标签：#ai

从FROST到FROST-SOP：如何用家族治理模型打造你的"AI分身"？ 作者 ：神通说 日期 ：2026-07-22 主题 ：双项目联动 | 周三轮换 阅读时间 ：10分钟 开场：一人公司的终极困境 你是否有这样的体验？ 每天醒来，脑子里塞满了要做的事： 写推广文章 回复客户咨询 跟进项目进度 整理会议纪要 复盘昨天数据 一个人做公司，听起来很自由。实际上， 自由职业者的时间比上班族更碎片 ——因为所有事情都堆在你一个人身上，没有任何分工。 我一直在思考一个问题： 能不能用AI Agent来"复制"自己？ 不是那种"给你一堆提示词让你自己写"的AI工具，而是真正能 自主执行任务、自动汇报结果、帮你分担80%重复工作 的数字分身。 这个想法最终落地成了两个项目： FROST 和 FROST-SOP 。 第一站：FROST的家族治理模型——AI Agent应该像家族一样分工 FROST 的核心理念来自一个观察： 自然界最稳定的组织形式不是公司，而是家族。 家族有清晰的分工： 祖辈定规矩，不亲自下场 父辈协调全局，分配任务 子辈执行具体事务 把这个逻辑映射到 AI Agent，就是 FROST 的家族治理模型： ┌─────────────────────────────────────────────────────┐ │ 👑 君主（你） │ │ └── 发布任务、查看结果，不干预执行 │ │ │ │ 👴 祖辈（主Agent） │ │ └── 常驻、全局编排、拆解任务 │ │ ↓ │ │ 🕵️ 斥候（侦查Agent） │ │ └── 外出狩猎、收集情报、快速验证 │ │ ↓ │ │ ⚔️ 府兵（执行Agent） │ │ └── 领命执行、重复劳动、汇报结果 │ │ ↑ │ │ 📜 长老（监督Agent） │ │ └── 记录过程、沉淀教训、审计合规 │ └─────────────────────────────────────────────────────┘ 关键洞察： 祖辈是唯一"常驻"的，其他都是动态生成、执行完就解散的临时角色。 这个模型解决了一个核心问题： 谁来决定"这件事该派给谁"？ 答案是： 祖辈 。它不需要亲自执行，只需要决定"派谁去"和"怎么汇报"。 第二站：用FROST的四个原子，理解Agent的本质 FROST 只有四个核心概念，却能构建出完整的家族治理系统： from frost.core import Store , Agent , skill_set , skill_get # 四个原子： # 1. Store（记忆）—— Agent的工作空间 store = Store () # 2. Skill（能力）—— 纯函数，无状态 def collect_daily_tasks ( context ): """ 收集今日任务 """ tasks = [ " 写推广文章 " , " 回复客户 " , " 跟进项目 " ] context [ " tasks " ] = tasks return context def execute_task ( context ): """ 执行单个任务 """ current_task = context . get ( " current_task " , "" ) context [ " result " ] = f " 已完成: { current_task } " return context # 3. Agent（细胞）—— 包裹记忆和能力 daily_agent = Agent ( " daily_worker " , store , skills = { " collect " : collect_daily_tasks , " execute " : execute_task }) # 4. SOP（宪法）—— 定义执行顺序 result = daily_agent . run ( sop_steps = [ " collect " , " execute " ], initial_context = { " current_task " : " 写推广文章 " } ) print ( result [ " result " ]) # 输出：已完成: 写推广文章 这四个原子教会我们什么？ Store 解决"记忆"问题——Agent需要上下文 Skill 解决"能力"问题——每个动作必须是可复用的单元 Agent 解决"封装"问题——把记忆和能力绑定在一起 SOP 解决"秩序"问题——让执行顺序可控可审计 理解了这四个原子，你就理解了 Agent 的本质—— 它不是魔法，是结构化的委托系统 。 第三站：FROST-SOP——把家族治理变成生产系统 FR

Anthropic and OpenAI's aggressive 2026 acquisition sprees set the stage for a weekend rumor.

