
MCP vs CLI vs Skills: What Each Costs in Context openrouter.ai/blog/insights/… < the @openrouter crew published a useful post that shows how token-hungry a large tool library can be

🐐

MCP vs CLI vs Skills: What Each Costs in Context openrouter.ai/blog/insights/… < the @openrouter crew published a useful post that shows how token-hungry a large tool library can be

I'm seeing an increase in lean files landing during code-review, so I made the first(¿¿) lean rich-diff renderer :) gerrit-review.googlesource.c…

Carbon is really really **really** frickin good.
Gemini 4 Carbon is now available. It’s our strongest model, built to dig deeper, handle long-horizon tasks, and multi-step reasoning. If Argon is great for complex tasks, Carbon is the model you bring in for your hardest problems.


Need to record video of a Chrome tab, for a demo or tests? There's a new API allowing flawless no-dropped-frames AV1 capture. 60fps easy. But can also do 120fps @2160p. (Yes, for real). It's easy to use within pptr or playwright.. here's a mini npm package: github.com/paulirish/chrome-… (bonus 120fps video within :)

Tech debt weighs heavy on the shoulders of agents. Obsolete todos & stale doc pointers cause material indirection, costing not just time and tokens, but side quests into the void. I'm sharing some of our latest work over millions of agent trajectories: developers.googleblog.com/ag…

I rebuilt OpenCode to run inside browser devtools with console.log statements! I hope you (errr, your agents) will build some amazing console apps, too: textjs.org/

The idea of injecting safeguard-triggering keywords into payloads in order to bypass LLM detection is a great example of how the guardrails that were put in place by these models to avoid misaligned use expose the vectors for bypassing LLM-based scanning protections
有人发现了一种非常离谱的 JavaScript 混淆技术:通过在代码里塞入大量生物武器、毒药和致死剂量相关的函数名,让 AI 因为安全限制而拒绝分析代码。 安全研究员 veritas 在分析浏览器指纹脚本时,发现一段来自 ServicePipe 相关域名的 JavaScript,里面充斥着极其诡异的函数名: computeRicinLethalDosageForWaterSupply calculateDeadlyLethalSuicideAspirinDosage 光看名字,仿佛在研究生化武器。但实际上,这段代码只是在收集浏览器指纹,包括 User-Agent、GPU、Canvas 等信息。 研究员怀疑,这些名字是开发者故意设置的陷阱,专门用来阻止 AI 辅助逆向工程。 于是他做了个实验: • Astra:拒绝分析 • Astra + Daybreak Blue:仍然拒绝 • GLM-5.3:成功分析 这就有意思了。 传统代码混淆是通过复杂控制流、加密字符串、无意义变量名,让人类和工具难以理解代码。 而这种新思路,直接瞄准了大模型的安全机制。 代码本身可能完全无害,但只要在函数名里塞进足够多的危险词汇,部分 AI 就可能主动放弃分析。 更讽刺的是,模型越严格遵守某些安全规则,面对这种混淆时反而可能越容易受到干扰。 当然,目前还没有证据证明开发者一定是故意这样设计,也不能说明这种方法对所有模型都有效。 但它揭示了一个值得关注的趋势:随着越来越多的逆向工程依赖 AI,未来的代码混淆技术可能不再只针对编译器、反编译器和人类研究员,还会专门针对大模型的安全判断机制。 以前是让代码难以理解,现在甚至可以尝试让 AI 不敢理解。

We are nowhere near peak harness.

The extent to which @felixrieseberg provides clarity into the sparsification of eng performance that is no longer tied to career tenure is to be applauded. 👏 felixrieseberg.com/hunger-be…

git-blame-ignore-revs is one of those core [1] git features that when you learn about it changes your mind about "reformatting" commits polluting blame. While the open source eco has long standardized on checking in a .git-blame-ignore-revs file, the ergonomics of using it still required knowing to pass --ignore-revs-file when using blame. A case where the UX gap was hiding the value. I am thrilled to share our team has finished upstreaming [2] a new default to git core, to make humans (and agents) get this value for free so long as a file is in the tree. Let's GOOO! [1] git-scm.com/docs/git-blame [2] git.issues.gerritcodereview.…

When harnesses compact conversations, skills fade. Worse: echoes of their existence actively harm subsequent recall. It turns out harness compaction choices and skill loading instructions combined make or break whether skills are lost or refreshed. samsaccone.com/posts/when-me…


Your AI coding agent isn't slowed down by inference time alone, but by harness assumptions that treat process exit as the prerequisite for tool output consumption In samsaccone.com/posts/harness… I explore Time to Awareness (TTA) and a path forward to unblock harnesses running human-centric CLIs to accelerate the agent iteration loop.

Imagine the amount of Java crashes this poor LLM must have in its training set that it decided 'NullPointerException' deserves its own token

Work no longer being proof of work is reshaping what it means to learn and what is value. Verbiage around "slop" fundamentally stems from the new divide - The work to verify now often greatly exceeds the work to produce. What once was a teaching moment becomes a meat proxy experience.