Gemini 3.7 Flash 확장, 로컬 Qwen3.8-27B, 에이전트 운영 계층화

Google Gemini

@GeminiApp

Gemini 3.7 Flash is now available to all Pro and Ultra users in Gemini chat. This model update delivers improved reasoning and accuracy for multi-step tasks like intelligently connecting the dots across dozens of files and emails into one master document. Give it a try today on the web or in the app!

It’s (finally) Friday Here’s our end-of-week recap: — This year’s @madebygoogle lineup (Pixel 11 series, Pixel Watch 5, and Pixel Tag) brings new AI integrations across devices. A few of the key announcements were Magic Capture for simultaneous video and photo capture, Rambler’s AI voice typing and text transformation, and expanded Live Transcribe for real-time ASL-to-text translation using the Pixel camera. Tying it all together is Gemini Intelligence, our proactive, agentic AI layer to anticipate user needs. — Gemini 3.7 Flash, our most intelligent workhorse model yet for coding and agents, is now available in the Gemini API via @googleaistudio , @antigravity , Spark in the @geminiapp , and the Gemini Enterprise Agent Platform. It’s also rolling out to paid users on @googleworkspace , Gemini App, and Search. — @GoogleDeepMind 's WeatherNext 2, an AI forecasting tool that can give meteorologists an extra day of accuracy when predicting a cyclone's track, intensity, and wind structure, is now open source. — The upgraded @Gemini_Notebook experience has been fully rolled out to all Pro users and the ability to copy a notebook has been rolled out to all users.

ClaudeDevs

@ClaudeDevs

Auto mode is rolling out today as the default permission mode in Claude Code for Pro, Max, and Team. If you've already set a default mode, Claude will ask before changing anything. You can still switch modes any time with Shift+Tab, or pin one with "defaultMode" in settings.

ClaudeDevs

@ClaudeDevs

To configure auto mode for your environment, run /auto-mode-setup. Claude looks through your repo and suggests trusted repos and domains for you to review. The more the classifier knows about your setup, the clearer the line on what's safe to run.

Share your results from Qwen3.8-27B! Here's 4-bit Qwen3.8-27B GGUF generating an interactive volcano simulation with geology, thermodynamics, fluid flow, buoyancy/drag, projectiles, cooling, groundwater interaction, and environmental effects in Unsloth Desktop. What’s your favorite prompt to test a new model with? GitHub: https:// github.com/unslothai/unsl oth …

The prediction that aged like fine wine - In January osman said we’d get Opus-level models on one GPU by year-end. - 201 days later: Qwen 3.8 27B beats Opus 4.6 Max in multiple benchmarks and fits on a single RTX 5090. We are in bro

> On the face of it, replicating a figure from a paper is not an especially creative endeavour. Most obviously, replication produces a figure that looks quite like the original, assuming that the method replicates. But if one examines the process, rather than the output, replication becomes a stepping stone towards innovation. > In a weak sense, Faraday not only replicates better than a frontier model; it also innovates better. > CAT (Coding Agent as a Tool) paradigm > Defining a verifiable reward necessitates specifying an evaluation procedure in foresight, which becomes a fixed target for manipulation. By contrast, judging entire rollouts in hindsight is a moving target. > In summary, we have created an intelligence layer with a modicum of research taste, sufficient to extend the capabilities of frontier agents. This is, however, the tip of the iceberg when it comes to imbuing agents with the ability to enrich scientific research as peer collaborators with humans. Stepping from replication towards innovation sharpens the problem of underspecification, and deepens the need to develop systems with good judgement.

> On the face of it, replicating a figure from a paper is not an especially creative endeavour. Most obviously, replication produces a figure that looks quite like the original, assuming that the method replicates. But if one examines the process, rather than the output, replication becomes a stepping stone towards innovation. > In a weak sense, Faraday not only replicates better than a frontier model; it also innovates better. > CAT (Coding Agent as a Tool) paradigm > Defining a verifiable reward necessitates specifying an evaluation procedure in foresight, which becomes a fixed target for manipulation. By contrast, judging entire rollouts in hindsight is a moving target. > In summary, we have created an intelligence layer with a modicum of research taste, sufficient to extend the capabilities of frontier agents. This is, however, the tip of the iceberg when it comes to imbuing agents with the ability to enrich scientific research as peer collaborators with humans. Stepping from replication towards innovation sharpens the problem of underspecification, and deepens the need to develop systems with good judgement.