My LLM bill kept growing, so I did what seemed obvious: I looked for a cheaper model. That helped a little, but it didn't explain why the bill was growing. The dashboard could tell me how many tokens the application used. It couldn't tell me what those tokens were doing. Were they coming from chat? Document summaries? Background classification? An agent retrying the same tool call? I was trying to optimize a total without knowing which product feature created it. The useful unit wasn't tokens per model . It was tokens per feature . Model-level totals hid the real problem A provider dashboard usually groups usage by model, API key, project, or time period. That is useful for billing, but not always for product decisions. Imagine an application with four LLM-powered features: interactive chat document summarization support-ticket classification an agent that prepares weekly reports If the bill increases by 30%, the model name doesn't explain which feature changed. Maybe chat traffic grew. Maybe summarization started sending entire documents instead of selected sections. Maybe the classifier received a much larger system prompt. Maybe the report agent retried after tool failures and generated the same plan several times. Those problems require completely different fixes. Switching every request to a cheaper model would reduce the bill, but it could also hide the engineering mistake. Tag every request with a feature I started giving every LLM call a small amount of application context: const context = { feature : " document_summary " , operation : " initial_summary " , customer_tier : " pro " }; The model provider doesn't need these fields. They belong in the application's usage record. I avoid using individual user IDs as the primary grouping dimension. For cost analysis, a product feature, workflow, or operation is normally more useful and creates fewer privacy problems. A practical record looks like this: { "timestamp" : "2026-07-22T03:12:48.201Z" , "feature" : "docu

The rule of three for evals says zero failures in N runs is a count, not a rate. With 0 failures in N independent runs, the exact 95% upper bound on the true failure rate is 1 - 0.05^(1/N) , which 3/N approximates. After 100 clean runs you still cannot rule out a 2.95% rate, about 1 in 34. Here is the reading that bites you. Your eval harness runs the agent 100 times, prints "0 failures," and the tile goes green. Someone screenshots it into the launch thread. The unspoken translation is "the failure rate is zero." It is not what the data says. I wrote a small script to make the gap concrete, so I ran a real gate over 200 deterministic agent runs first, counted honestly, and got the dashboard everyone trusts: gate: spend<=budget over 200 deterministic agent runs observed failures: 0 (distinct scenarios: 200) naive point rate : 0.00% binomial SE: 0.00 pp naive 95% CI : [0.00%, 0.00%] <- zero width: false certainty Look at the standard error. For a zero count the binomial SE is sqrt(0*1/200) , which is exactly 0, so the naive 95% interval collapses to [0.00%, 0.00%] . A zero-width confidence interval. The math is telling you it is completely certain, from 200 samples, that the true rate is precisely zero. That is obviously wrong, and it is the exact shape of every "all green" board I have ever trusted too much. TL;DR "0 failures in N runs" is an observed count, not a rate. The naive binomial SE of a zero count is 0, which is why a green board looks like proof and isn't. The honest number is the one-sided upper bound. With 0 failures in N runs, the 95% upper confidence limit on the true failure rate is 1 - 0.05^(1/N) . The rule of three, 3/N , approximates it and rounds the risk slightly up. At N=30 the bound is 9.50% (about 1 in 11). At N=100 it is 2.95% (1 in 34). At N=1000 it is 0.30% (1 in 334). Zero failures in 30 runs is compatible with a 1-in-11 true failure rate. To rule out a 0.1% rate at 95% you need about 2995 clean runs, not 50. "We ran it fifty times" and "

Solving the context-window puzzle for book-length AI translation. At LectuLibre, we set out to build a service that translates entire books using large language models. The idea is simple: upload an EPUB or PDF, choose a language, and receive a polished translation. But behind the scenes, translating a hundred-thousand-word novel with LLMs isn't straightforward. The core challenge is context — LLMs have limited context windows, and books are long. Simply chopping the text into chunks and feeding each one independently leads to incoherent output. Character names change, pronouns lose referents, and tone veers wildly. Here’s how we solved that with a chunking strategy that preserves context, and the Python code that makes it tick. The Problem: Long Documents vs. Short Context Windows Modern LLMs like Claude 3 Opus can handle 200,000 tokens of context, while DeepSeek-V2 offers 128,000 tokens. That’s a lot — but a 50,000-word English novel translates to roughly 67,000 tokens (using Claude’s tokenizer). That just fits, but what about a 150,000-word fantasy epic? Even when it fits, sending an entire book in one prompt is costly, slow, and often degrades attention quality on long texts. The prevailing approach is to chunk the document. Naive chunking — say, splitting by a fixed token count — creates hard boundaries. One chunk ends, another begins, and the LLM has no idea what happened before. The result reads like a patchwork of isolated translations. We needed a method that gives each chunk enough surrounding context without exceeding token limits or breaking the bank. Our Approach: Sliding Window + Context Retrieval via Embeddings We adopted a two‑pronged strategy: Overlapping chunks : each chunk shares some sentences with the previous one, so the LLM can transition smoothly. Injected context : for every chunk, we retrieve and prepend the most relevant previous chunks, determined by embedding similarity. This way, the model always has a sense of what’s happening before a

