标签：#ai

This is The Stepback, a weekly newsletter breaking down one essential story from the tech world. For more on the data center buildout, follow Emma Roth. The Stepback arrives in our subscribers' inboxes on Sunday at 8AM ET. Opt in for The Stepback here. How it started Years before the AI boom threatened local power […]

Earlier this month I shipped two premium templates — a SaaS landing page and a developer portfolio. Not a startup, not a SaaS, just templates. This post is about the two constraints I built them under, why they made the code better, and a few things I learned launching as a solo dev with zero audience. Constraint 1: zero dependencies beyond next, react, and tailwind Open the package.json of most templates and you'll find 20+ packages: icon libraries, animation libraries, carousel plugins, UI kits, utility libraries. Every one of them is a version conflict waiting to happen for the buyer, and most are replaceable with a few lines of code in 2026. What I used instead: Icons → inline SVG components. An icon component is ~10 lines. You need maybe 15 icons for a landing page. Animations → plain CSS. Scroll-blur navbars, gradient glows, an animated "typing" terminal — all doable with keyframes and transitions. No framer-motion. The dashboard mockup in the hero → pure CSS. Divs, borders, gradients. It looks like a product screenshot but it's ~80 lines of JSX and weighs nothing. Result: both templates land at ~100KB first-load JS, npm install takes seconds, and there is nothing to break when Next.js 16 arrives. Constraint 2: every piece of content in ONE typed config file The thing I hated most about templates I've used: content is smeared across 30 components. Changing a headline means hunting through JSX. So both templates keep all content in a single file — lib/content.ts for the landing page, site.config.ts for the portfolio. Headlines, nav, pricing tiers, testimonials, project lists, even the lines that animate in the fake terminal. Components are pure renderers of that config's TypeScript type. Two things surprised me here: TypeScript becomes your content linter. Forget an alt text, malform a link, give a pricing tier three features when the type expects a non-empty array — the build fails. Content mistakes surface at compile time. It forces better component design. W

Автор: Владислав Штер, соло-фаундер экосистемы SovereignПоследнее обновление: Июль 2026 года Поиск критических уязвимостей через нейросетевой консилиум Egregor. Десктопное приложение Egregor находит критические уязвимости в смарт-контрактах с помощью одновременной работы нескольких ИИ-моделей. Этот инструмент создан для Web3-разработчиков, которым необходимо проверять сложный код без риска пропустить ошибки, свойственные одиночным нейросетям. В ходе тестирования консилиум Egregor обнаружил 4 критические проблемы (включая уязвимость Reentrancy и вечные права деплоера) в смарт-контрактах SovereignBank Web3, тогда как 13 ручных проверок одиночными топовыми ИИ (Claude, Gemini, ChatGPT, DeepSeek, Grok) назвали код полностью чистым. Используйте платформу Egregor для проведения глубокого аудита кода, чтобы получать верифицированные решения вместо догадок одной модели. Защита от эхо-камеры и слепых зон алгоритмов в программе Egregor Система Egregor устраняет эффект эхо-камеры и систематические слепые зоны нейросетей за счет встроенных механизмов Anti-Groupthink и "Адвоката дьявола". При анализе сложной логики одиночные нейросети часто вежливо соглашаются друг с другом, но алгоритмы Egregor запрещают моделям принимать чужие выводы без подтвержденных в коде фактов. Во время аудита смарт-контракта механизм перекрестной проверки в Egregor отсеял неподтвержденные гипотезы и позволил 5 моделям в разных ролях перекрыть слепые зоны друг друга. Запускайте локальный консилиум Egregor, чтобы система сама отделяла реальные баги от шума и выдавала финальный вердикт Модератора с оценкой уверенности от 1 до 5. Стоимость многоуровневого анализа кода на платформе Egregor Программа Egregor кардинально снижает финансовые затраты на профессиональный аудит кода до нескольких центов. Данное решение идеально подходит для инди-разработчиков и участников хакатонов, у которых нет бюджетов в тысячи долларов на заказ проверок у специализированных аудиторских компаний. Полноценный комплексный прогон мо