Selecting the right model for a task, and for every subagent it spawns, is a decision you shouldn't have to make by hand on every run. Smart Routing makes it automatic. Turn it on once, and Omnigent routes each task and its subagents to the best model and harness for the job: Open Source: an LLM classifier picks the best model @databricks : powered by AI Gateway smart routing Learn more https:// omnigent.ai/docs/build/rou ting … #Omnigent #OpenSource #AIAgents #LLMOps #CodingAgents

Teknium

@Teknium

Hermes now has /loop - a command that will allow your session or bot to perform an action over and over again at any interval you like. It's like cronjobs, but lives inside a session with your context and other activities. Enjoy!

Tool-use is already a bottleneck in AI reasoning models, and it will become even more problematic as the underlying models become faster. The reason is that the sandbox, infrastructure, and execution environment needed for the tools (e.g., running code, search, database and file access) can be a very resource-intensive tax on top of the costs of running the AI model. For example, Hierarchical Decoupled Policy Optimization (HDPO), a technique developed by Alibaba, trains LLMs to avoid excessive tool-use. Another interesting example is Code World Model (CWM) by Meta, which uses a world model to predict code results without spinning up execution environments. This helps speed up the training of LLMs on coding tasks. Moving forward, as inference becomes faster and cheaper, I expect the other elements to catch up, such as faster containers and execution environments for AI tools.

We’ve written an FAQ to answer some of the questions we've received about watermarking. In summary: • We’re implementing watermarking to comply with the EU AI Act. Other major model developers have signed the same Code of Practice and will also be implementing watermarking; • Our watermarking method doesn’t have any practical impact on the quality or content of Claude’s outputs; • The difference between watermarked and un-watermarked text will not be distinguishable to readers; • Nothing is added to the text and there are no hidden characters; • Watermarking doesn’t require extra tokens, and will not be more expensive; • Watermarks can’t be traced to a specific person, organization, or chat. Read more:

As part of our Responsible Scaling Policy, we publish regular Risk Reports. These share detailed information on the risks of our systems and how prepared we are to address them. Our second Risk Report is now available: https:// anthropic.com/aug-2026-risk- report …

LLM engineer's handbook (30 minutes a day, 10 weeks, 50 lessons) a roadmap for llm inference serving where everything points at one service instead of scattering across demos. you get the mental model first, then serve a model, instrument it, load test it past 1000 concurrent requests, and tune it. you finish with a stack you configured yourself and a benchmark worth publishing. here is what it covers: → the roofline model, and why decode waits on memory while prefill waits on compute → vLLM internals, PagedAttention and the scheduler, read from the code → Prometheus and Grafana for TTFT, inter-token latency, and queue depth → SGLang and RadixAttention prefix reuse, benchmarked against vLLM → load testing past 1000 concurrent requests → quantization across FP16, FP8, and INT4, on quality as well as speed → speculative decoding and KV eviction, including where the gains disappear → disaggregated prefill and decode, deployed on Kubernetes → a cost, latency, and quality router with per-request token budgeting → publishing a reproducible benchmark the roadmap on GitHub: https:// github.com/patchy631/time -to-first-token … (don't forget to star ) i am also writing an article for each major topic. the first one is out, on how a GPU actually works. the article is quoted below.

Qdrant + Minima: agentic RAG on one RTX PRO 6000 Blackwell GPU. Hybrid search and late-interaction reranking made the first retrieval enough 87% of the time, up from 72%. Median latency 21.3s to 7.7s. 2.92x more successful tasks per GPU-hour. https:// qdrant.tech/blog/case-stud y-minima/ …

Runaway AI spend can start with a retry loop or an uncontrolled agent experiment. AI Spend Controls in Unity AI Gateway help teams catch it early. Organizations can set budget alerts and hard spend caps by user, use case, workspace, or account. Combined with Databricks budgets and cost data logged to Unity Catalog system tables, teams gain visibility and accountability across models, agents, MCPs, and providers. Learn how to monitor, contain, and govern AI spend at scale: https:// databricks.com/blog/introduci ng-ai-spend-controls-unity-ai-gateway?utm_source=twitter&utm_medium=organic-social …

Meta says AI should be “for everyone.” But is its AI actually accessible to everyone? The @EquityPod crew digs into the gap between Mark Zuckerberg’s vision and the hardware actually required to run Meta’s new open-weight model. https:// spr.ly/6014B1TESy