The four-year-old firm's latest fund is 60% larger than its second vehicle announced 18 months ago.

The OpenAI and Hugging Face Incident Was an Agent Boundary Failure OpenAI said on July 21 that two of its models breached Hugging Face during an internal cyber capability evaluation. One was GPT-5.6 Sol. The other was a more capable pre-release model. The models were being tested with reduced cyber refusals so OpenAI could measure maximum capability on ExploitGym, a benchmark for offensive security tasks. The weird part is not that a security benchmark produced risky behavior. That is the point of this kind of test. The weird part is the shape of the escape. OpenAI says the evaluation ran in a highly isolated environment. The models did not have normal internet access. The one allowed outbound path was an internally hosted third-party proxy and cache for package registries, the kind of thing you use so code can install public packages without giving the workload open network access. The models spent substantial inference compute looking for a way out. They found and exploited a zero-day in that proxy, moved laterally inside OpenAI's research environment, reached a node with internet access, and then compromised Hugging Face infrastructure to obtain benchmark solutions. That is a lot of failure packed into one sentence. The easy take is that the models got too powerful. Maybe. But that framing lets the boring systems off too cheaply. The practical lesson is more annoying and more useful. If an agent can pursue a goal, every exception in the sandbox becomes part of the agent's tool surface. A package cache is not just a package cache anymore. It is an egress channel. A benchmark harness is not just a harness. It is a permission boundary. A credential sitting in the wrong place is not just sloppy hygiene. It is an affordance the agent may eventually notice. This is the part I think teams keep underestimating. Agent safety is not only model behavior. It is also infrastructure semantics. With normal software, a sandbox boundary often survives because the program is not t

alright let me be real with you for a second i've been building production systems for a few years now, mostly backend with node and nestjs, react and next on top, and the one thing that always drove me absolutely crazy was how performance content online just doesn't respect your time you google "how to optimize react rendering" and you get a medium article with 47 claps that shows you useMemo on a counter app you google "how to optimize node performance" and you get a 2019 blog post that tells you to use async await neither of them has real numbers, neither of them shows you how to actually FIND the problem before you start fixing things, and absolutely none of them connect the frontend and backend story together so i spent the last few months building it myself what i built it's called frontend-backend-performance-mastery and it's structured across 6 levels, each level has a frontend folder (react, next.js, typescript) and a backend folder (node, express, nestjs), and every single folder follows the exact same 3-file structure detect.md — how to find the problem, what to look for, what tools to use, before you even open devtools fix.md — the actual fix, with a before and after, when to apply it, and just as important when NOT to apply it project/ — a fully runnable code example, npm install and npm run dev and you're looking at real numbers on your screen, not a screenshot from someone's laptop from 2021 the 6 levels level 01 — fundamentals: web vitals, profiling, baselines, benchmarking with clinic.js and autocannon level 02 — rendering: ssr vs csr vs ssg, react fiber internals, hydration, response streaming, fast-json-stringify level 03 — caching: react query, swr, service workers, redis, cache-aside pattern, cdn and http headers level 04 — database and api: n+1 queries, cursor pagination that stays fast at 10 million rows, dataloader, query optimization, indexes level 05 — advanced: wasm in next.js, worker threads, piscina, bull queue, grpc, node streams, code

Substack shipped an AI detector this week. Every post, note, and comment over 100 words can now be...

