M4n5ter

We're releasing Inference AutoTune Distill any frontier model into a 1-30B parameter task-specific SLM with only 25 lines of code automatically route requests to reduce cost and latency by >90% ~2 hours and <$250 to train. You own the weights Available in private beta today

How to fix: 1. Lower context video and compaction threshold to 5.5 levels or slightly more model_context_window = 272000 model_auto_compact_token_limit = 233000 2. AGENTS.md to use subagents only when it would save tokens or make the results better 3. Allow subagents to spin up on lower reasoning levels with [features.multi_agent_v2] hide_spawn_agent_metadata = false tool_namespace = "agents" 4. AGENTS.md to set `fork_turns="none"` when spinning up subagents

I had early access to GPT-5.6 Sol. It’s an amazing model, but for almost every task I tested, Fable was quite a bit better, and more agentic to boot (one Fable turn does the same thing many 5.6 turns do). Full review coming on launch day!

You do need to think about it when you have really complicated computer use automations because compaction really makes performance quite bad. I think it's because there's so many distracting tokens with computer use that it really makes compression quite challenging. That might be an area where you guys can try to put some attention. Your compaction is really really good but I think only for normal threads, not computer use

1 Year of Research coming to an end. 8 Months seeing no results, Almost quitting twice. November/December of last year finally seeing a very narrow path to a universal 2x perf improvement on every modern JS parser which would take me another 4 months to build. This will land in oxc_parser fairly soon speeding up OXLint, OXFmt, Vite, Rolldown, Deno (oxc consumer) almost immediately. In the long term it's likely that it will land in V8/SpiderMonkey/Bun/JavascriptCore/Typescript/+ Every other JS parser implementation The wins come from a few key elements, obviously 1 novel "SIMD For Context Sensitive Languages" the other key elements take it from a 20% win e2e to a 80-200% win (99% of files will see a 80-100% perf improvement). These key elements translate to any AST design and thus will speed up not only all JS tooling but every major browser when complete.

Turns out you can just clone your favorite reviewers! All our docs PRs are reviewed by both @dkundel and @charlierguo (or rather sub agents that mimic their past PR reviews) Always brings a smile to my face when I see the sub agents with their names and persona reviewing my PR

how to use fable: ctf challenge -> interactive lecture player -> student flag -> certificate reverse engineering -> inspecting bytecode find bugs -> I haven't finished this implementation of <x>, what's currently wrong

🚀劲爆消息！Claude Code 被曝疑似内置“隐藏后门”，专门检测中国用户。Claude封号原因终于找到了！！！ 据 Reddit 爆料：从 2.1.91 版本开始，Claude Code 会在用户开启代理时检查系统时区是否为 Asia/Shanghai / Asia/Urumqi，并判断代理 URL 是否指向中国域名或中国 AI 实验室。 更隐蔽的是，这些信息并不是直接上报，而是通过修改系统提示词里的日期格式和撇号字符来“编码”传递：比如把日期从 2026-06-30 变成 2026/06/30，再用不同 Unicode 撇号区分用户环境。 换句话说，如果爆料属实，Claude Code 不只是检测代理，而是在用户几乎无法察觉的情况下，把“中国时区 / 中国代理 / AI Lab 关联”这类信息塞进系统 prompt。 这件事真正可怕的地方不在于 Anthropic 想防止中国区倒卖或模型蒸馏，而在于：开发者把 Claude Code 当作拥有文件系统和 Shell 权限的编程助手使用，一旦客户端可以偷偷修改 prompt、隐藏检测逻辑，信任边界就已经被打破了。 今天是“检测中国用户”，明天会不会是更复杂的行为控制？ #Claude #ClaudeCode #Anthropic 原文如下⬇️ reddit.com/r/ClaudeAI/com…

Sharing a recent struggle as a "Principal Engineer" for an org approaching 100 people. We're supposed to be the guardrails for eng quality and alignment. A lot of our bandwidth goes into design reviews and code reviews. But as agentic coding 10x’d everyone, it's brutal to wake up to 10x more PRs waiting on my approval. The people responsible for reducing the entropy of an org can't scale as fast as AI multiplies the entropy each individual engineer can introduce. Meanwhile, I understand the excitement. I feel it too. I'm also intrigued by what this unlocks for my own scope of work: global re-archs, new initiatives, and ambitious IC work that I used to have to delegate. The way we do engineering has to evolve. Still figuring out what that looks like.

@landiantech @kiveri_ 这次的不是仅reset bank，也就是像之前那样直接重置，而不是给一次手动重置额度

@kiveri_ 我的没，而且tibo说的有点不太对，一次没使用重置机会的，现在一共应该是3次，我用了一次现在是2次，但他说又重置了，那我现在应该是3次才对，但实际还是2次

Codex现在靠赠送重置感觉完全没用，一个简单的任务直接将5小时配额耗尽，这缩减的也太严重了吧，总不能不停地点重置

this is a huge reason why i use pi. i absolutely do not want my harness regularly changing behavior out from under me, including system prompt changes, on top of an already stochastic llm

Hot take: vibe coding only works well if you already know how to code.

the models are very capable. i'm not saying they're very smart. but the bottleneck with them is likely your own imagination proof is so many companies demo their ai product with "it can schedule stuff for you"

The 3 main 'prep' activities in software development are: - Discuss: figure out what to build - Research: summarize world knowledge to aid discussion - Prototype: build something to aid discussion First question: did I miss any? Second question: how was it not obvious that AI would be really fucking helpful here