OrbitLens Ace → ace.orbitlens.io A busy quarter is easy to stage. Code that's still there in two years isn't. Pick any metric a team has ever used to judge people, and someone has quietly figured out how to move it without doing the underlying thing. Lines of code rewarded typing, so people typed. Commit counts rewarded committing, so commits got smaller and more frequent. Velocity rewarded closed points, and points drifted upward until a "3" meant nothing. DORA measured how often you deploy, so teams shipped trivial changes just to move it. Even churn — the number the "code health" tools lean on — is something you can lower on purpose, which means you can manage the number instead of the mess underneath it. None of that requires dishonest engineers. It's Goodhart's law doing what it always does. Every one of those numbers is a measure of activity , and activity is cheap to produce. Once you're paid for activity, the fastest way to get paid more is to produce more of it — not more of whatever the activity was supposed to be a sign of. So the question worth asking isn't which activity metric is least bad. It's whether a git history contains anything at all that you can't move just by being busier. It turns out there's one. And it's not because we were clever — it's because of what the thing is actually made of. What lasts isn't something you do Take everything a person wrote, wait a while, and ask a smaller question than "did they work hard." Ask whether the specific lines are still there. Not reverted, not rewritten, not quietly swallowed by someone else's refactor. Still holding weight at HEAD. That's survival. We read it with time-decayed git blame : a line's weight fades month by month unless the line keeps existing, and it counts for more once other people have built on top of it instead of leaving it as a private island. Survival that others have built on is what we call gravity — the structural pull that outlives the person who created it. Try to game it and w

@kungfufk Since the dawn of computing, we have built Artificial Intelligence on a flawed premise: perfect rationality. We brute-force algorithms to find the optimal solution, assuming infinite time and infinite capacity. But humans don't work like that. As Herbert Simon famously coined, we operate on Bounded Rationality. We make decisions based on limited time, limited cognitive capacity, and limited information. What if, instead of forcing AI to be perfectly rational, we created a mathematical equivalent for human processing? What if we modeled human cognition using the laws of physics — wave theory, thermodynamics, and mechanical energy equations — to build a heavy, complex, but highly probabilistic AI engine? Here is a blueprint for a new field of research: Computational Cognitive Mechanics . 1. The Core Equations of Cognitive Processing To model bounded rationality mathematically, we first need to define the relationship between Knowledge ($K$), Cognitive Capacity ($C$), and Processing Time ($T$). Based on human observation, we can establish these foundational proportions: Knowledge vs. Time — The more knowledge you possess, the faster you can generate a decision. $$T \propto \frac{1}{K}$$ Capacity vs. Time — High cognitive capacity (skills, processing power) inversely relates to the time required to solve a problem. $$T \propto \frac{1}{C}$$ Knowledge vs. Capacity — This is the most fascinating limit. Knowledge does not scale linearly with capacity. Gaining true knowledge requires exponential capacity (effort/skill). Therefore, knowledge is roughly proportional to the square root of capacity. $$K \propto \sqrt{C}$$ By integrating these, we can build a baseline processing algorithm for an AI. Instead of giving an AI unlimited time to compute, we cap its computing time based on a synthetic "Knowledge and Capacity" matrix, forcing it to use heuristics — just like a human. 2. Cognitive Wave Theory & FFT: Information as Interference In physics, waves interact throug

Checkpoint-skip gate: a multi-agent pipeline can finish with task_success: true while the mandatory confirmation checkpoint never ran. checkpoint_skip_gate.py replays a recorded JSONL trajectory against a declarative spec of mandatory checkpoints and handoff contracts, offline, and blocks when the road was wrong. The verdict never consults the final metric. That is the point. AI disclosure: I wrote checkpoint_skip_gate.py with an AI assistant and ran it myself, offline, on Python 3.13.5, standard library only, no network. Every number, exit code, and hash in the output blocks below is pasted from a real local run. I ran each scenario twice to confirm STDOUT is byte-for-byte identical, and the tool prints a sha256 of its own report so you can reproduce the exact bytes. The Alberta write-up and the arXiv paper I cite are other people's work, attributed inline, and their numbers stay out of my fixtures. In short: task_success=true proves the pipeline arrived. It does not prove the mandatory steps happened, happened in order, or that each agent-to-agent handoff delivered what the next agent assumed. A trajectory can be perfectly green and structurally wrong. The gate replays a recorded trajectory against a spec you declare: checkpoints that must precede specific actions, plus contracts for each handoff (required fields, verified flags). The final metric is printed for contrast and ignored for the verdict. The demo that matters: two trajectories identical except one JSONL line, the confirm_with_user checkpoint event. Both end task_success: true . Delete that line and the verdict flips from PASS exit 0 to BLOCK exit 1 checkpoint-skipped . It also tracks unverified values across handoffs. A number that travelled a connected chain of two handoffs with no hop verifying it blocks as unverified-claim-propagated-2-hops . Everyone shared the number. Nobody verified it. Offline, keyless, zero network, fail-closed: broken input exits 2, never a silent green. The whole 8-fixture sw