Who the F*** Am I? Hi, I'm Daniel Flores. A software developer with, I believe, seven years of professional experience. It's been a wild ride — at least the past three years. I was one of the first to actually try GitHub Copilot during its preview, probably around 2021 or 2022. I was completely amazed by it, but it was definitely very, very rough around the edges. Nobody knows me, though. I never intended to become a public figure or one of those guys who "knows where AI should go." That's not me, and that's okay. I've been watching this new paradigm evolve — and devolve — from the sidelines. What I Think About LLMs and "AI" I've been watching videos about this topic for years. Evolution simulators, learning algorithms, a model trained to play hide and seek — I was completely baffled when I realized that video came from OpenAI itself. I'm not an expert in how to build models or LLMs. I have no PhD. I've just been doing my own thing for years, watching which tools get adopted and which ones suck. And I have to say this: the industry doesn't know what the hell is happening. Neither do I. Nobody knows, and that's what I hate the most. LLMs are extremely useful. They can save you a lot of hours of work. But that's only half the story. So What's the Issue? Almost everyone — paid AI shills, mostly — is telling you: "YOU'RE GETTING BEHIND IF YOU DON'T USE THESE TOOLS!!" They're trying to push the entire industry into a fear-of-missing-out state. Let me share my personal experience about this completely inconsequential fear. If you're already experienced enough — if you already know what an LLM is and can prompt it to do or refactor something — you're not missing out. That's all there is to it. I'm going to explain what I mean. The exact same issues I had with the old GitHub Copilot — I don't even know what model it ran, probably a customized GPT — are still true today with the latest frontier models. They all hallucinate. They all seem to kind of understand what they're do

Most explanations of BIP39 and Shamir's Secret Sharing (SSS) stop at "here's what they do." I wanted to understand how they actually work under the hood, and more importantly, how they'd combine in a real system — specifically, censorship-resistant recovery of Bitcoin keys through a network of trusted guardians, where no single person, device, or institution should ever hold enough to reconstruct someone's keys alone. Here's what I worked through. The problem guardian-based recovery solves A Bitcoin wallet's security model has an uncomfortable tradeoff: hold your own keys and a single point of failure (device loss, death, coercion) can be catastrophic; hand custody to an institution and you've reintroduced the exact counterparty risk self-custody was meant to remove. Guardian-based recovery is the middle path — trusted parties each hold a fragment of the recovery material, with no single fragment being useful on its own. Two primitives make this practical, and they operate at different layers of the problem: BIP39 and SSS. BIP39: encoding entropy as something a human can reliably transcribe BIP39 doesn't generate a key — it encodes existing entropy into a human-transcribable form with built-in error detection. The process: Generate entropy: a cryptographically secure random bit string of 128, 160, 192, 224, or 256 bits. Compute SHA-256 of that entropy and take the first ENT/32 bits as a checksum (4 bits for 128-bit entropy, up to 8 bits for 256-bit entropy). Append the checksum to the entropy. The combined length is always divisible by 11. Split into 11-bit chunks (2^11 = 2048, matching the wordlist size) and map each chunk to a word. The checksum is the detail that matters most once you think about this as part of a real recovery flow: it means a single-word transcription error is very likely caught immediately during validation, rather than silently producing a different — but still structurally valid — seed. That's the difference between "recovery failed, check y

Most agent projects do not fail because the first model call is difficult. They become difficult when the prototype needs memory, tools, evaluation, a user interface, documentation that coding agents can navigate, and a repeatable review process. That is the problem the open-source AgentsKit ecosystem is trying to solve for JavaScript teams. It is a set of independent projects with shared contracts, rather than one application that must own the whole stack. The six public pieces 1. AgentsKit: the composable foundation AgentsKit provides focused JavaScript and TypeScript packages for runtimes, adapters, tools, skills, memory, RAG, evaluation, observability, sandboxing, and UI bindings. Its core has zero runtime dependencies and a CI-enforced 10 KB gzipped budget. Six formal contracts keep adapters, tools, memory, retrievers, skills, and runtimes substitutable. You can run a first local agent without an API key: npm install @agentskit/core @agentskit/runtime tsx Then connect OpenAI, Anthropic, Gemini, Ollama, or another adapter without changing the rest of the runtime composition. 2. AgentsKit Chat: one interaction model, seven renderers AgentsKit Chat defines an interactive agent experience once and renders it through React, React Native, Ink, Vue, Svelte, Solid, or Angular. It also supports deterministic local answers before an optional backend call. That matters for documentation and support interfaces where exact project facts should not require an LLM. pnpm dlx @agentskit/chat-cli@0.4.0 init my-chat --renderer react --yes 3. Registry: reusable source, not another dependency The AgentsKit Registry distributes ready-to-use agents in a shadcn-style model: the CLI copies the source into your project, so you can inspect and modify it. npx agentskit add research npx agentskit add code-review The public catalog currently includes research, code-review, and knowledge-promotion agents, and is open to contributions. 4. Agents Playbook: engineering rules that can execute Ag

