에이전트 운영 계층·적응형 추론·멀티모달 생성 모델의 동시 확장

Claude Code 2.1.233 has been released. 20 CLI changes Highlights: • Added opt-in memory cgroup via CLAUDE_CODE_TOOL_MEMORY_LIMIT to keep Bash tool runs from stalling sessions • GitLab MR URLs accepted by --worktree and shown as !N in claude agents, enabling MR linking/navigation • Windows path handling blocks \??\ NT path bypass and revalidates UNC paths to prevent NTLM credential leaks Full details are in thread ↓

Claude Code CLI 2.1.233 changelog: New features: • Added GitLab merge request URL support to the --worktree flag and the claude agents view (where MRs display as !N) • Added an opt-in forward_user_identity apps gateway setting on Anthropic upstreams that sends the signed-in user's identity as headers, so a proxy behind the gateway can attribute spend per user • Added opt-in memory cgroup support for Bash tool commands on Linux (CLAUDE_CODE_TOOL_MEMORY_LIMIT) so a runaway build can't stall the session • Added CLAUDE_CODE_WEBFETCH_CACHE_TTL_MS environment variable to configure the WebFetch session URL cache TTL (default unchanged: 15 minutes) Fixes: • Fixed cloud sessions occasionally being marked as lost when the environment shut down while Claude was waiting on a permission prompt • Fixed MCP v2 connections endlessly reopening the subscriptions/listen stream against servers that terminate long-held streams on a fixed timeout (e.g. serverless hosts) • Fixed Notification hooks not firing for permission prompts when running under Claude Desktop or VS Code • Fixed idle sessions on Linux sometimes keeping one CPU core at 100% when sandboxing is enabled • Fixed bundled skill aliases like /checkup and /review reporting "Unknown command" in -p mode or with plugins/MCP loaded when a user or project skill shadows the bundled skill • Fixed skill/command argument substitution to prevent argument values from being re-expanded as template markers • Fixed Windows paths spelled with the NT \??\ device prefix bypassing UNC path validation, closing an NTLM credential-leak vector • Improved apps gateway error forwarding: 400/413 errors from Vertex, Foundry, and Claude Platform on AWS upstreams now carry the upstream's own message; fixes a bug with auto-compact on apps gateway • Windows: fixed auto mode repeatedly stopping for manual approval on ordinary cd <dir> && <command> > file Bash commands (a 2.1.232 regression) Improvements: • Improved claude self-hosted-runner session start time: the session branch is now created without rewriting the working tree, and two server round trips no longer block the agent's launch • Improved claude plugin validate to check a bare .claude/skills directory, reporting SKILL.md files whose frontmatter fails to parse • Improved screen reader mode: the /effort selector renders as a numbered list with a typed-number prompt, and hint and dialog text is no longer clipped • Improved print mode diagnostics: a [claude-code:unrecognized_model] line is written to stderr when a request goes out for a model ID Claude Code doesn't recognize; map it with modelOverrides to silence Security/safety change: • Reverted the 2.1.232 Bash permission changes for Cygwin-style symlinks on Windows and for input redirections (< file); a narrower version will return in a later release Other changes: • Changed the GitHub app setup tip to no longer appear in repositories whose origin remote is on http:// gitlab.com or http:// bitbucket.org; the enterprise marketplace tip now covers non-GitHub internal git hosts • Todo/task-tracking tools (TaskCreate/Get/Update/List, TodoWrite) are no longer available on Opus 4.8, Sonnet 5, Fable 5, Mythos 5, and newer models; set CLAUDE_CODE_ENABLE_TODO_TOOLS=1 to bring them back Source: https:// github.com/anthropics/cla ude-code/blob/main/CHANGELOG.md#21233 …

AI just fixed Will Smith eating spaghetti... 3 years later Pika Soundtrack can add music, speech + motion-aware SFX to any video. It's one of 4 new Pika Audio Models: Soundtrack. Speech. SFX. Music. RIP silent AI videos.