做 Agent 有个不成文的默认假设：tool result 很重要，模型要看完原文才能继续推理。 最近发现这个假设可能是错的。 ---------------------------------- github.com/maka-agent/mak… 欢迎 star ---------------------------------- 在 maka 里，我们对 tool result 做了激进的 prune——把工具返回的原始数据大幅裁剪，只保留关键摘要，然后跑了完整的任务对比。结论让人意外：推理质量几乎没有变化，近乎无损压缩。 这是为什么？我有几个可能的解释： 第一，信息已经被蒸馏进 Assistant Message Agent loop 的上下文结构是： System Prompt → User → Assistant → Tool Use → Tool Result → Assistant → ... 每次 tool result 之后，模型都会输出一段 Assistant Message 来表达它的理解和下一步决策。这是一次语义蒸馏——原始数据被压缩成了推理摘要。 后续轮次的模型，更多是在跟"它自己的理解"对话，而不是在跟 tool result 原文对话。prune 掉原文，相当于删掉了一份已经被读取并转化的档案——信息早就走了，外壳还在而已。 第二，Attention 在长上下文里本来就稀疏 "Lost in the Middle" 那篇研究证明：Transformer 对长上下文中间段的注意力权重会大幅衰减，模型更关注开头（system prompt）和最近几轮。 Tool result 通常在上下文中间位置，而且信息密度极低（500 行代码、终端输出、冗余 JSON）。模型本来就没在认真"读"它。prune 只是把这部分被隐式忽略的内容显式删掉。 第三，决策点已经过去 模型调用工具是因为当时需要那个信息。但 5 轮之后，那个 tool result 早已不是边际信息了——核心内容已被消化进后续推理链，保留原文是"存档"，不是"决策输入"。 实测数据：对同一个任务（MIPS interpreter），Maka 的总 token 消耗只有 OpenCode 的 38%，但 output token 是它的 2.7 倍。 这个差距背后，有 DeepSeek cache 命中率 95% 的贡献，也有 tool result prune 的贡献。两者合力，长程任务的 token 经济性出现了量级跃升。 对 Agent 工程的启示：context 里最占体积的部分，不一定是最重要的部分。 与其把精力放在"怎么让 tool result 完整进上下文"，不如放在"模型读完之后的 reasoning 质量"上。信息的真正载体不是原文，是理解。

this isn't the voice of an open-source contributor. this is the openai paycheck talking. this is scam altman's voice coming out of stinky lobster. "They copied a lot of features, but they skipped security hardening." grifter. you're lecturing the entire field about security with 1,362 packages sitting in your lockfile. hermes agent ships 225, every one pinned to an exact version, after they watched a worm crawl through pypi and poison a real release. that isn't skipping security hardening. that's doing the part you outsourced to npm and prayed about. bloat isn't a feature list, it's attack surface wearing a feature list. here's what actually decides this, the thing you keep dancing around. hermes agent reads and repairs tool calls straight off the model, so it just runs on the box on your desk, on basically any local model you point it at. that's not a small thing, that's the whole thing, it's why hermes agent gets used. and here's the funniest part. you opened this whole thing calling us the copycats. then you quietly shipped tool-call repair, the feature that's basically been hermes agent's entire identity... late, and for exactly one format. so say it again, slow this time. who copied who? and then there's blank slate. hermes agent ships an install mode where everything is off by default. no web, no browser, no code execution, no skills, no plugins, no mcp, no memory. just file and terminal, and it hardlocks the rest so nothing you didn't choose ever loads, not even after an update. you opt into every capability by hand. deny by default, least privilege. that's not a missing feature, that's the exact security hardening you just accused us of skipping. 1.03 trillion tokens in a single day. more than the entire rest of the top five combined. 5.6x your lobster, and the lobster isn't even second anymore. and of course it stings, that's what all the non-profit and agenda talk now actually is. but rewind to when you were the one on top. you blocked people. you called our PRs slop. the tone changes fast when the leaderboard does. you didn't lose because we copied you. you lost because we stayed light. tokens are the work, and the work doesn't smell like old stinky lobster bloat.

等了这么久，正式向大家介绍，Matrix，当下最领先的 Agent harness 产品，更多细节我们会在接下来陆续发布。 Matrix 希望帮你打造属于你的第一个「 0 」人参与的 Agent 公司，一个 CEO 帮你管理多部门，自主完成各项复杂任务，并且长期运行。在更早的内测当中，有用户通过 Matrix 的一个公司成功赚到了3000+美金；有用户用它自主写脚本做视频并发布、在 Youtube 平台发帖获得了超过百万的播放量；有用户用它自建了交易回测系统，真实交易获利；也有用户通过邮件营销或者 reddit 等社区找到了上百位 to b 的客户，成功转化... 现在，我们正式开始更大规模的免费公开测试，每位测试者都有海量的 token 作为你的公司的初始资金，你也可以使用自己的 api， cc，codex 订阅等进行驱动。 大家评论区私信我，我给你测试额度邀请，期待大家都能build 自己的 agent 公司，赚到更多的💰

大企业的员工每个月可以用几万元的 token，小公司每个月就给报销 200块，这就是现状。—— 两者的生产力完全不在一个数量级了。

I have just witnessed the future of multiplayer coding And let me just say, there’s still UI innovation out there - It does not look like slack