Grok 에이전트 실행, ChatGPT 작업 기억, 하드웨어 인식 최적화

My workflows changed when I started treating each Grok Bot like an application Client: chat Server: @bot DB: anything (gh, sheets, SaaS) I've been white-pilled on "agents for everything" - thin client, thick server

Grok @bot visualized human migration using @imagine to generate assets, then wrote the visualization of human migration. It also cut and edited the videos, making sure the message is coherent. Pretty good video editing tool just using chat.

Elon Musk

@elonmusk

Grok 4.6

BREAKING: Grok 4.6 just tied for #1 on the Artificial Analysis Agentic Index. • Grok 4.6 (High) — 59 • Claude Opus 5 (Max) — 59 Grok 4.6 is now at the top of the benchmark, ahead of other leading models. The Agentic Index focuses on capabilities that matter for real-world AI agents — tool use, planning, autonomy, and complex problem-solving. This is especially significant as Grok expands into Grok Build and Grok Bot, where AI needs to do more than answer questions — it needs to use tools and complete tasks. Source: Artificial Analysis

OpenAI Developers

@OpenAIDevs

Europe, two updates for you. Computer History is now available in the EEA, UK, and Switzerland for ChatGPT Pro, Business, and Enterprise users on Mac. ChatGPT can remember activity across your apps and websites, so future conversations feel more personalized and require less explanation.

OpenAI Developers

@OpenAIDevs

Record & Replay is also available in the ChatGPT app for macOS across the EEA, UK, and Switzerland. Turn your go-to workflows into reusable skills by showing, not just telling, ChatGPT Work and Codex what to do.

We’ve seen an explosion of new ML chips with unique architectural features, but software support remains the critical bottleneck Achieving peak performance increasingly relies on hardware-specific optimizations in the kernels, but we observe that coding agents are particularly weak at this Introducing Hawkeye, a framework that brings hardware-awareness to coding agents by grounding them in a minimal and comprehensive taxonomy of optimization strategies For new GPU or ML accelerator architectures, you only need to write 10 unit tests and solution kernels (one per optimization strategy), and we show that coding agents can effectively scale test-time compute with this minimal supervision to write hardware-aware kernels Hawkeye can port kernels across architectures (Ampere, Hopper, Blackwell), vendors (NVIDIA, AMD), and precisions (FP8, NVFP4, MXFP4) while consistently leveraging hardware features and approaching expert kernel performance Work co-led with @keramakr and done in collaboration with Alexander Ingare @simonguozirui @18jeffreyma @ZishenW @simran_s_arora @Azaliamirh @profvjreddi

Waymo has built its own custom chip for its robotaxis, an application-specific integrated circuit made by TSMC on a 5-nanometer process that is capable of more than 1,000 TOPS. The chip is in production in Waymo's latest generation of vehicles and reduces its reliance on outside suppliers including NVIDIA.

I’ve been looking forward to today for 3 years Today we’re announcing Foundation - Chroma’s solution to memory Our research preview of this technology builds self-improving memory from your agent sessions. Try it out at https:// trychroma.com/foundation

Harrison Chase

@hwchase17

very cool launch we did a webinar with jeff on "wiki" style memory and it's clear he'd thought about this problem a lot (webinar here: https:// youtube.com/watch?v=Lsut4T Cfygw …)

AI at Meta

@AIatMeta

Muse Spark 1.2 supports a broad range of multimodal tasks, from turning visuals into working code to translating perception into physical action. It also brings robust audio-visual understanding to enable video-heavy workflows common in real-world enterprise use. Today, we’re sharing new evals and demos that illustrate the breadth of the model’s visual understanding and reasoning capabilities. Let’s start with a demo that shows how Muse Spark parses multimodal observations and calls tools to guide a robot to navigate in an unstructured environment to find a rubber duck.

AI at Meta

@AIatMeta

Muse Spark brings spatial intelligence to robotics. A specialized variant of Muse Spark acts as the robot brain and orchestrator: it takes a user instruction, decodes tool calls, observes the results, and loops until the task is complete. As shown in this demo, Muse Spark 1.2 can plan sub-tasks for a bimanual robot tidying a desk, distinguishing a hair brush from a makeup brush and placing the lipstick in a drawer.

Finally a good paper testing whether agents can really post-train other agents. (bookmark it) They analyzed a large corpus of publicly released post-training trajectories. Across tasks, the agent locks in its training strategy at the very first step and spends the entire remaining budget on local adjustments inside it. They then tried three escalating fixes. An experience-driven scaffold lifted execution broadly, worth 12.6 points on GSM8K and 40.8 on HumanEval, and the strategy stayed frozen. Human guidance redirected the opening choice, and the agent slid back into local loops once training began. Extra inference compute paid off on easy tasks and did almost nothing on the hardest one. What agents lack here is a way to reconsider strategy while execution is still running. Paper: https:// arxiv.org/abs/2608.19072 Track more trending AI papers in our academy: https:// academy.dair.ai

dots3-note Preview is an open-weight model built for tasks that run for hours—sometimes days. → 280B total / 16B active parameters → 512K context → Text, vision, and speech → Reasoning, coding, and tool use But the most interesting part isn’t the model size. It’s how the model learns to evaluate its own progress before a long task is finished.

The Embedder’s Dilemma— the ultimate question! Find out when you should use an embedding model vs an LLM!