ongoing sprint fixes authentication for Codex within OpenClaw, restoring reliable harness behavior and improving agent performance; enables usage with ChatGPT subscriptions and signals further updates next week. Source-x

Fashion Tech & Vision

** Tstars-Tryon 1.0: Realistic Virtual Try-On at Scale** — Introduces a commercial-scale virtual try-on system described as robust and efficient, capable of handling extreme poses and lighting for scalable fashion item try-ons. Source-huggingface

⚡ Quick Bites

AnyRecon Enables Arbitrary-View 3D Reconstruction via Video Diffusion — Demonstrates converting video diffusion outputs into arbitrary-view 3D reconstructions. Source-huggingface
TEMPO Scales Test-Time Training for Large Reasoning Models — Proposes TEMPO to scale test-time adaptation for large reasoning models. Source-huggingface
Claude Context MCP Enables Semantic Code Search Across Codebase — Claude Context MCP enables semantic code search across large codebases. Source-github
RAG-Anything Launches Multimodal RAG Framework with VLM — Multimodal RAG framework integrating vision-language models for retrieval-augmented generation. Source-github
Dense vs MoE gap narrows with 3.6-27B release — The performance gap between dense and Mixture-of-Experts models narrows with the 3.6-27B release. Source-reddit
Qwen3 TTS Underrated, Real-Time Local Streaming Achieved — Qwen3 TTS achieves real-time local streaming, highlighting efficient on-device speech synthesis. Source-reddit
Recent Open LLMs (Nov 2025–Apr 2026) Overview — Survey of open LLMs released in late 2025 through early 2026. Source-reddit
Qwen 3.6-27B LLM Released — Qwen 3.6-27B released, expanding open-model options. Source-reddit
MiMo-V2.5 Released by Xiaomi — Xiaomi releases MiMo-V2.5, expanding on-device multilingual model capabilities. Source-reddit
CoInteract Enables Physically-Consistent HOI Video Synthesis — CoInteract advances physically-consistent human-object interaction video synthesis. Source-huggingface
AgentSPEX: Specification and Execution Language for Agents — Introduces AgentSPEX for formal specification and execution of autonomous agents. Source-huggingface
Microsoft Launches AI Agents for Beginners Course — Educational course introducing AI agents for beginners. Source-github
TrendRadar: AI-driven public opinion and trend monitor — Opensource tool for tracking AI-related public opinion and trends. Source-github
Why a 27B Model Could Beat 397B MoE in Qwen — Discussion on why smaller dense models can outperform large MoE scales in Qwen. Source-reddit
Claude’s high demand tests supply constraints — Reports on Claude demand stressing supply constraints. Source-x
Anthropic’s Claude change widely visible and buggy — Claude updates are visible but buggy, drawing scrutiny from users. Source-x
Claude Code token budget at work now higher than gross compensation — Claude Code token budget cited as larger than many compensation figures. Source-x

Generated by AI News Agent | 2026-04-22