Pika just killed the silent AI video problem. Pika just dropped 4 Pika Audio Models: Soundtrack. Speech. SFX. Music. And Pika says they’re the least expensive audio models on the market.

Six weeks ago, serving DSpark in vLLM meant choosing a draft length for your traffic and living with it. Now you set that length once, and vLLM decides how much of the draft to verify every step. On DeepSeek-V4-Pro-0813, the first token of a 7-token draft survives verification more than 70% of the time. The last one, less than 10%. One config, adaptive verification on with num_speculative_tokens 7, holds the Pareto frontier from concurrency 1 to 256 on 8×B300. Long draft at low load, short at high load. On main behind enable_adaptive_verification, for DSpark with a confidence head on Flash Attention or DSV4 attention. More backends and models are in bring-up. Thanks to Lucas Wilkinson ( @RedHat_AI ) and Ben Chislett ( @NVIDIAAI ), and to @deepseek_ai for DSpark and the varlen DeepGEMM indexer kernel. http:// vllm.ai/blog/2026-08-1 4-dspark-adaptive-verification …

Jerry Liu

@jerryjliu0

23일 전

There's a lot of real-world documents that are scanned, rotated, handwritten, or some combination of any of these elements. This week we created a comprehensive document extraction benchmark that contains documents tagged with various "perception challenges", along with other tags denoting task challenges, table structure, business domain. These docs include regulatory filings, hand-filled tax forms, photocopied docs, sensor noise, and more. Codex is surprisingly good at scans, but not great on rotated docs. OCR solutions are reasonable on rotations/handwriting but struggle on more general scans. Check out ExtractBench! ArXiv: https:// arxiv.org/pdf/2607.29677 Site: https:// extractbench.ai

Did you know Hermes Desktop can run against a Hermes Cloud agent? Point Hermes Desktop at Hermes Cloud in Settings > Gateway and get an agent that keeps working after you close the laptop, reachable from anywhere. Deploy your own Cloud today at https:// portal.nousresearch.com/cloud

Securing AI agents takes more than locking down an API key. AI agents need more than static API keys. In this Agents of Dev episode, Snowflake Chief Security and Trust Officer Mayank Upadhyay explains how workload identity, model routing, cost management and MCP tool governance change the security model for agentic workflows.

Elon Musk

@elonmusk

23일 전

Grok 4.6 now in Copilot

"From RLVR to RLSVR: Task Transformation Induces Self-Verifiable Rewards for Open-Ended LLM Self-Improvement" RLSVR extends RLVR beyond math and code by making rewards verifiable through task transformation. The researchers used SpyRL to instantiate this with information-asymmetric self-play. One agent gets degraded context, all agents perform the same task, then vote on the spy. The known spy identity gives an exact detector reward, while vote counts become a relative performer reward. This converts subjective output quality into a self-generated RL signal without reward models or external judges. https:// alphaxiv.org/abs/2607.23802

This is quite an interesting paper... DEXA scans are used to measure bone density and body composition. Here, authors trained a LeJEPA model on 11,540 DEXA scans which can then be used to extract clinically meaningful, whole-body representations of body. For example, the model can predict Type 2 Diabetes, osteoporosis, biological age, and so much more. Link to paper: https:// arxiv.org/abs/2608.02208

"Massive Activations in Hybrid Linear Attention Large Language Models: Pre-Attention Spikes and Inter-Spike Plateaus" This paper finds hybrid LLMs seem to organize computation around their expensive full-attention layers. So the linear-attention layers are not just doing the same computation more cheaply. They seem to be building up and carrying information toward full-attention layers, where that information is globally mixed, then reset or redistributed afterward. These activation spikes are like markers for where important computation is happening inside the network, potentially making hybrid models easier to interpret and potentially help researchers decide where full attention is actually necessary instead of inserting it heuristically. https:// alphaxiv.org/abs/2608.12149 v1 …