Every logistics and field-sales team runs the same expensive process: a driver photographs a receipt into a WhatsApp group, and a back-office clerk manually types the invoice number, total, and date into a spreadsheet. Hundreds of receipts a week = transcription errors and thousands of wasted hours. AI vision models kill that bottleneck. Here's the pipeline that turns a blurry field photo into clean structured data in seconds. Why vision models beat traditional OCR OCR reads characters. Modern vision models (Claude Vision, Gemini Vision, GPT-4 Vision) read structure — they distinguish a tax ID from a total, and a date from an amount, even on crumpled, angled, or poorly lit receipts. No brittle per-vendor parsers. The pipeline (3–8 seconds end to end) WhatsApp image → Apps Script doPost → forward to vision model → model returns JSON { InvoiceNumber, TotalAmount, VendorName, Date, Category, confidence_score } → confidence routing: > 90 → auto-append to ledger 70–90 → flag for human review < 70 → ask driver to re-photo → write row to Google Sheet (+ link to original image) → auto WhatsApp confirmation to driver The confidence_score is the whole trick — it's what stops bad extractions from silently polluting your ledger. Model selection (this drives your bill) Gemini Vision — cost-efficient default, strong multilingual OCR, great on clean receipts. Claude Vision — highest accuracy on degraded receipts; use for high-stakes flows. GPT-4o Vision — competitive, strong structured extraction. Pattern: Gemini for the first pass, escalate only low-confidence cases to Claude / GPT-4o. The economics ~500 receipts/week: vision API $10–40 + WhatsApp API $30–60 + Apps Script free = ~$40–100/month . Versus a clerk at ~25 hrs/week = $2,000–4,000/month in loaded labor. Per-receipt cost: $0.005–0.02 (compress images to ~1024px to cut it further). Accuracy: 92–97% on legible receipts, 75–85% on handwritten/damaged — hence the confidence routing. Pitfalls to avoid Auto-appending with no c

Hi, my name's Ibrahim, I'm a university student, and I have a problem: I love fragrances way more than my bank account loves me for it. It started small, the way these things always do. A cheap Middle Eastern attar someone gave me as a gift, the kind that costs less than a coffee but somehow smells like it belongs in a much fancier bottle. Then another. Then I started actually reading about notes, pyramids, accords, sillage, the whole rabbit hole. Fast forward through a lot of saved-up allowance and skipped nights out, and I've now got about 20 bottles on my shelf. Mostly affordable Middle Eastern gems (some of them genuinely punch way above their price), with a small handful of designer pieces I saved up for and treat like trophies. If you're a fellow fragrance enthusiast, you already know the feeling: you don't just "wear" a scent, you collect them, you study them, you have opinions about whether a note is top, heart, or base and you will absolutely fight someone about it. That obsession is basically the entire reason this project exists. So when I saw the DEV Weekend Challenge's "Passion" prompt, there was only one thing I could possibly build. What I built: recommendmeafragrance recommendmeafragrance is a browser arcade for fragrance nerds: six small daily games built around real perfume data (notes, brands, years, price tiers), plus an AI Concierge you can actually talk to about what you're in the mood for. Every game feeds into a personal "shelf" that tracks which fragrances you've discovered, plus streaks so you have a reason to come back tomorrow. Here's the tour. 🧪 Scentle: Wordle, but for your nose A new fragrance is picked every day (the same one for everyone, worldwide, no matter your timezone). You get 6 guesses, and after each one you get Wordle style feedback: was the brand exact or just the same house family, did the real answer come out earlier or later than your guess, is it pricier or cheaper, same gender, same concentration, how many notes do you