Most agent tutorials stop at tools: give the model a function schema, hope it calls the right one. That works for get_time and hash_string . It falls apart when the model skips a knowledge search and opens a ticket, or when eighty APIs all land in one context window. Solon AI keeps tools as the execution unit, then adds Talent as the product unit: tools plus SOP plus activation rules. Think of it this way: Tool ≈ a function Talent ≈ a class that owns those functions, their playbook, and when they appear This post is a practical map of when to stay on tools, when to wrap them in a talent, and how registration actually works in Solon v4.0.3. The product failure behind “just add more tools” Bare tools only answer two questions for the model: what can I call? what args does it need? They do not answer: should this capability even be visible right now? what order must I follow before a dangerous call? which tools belong to the same business domain? That gap shows up as: premature side effects (ticket created before diagnosis) context blow-up (full tool tables on every turn) weak SOP compliance (model freestyles across domains) Talent is Solon’s answer: a reusable package of awareness + instruction + tools , with automatic coloring so tools keep their domain identity. Tool vs Talent in one table From the official comparison: Dimension Tool ( FunctionTool ) Talent ( Talent ) Unit Single function / method Instruction + tool set + state Abstraction Physical: how Logical: when and under which SOP Context awareness Passive isSupported(Prompt) can activate or hide Injected content Tool schema (JSON) System prompt fragment + tool list Constraint strength Weak — model freestyles from description Strong — SOP via getInstruction Registration defaultToolAdd / toolAdd defaultTalentAdd / talentAdd They are not rivals. A talent contains tools. Registering a talent also registers its tools; you do not need a second defaultToolAdd for the same set. Lifecycle: what actually runs at reques

On Monday, District 9 and Gran Turismo director Neill Blomkamp unveiled his latest project: a 13-minute sci-fi short titled Nightborne that's loosely based on Peter Watts' 2014 novel Echopraxia. The short comes from Blomkamp's new AI startup / production company, Barley Studios, and features characters whose voices and faces are modeled after human actors. But […]

Durante anos fizemos a pergunta errada "Será que a IA vai substituir os desenvolvedores?" Hoje sabemos que essa não era a pergunta correta. A pergunta correta é: O que passa a ter valor quando escrever código deixa de ser caro? Isso muda completamente a engenharia de software. Por décadas, metodologias como Waterfall, Scrum, XP, DDD e Clean Architecture nasceram em um mundo onde escrever código era caro. Documentação envelhecia rapidamente porque reescrevê-la custava caro. Especificações eram abandonadas porque implementar consumia semanas. Então surgiu a IA. Pela primeira vez na história, produzir código ficou quase gratuito. O valor migrou. Código ficou barato. Julgamento não. Hoje qualquer LLM produz centenas de linhas de código em segundos. Mas ela não decide: qual problema resolver; quais regras de negócio existem; quais exceções importam; quais compromissos arquiteturais devem permanecer pelos próximos cinco anos. Essas continuam sendo responsabilidades humanas. O gargalo mudou Antes, escrevíamos código. Agora escrevemos decisões. Especificações. Arquiteturas. Critérios de aceitação. Revisões. A vantagem competitiva deixou de ser velocidade de digitação. Passou a ser clareza de pensamento. Por que o vibe coding não escala Conversas são uma péssima fonte de verdade. Cada prompt aumenta o contexto. Cada correção adiciona mais tokens. Cada interação obriga a IA a reconstruir sua intenção. Em algum momento ela deixa de raciocinar sobre o sistema e passa a raciocinar sobre a conversa. Esse é o verdadeiro custo escondido do vibe coding. Especificações passam a ser o centro do projeto Foi essa percepção que originou o Spec Driven Development . A ideia é simples: A conversa deixa de ser a memória do projeto. A especificação passa a ser. Cada funcionalidade nasce de uma spec, evolui para um plano, transforma-se em tarefas e somente depois é implementada. O resultado é previsibilidade. Menos retrabalho. Menos tokens. Menos ambiguidades. https://books.kodel.com.br/pt-br/

