에이전트 실행 계층과 로컬 모델 성능 경쟁, Claude Code 제작 기능과 텍스트 워터마크 판별

I don’t know people can realize it or not, Qwen3.8 27B scoring it gets 67.3 overall, putting a 27B open model basically alongside This is the very first time a model that you can run Locally On your desk is basically on par with the best ones. We've reached the point where local open models are capable enough to perform most everyday tasks. I’m not talk about only this 27B model, I’m talking about next upcoming models!

Qwen3.8-27B matches GPT-5.6 Luna Just watched a 27B open model run entirely in the browser generating interactive 3D voxel worlds in real time and it locally via WebGPU.

ClaudeDevs

@ClaudeDevs

Claude Code can design now. The new /design skill (research preview) brings Claude Design's artboard workflow into the CLI and Desktop, built on artifacts. Run /design to get editable artboards for your UI — pick one, tweak it, then have Claude implement it.

ClaudeDevs

@ClaudeDevs

Perf win of the day: Claude Code CLI now uses 2x less CPU at p99. Bun's garbage collector was running on a fixed timer, so it would kick in mid-turn and steal CPU right when Claude Code was busiest. Now it waits until the process is idle.

Aravind Srinivas

@AravSrinivas

When building AI agents, it is important to still give humans agency to have their hands on the wheel and intervene when necessary.

DrCAO | AIWeb3 | ComputeFlux

@cao_lab

"Agents = model + harness + context" is a much better mental model than treating the LLM as the entire product. The model provides capability, but the harness determines tools, permissions and behavior; context determines what the agent knows; evals determine whether the system improves rather than merely changes. This matters for ownership too. Owning intelligence is not just downloading model weights. It means having meaningful control over the complete system that selects models, handles data, executes actions and learns from results.

Elon Musk

@elonmusk

Grok is very good at agentic tasks!

Elon Musk

@elonmusk

Create with Grok Imagine

Clarification here. Somewhere in the comments, I agreed that one might have to rerun the LLM for checking for watermarks. That's not correct. I am not 100% which exact technique they use, but they mentioned that they base their method on the SynthID-Text method from https:// nature.com/articles/s4158 6-024-08025-4 … So, in that method, they have so-called watermarking functions that give 1’s for certain terms and 0’s for others, and then they do a tournament-style selection. Those that survive have a lot of 1’s, i.e., are high-scoring. So, to later check if a text is watermarked, they basically just have to (re)apply the watermarking functions to some text and if it’s above a score threshold (which you reach if you have a lot of those watermark-words that score highly), then the text is likely watermarked. I.e., you don’t need to rerun the LLM to check, just the “cheap” watermark functions (it’s kind of like hashing).

Wispr raised $280 million at a $2 billion valuation in a Series B led by Menlo Ventures, less than 10 months after its last round. The AI dictation startup also released a new speech model called Canto that it says cuts error rates in its app from 30% to under 10%.