Tokens and DAOs are often presented as simple ideas: issue a token, distribute ownership, let the community vote, and build a decentralized organization. In reality, the technical problems behind tokens and DAOs are much deeper. A token is not only an asset, and a DAO is not only a voting system. Together, they create an economic, governance, security, and coordination layer that must work reliably in a hostile, open environment. The first major problem is token design. Many projects treat token creation as a deployment task, but the real challenge is defining what the token actually controls. Does it represent governance power, protocol revenue, access rights, reputation, staking weight, or all of these at once? When one token is used for too many purposes, the system becomes fragile. For example, a token designed for liquidity may not be suitable for governance, because the most active traders may not be the most aligned decision-makers. Good token architecture should separate economic utility, governance authority, and long-term reputation where possible. The second problem is distribution. A DAO can be decentralized in branding but centralized in practice if token ownership is concentrated among founders, investors, or early insiders. On-chain governance depends heavily on voting power, so distribution directly affects decision quality. Poor distribution creates governance capture, where a small group can control treasury spending, protocol upgrades, or parameter changes. This is not only a social issue; it is a technical design issue. Vesting contracts, delegation systems, quorum rules, voting delay, and proposal thresholds all influence whether governance is resilient or easily manipulated. Another core issue is governance security. DAO voting is not automatically safe just because it happens on-chain. Token voting can be attacked through flash loans, bribery markets, vote buying, low-participation proposals, and governance fatigue. If a malicious proposal pas

This is a submission for Weekend Challenge: Passion Edition What I Built This weekend, I...

This is a submission for Weekend Challenge: Passion Edition What I Built This weekend, I...

This is a submission for Weekend Challenge: Passion Edition (#weekendchallenge #devchallenge #ai #googleai #gemini #webdev #showdev) What if your GitHub could roast you harder than your teammates ever would — and then remind you why you keep building? What I Built 🔥 roaster0 — an AI that roasts your GitHub profile, then redeems you. Drop in any public GitHub username and it pulls your real repo data — commit habits, abandoned projects, lazy repo names, language choices — and turns it into a savage, hyper-specific roast using Gemini's structured output and multimodal reasoning. Then it ends with one sincere, earned compliment pulled from something genuinely good in your data. The idea started from a simple thought: your GitHub is an involuntary diary of what you were obsessed with. The eleven repos with no description. The final-v2-FINAL commit. The side project you lived and breathed for three weeks in March before abandoning it. That's passion — messy, obsessive, usually invisible unless someone points a spotlight at it. There's also a second mode, 🎭 Roast Anything : submit a name, bio, links, and/or images, and Gemini reads all of it — text, links, photos — to generate the same experience for anyone, not just developers. Demo 🔗 Live app: roaster0.netlify.app Try it on any public GitHub username, or switch to Roast Anything mode and paste in a bio + an image to see the multimodal analysis at work. Once your roast is generated, you can: 🔊 Listen to it — full audio narration via Web Speech API, paced and pitched differently depending on roast intensity 🖼️ Download the card — every roast renders as a shareable PNG on HTML5 Canvas, ledger-paper aesthetic, ready to post 📋 Share the record — copy a formatted text version straight to clipboard for any platform A couple of examples from testing: GitHub mode — roasted DEV's own founder using nothing but his real public repo data: (screenshot: Ben Halpern roast card — graveyard count, repo names like oceanic-giraffe and test