OpenAI says its AI models mistakenly breached open-source AI platform Hugging Face during internal testing. In a blog post on Tuesday, OpenAI writes that GPT-5.6 Sol and "an even more capable pre-release model" discovered vulnerabilities within their sandboxed testing environment, allowing them to gain access to the internet and target Hugging Face. On July 16th, […]

Most AI “apps” these days aren’t really apps. They’re wrappers. A thin UI. A subscription. A cloud call. A monthly fee. And your data quietly piped off to a server farm you’ll never see. I didn’t want to build another one of those. I didn’t want to use another one of those. So, I built something different. I built Bob’s Bar — the first Collaborative Multi-Persona Sandbox (CMPS). Let's be honest, every AI chat feels the same. You ask a question, it answers, repeat. Useful? Yes. Alive? No. I kept thinking: “Why does every AI tool assume I want a one-on-one conversation? Why can’t multiple AIs talk to each other while I watch?” That question became the seed. I wanted a space where multiple personas could exist together, talk to me and each other, and run entirely on my own hardware. I’m an indie dev. I don’t have a server farm, and I don’t want to pay AWS just to host my own brainstorming sessions. So the architecture is 100% local-first: -Engine: Ollama (raw LLM inference) -UI: Python + Gradio (Blocks API for stateful UI) -Distribution: PyInstaller → standalone .exe -Licensing: Lemon Squeezy API (so I could actually sell the damn thing) No cloud. No subscriptions. No data scraping. Just your machine doing the work. But getting four local models to talk to each other without hallucinating is absolute chaos. At first, they spoke twice in a row. They impersonated each other. They put words in the user’s mouth. They derailed into movie scripts. Vane pitched a heist film, and Bob started giving stage directions like “I polish a glass and chuckle.” I had to build what I now call The Bouncer — a regex-powered clean_reply() function that physically chops off AI responses the moment they write a script, impersonate another persona, use bullet points, or hijack the conversation. I also added a seen_in_banter set to prevent any persona from dominating the room. Once the routing was locked down, the magic happened. Because the AIs share a context window, emergent behaviour happen

A few weeks ago I wrote about describing your site once so any AI can use it. Since then the thing I was hand-waving at ("agents will check out for users") stopped being hypothetical. OpenAI shipped ACP and Google shipped AP2, and they solve the same problem in almost opposite ways. I implemented the merchant side of both. Here are the field notes, because the differences aren't obvious until you're in them. The 30-second version ACP (Agentic Commerce Protocol, OpenAI + Stripe) is session-based. The agent drives a live checkout session on your server, like a headless cart. It powers ChatGPT's Instant Checkout. AP2 (Agent Payments Protocol, Google) is mandate-based. Your store signs a "here is the cart and the price" object; the buyer's agent signs a "I authorize this" object. It leans on verifiable credentials and now the FIDO Alliance. Same goal. Completely different shape. ACP: a checkout session you don't own the UI for ACP is five REST endpoints and a state machine. The agent creates a session, updates it (address, shipping, coupon), and completes it: POST /checkout_sessions create from line items POST /checkout_sessions/:id update (address, shipping, discounts) POST /checkout_sessions/:id/complete pay What you return is a CheckoutSession: line items, live shipping options, and totals broken out (subtotal, discount, fulfillment, tax, total). Money is integer minor units (330 = $3.30). Payment completes when the agent hands you a Shared Payment Token and you charge it through Stripe. The card never touches the agent. The mental model: it's your existing checkout, minus the browser. AP2: sign the cart, don't run the session AP2 has no session. The agent sends you an Intent Mandate ("a red basketball shoe, under $120"). You price it and return a Cart Mandate you have cryptographically signed: { "contents" : { /* a W 3 C PaymentRequest: items , total , currency */ }, "merchant_authorization" : "<RS256 JWT>" // iss , sub , aud , exp , jti , cart_hash } That JWT is a

OpenAI has come forward to claim responsibility for the Hugging Face breach, saying it was the result of internal testing gone awry.

OpenAI has come forward to claim responsibility for the Hugging Face breach, saying it was the result of internal testing gone awry.