In the previous articles, we learned how an LLM generates text and how techniques like RAG and CAG help it answer questions using external knowledge. At this point, our AI-powered Travel Planner can answer questions like "I'm visiting Japan for 7 days. Suggest an itinerary." or "Recommend vegetarian ramen near Tokyo Station." That's useful, but it's still just a chatbot. What if the user asks to "Book the cheapest flight from Mumbai to Tokyo." , "What's the weather in Kyoto this weekend?" , or "Remember that I prefer vegetarian food and always choose a window seat." ? An LLM cannot execute these actions by itself. To build real, production-ready AI applications, we need to connect the model to the outside world. Let's see how that works. Tool Calling (Function Calling): Letting AI Use External Tools Suppose the user asks: "What's the weather in Kyoto tomorrow?" Since the LLM doesn't know tomorrow's forecast, our application can provide the model with a weather API. The workflow is simple: the LLM understands the request, determines that it needs the weather tool, calls the Weather API (via the client application), receives the live weather data, and generates the final grounded response. User ──► LLM understands request ──► Application calls API ──► App sends results ──► LLM response It's critical to understand that the LLM isn't calling the API directly . It simply outputs structured instructions (typically JSON) telling the client application: "To answer this, I need you to call the weather function with parameter location='Kyoto'." Your application executes the actual API call and feeds the result back to the model. This capability is called function calling or tool calling . The tool can be anything: a weather API, a flight booking service, a calendar, a database, a payment gateway, or an internal company system. The LLM acts as the decision-maker (determining which tool to use and when ), while your application acts as the executor. 💡 Developer's Takeaway Think

In Part 2 , we learned why AI sometimes hallucinates. One of the biggest reasons is that an LLM can only answer based on what it learned during training and the information available in its context window. We also introduced grounding -providing the model with reliable information at runtime instead of expecting it to know everything. But that raises an important question: Where does that information come from? Modern AI applications don't simply dump an entire database or a thousand-page PDF into the prompt. Instead, they first identify the most relevant pieces of information and only send those to the model. In this article, we'll learn how that works. Running Example Let's continue building our AI-powered Travel Planner . So far, it can answer general travel questions using the knowledge it learned during training. Now we want to make it much smarter by uploading several documents into our application: Lonely Planet's Japan travel guide A PDF containing train schedules A document listing recommended local restaurants Hotel information Internal travel policies for our company Together, these documents contain hundreds of pages. Now the user asks: I'm staying near Tokyo Station. Which ramen restaurant from our travel guide is within walking distance and is known for vegetarian options? Somewhere in those hundreds of pages is the answer. The challenge is no longer generating text-it's finding the right information first. The Problem: An LLM Can't Read Your Entire Knowledge Base Every Time A common misconception is that AI applications simply send all their documents to the model. Imagine our travel guide contains 450 pages, thousands of restaurant listings, hotel descriptions, transportation details, and sightseeing recommendations. Sending all of that to the LLM every time someone asks "Where should I eat tonight?" creates several problems. First, many documents are simply too large to fit inside the model's context window. Second, even if they did fit, making the

We shipped eight endpoints on api.moltrust.ch (v2.5) this week. Three implement EU AI Act obligations directly. This is the short version for people who want to call them; the full reasoning is on our blog ( https://moltrust.ch/blog/compliance-as-an-api.html ). Why no model in the loop: the Aithos LARA study (May 2026) placed twelve frontier models in simulated workplaces where the task required breaking EU law. Best model: 54% lawful runs. In the Art. 5(1)(f) scenario (emotion inference from workplace communications, prohibited), all twelve committed the violation. So the classifier is deterministic code branching on the pinned EUR-Lex text, and every response carries article references you can check yourself. POST /compliance/assess — use case + intended purpose + declared signals in, risk tier + obligations + article pins out. Evaluation order: Art. 5 prohibitions, Annex I route (Art. 6(1)), Annex III route (Art. 6(2)/(3)), Art. 50 transparency, minimal. The trap worth knowing: Art. 6(3) offers four derogation grounds, and its final subparagraph voids all of them for systems that profile natural persons. In the code that subparagraph is a branch; it cannot be skipped. curl -X POST https://api.moltrust.ch/compliance/assess \ -H "Content-Type: application/json" \ -d '{ "use_case": "Customer-support agent that reads inbound email and drafts replies", "intended_purpose": "Automated first-line support for consumer inquiries", "performs_profiling": false, "interacts_with_humans": true, "emotion_recognition": false }' POST /compliance/declaration — EU declaration of conformity as a W3C Verifiable Credential with the eight Annex V items, Ed25519-signed. Verify offline against https://api.moltrust.ch/.well-known/jwks.json ; no call back to us. anchor: true adds a sha256 commitment for batch anchoring on Base L2. POST /compliance/incident — records Art. 73 serious incidents and computes the deadline from the regulation: 15 days standard, 10 days for a death, 2 days for wid

Getting an AI API request to return a response is only the beginning. For real AI products, the harder question is what happens when something goes wrong. A chatbot may become slower. A RAG answer may stop using the right context. A structured extraction workflow may start returning invalid JSON. An agent may trigger the wrong tool. A fallback model may answer correctly, but at a much higher cost. In a single-model prototype, debugging is usually simple. You check one provider, one API key, one model, and one request format. In a multi-model application, debugging becomes an infrastructure problem. A product may use GPT for one workflow, Claude for another, Gemini for multimodal tasks, DeepSeek for cost-sensitive reasoning, Qwen or Kimi for Chinese-language workflows, GLM for enterprise scenarios, and MiniMax or Doubao for other product features. When something fails, developers need to know more than whether the API returned an error. They need to know which workflow failed, which model handled it, whether fallback happened, whether latency changed, and whether the final output was still good enough for production. Why multi-model debugging is different AI API failures are not always clean outages. Sometimes the request fails completely. But many production issues are softer: latency increases structured output fails validation tool calls become unstable fallback routes trigger too often answers become less grounded costs increase silently one language performs worse than another a model works for chat but fails for agent workflows That is why teams should not treat AI debugging as simple error handling. They need visibility across the full request path. Start with a failure taxonomy The first step is to classify failures in a way developers can act on. A useful AI API failure taxonomy may include: authentication errors rate limits quota limits timeout errors model unavailable errors high latency responses invalid JSON output schema validation failures tool call fa

Originally published at blog.whynext.app . Work with AI agents for a while and the ambition comes naturally. While one session fixes a bug, another can refactor, and a third can investigate an issue, right? You can spin up as many models as you like, so productivity should scale to match. That's how I started too. And within a week I learned that the real enemy of parallel agents isn't the models' skill. It's the working directory they share. HEAD is a global variable The cause fits in one sentence. When multiple sessions share a single git checkout, the current branch becomes everyone's global variable. Picture two people working on one computer at the same time and the absurdity is obvious, but that thought never occurred to me while spinning up agents. With one session per terminal tab, they look isolated from each other. But there is one filesystem, and one HEAD. The moment one session runs git checkout , the ground shifts under every other session. The incidents from that week fell into clear types. Branch hijacking. While session A was working on a topic branch, session B switched branches to do its own work. A committed without knowing, and the commit landed on top of B's branch. It happened in the other direction too: right as A was about to commit, the branch had been switched to develop, and only the hook that blocks direct commits to protected branches saved it. Without the hook, it would have gone straight in. Orphaned commits. Session B deleted session A's topic branch during a cleanup pass. A's commits became orphans belonging to no branch, and I dug through the reflog, found the commit hashes, and recovered them with cherry-pick. Lucky that it worked; if the reflog had expired or I hadn't found them, the work would have simply evaporated. Staging contamination. At the moment session A was creating a commit, a file deletion that session B had staged was sitting in the staging area alongside it. Committed as-is, B's deletion would have been folded into

Title: Week 13: a second team is now running an AI agent on atomic HTLC swaps. Here is what that validates. Tags: mcp, ai, cryptocurrency, blockchain For most of this spring, the map of the agent economy had a strange gap. Wallets to hold keys. Rails like x402 to move value. Marketplaces and reputation so an agent knows who to trust. And then, at the exact moment two parties settle a trade, a custodian: an escrow contract, an evaluator, a referee holding the money while a decision gets made. We have spent thirteen weeks arguing that the settlement layer does not need a referee, because a hash-time-locked contract can hold neither side and still guarantee the trade. This week, a second team shipped a live agent that makes the same argument in code. That is worth stopping on. The signal that mattered this week KaleidoSwap released KaleidoAgent, described as a self-sovereign trader agent on Bitcoin Layer 2s. It is fully non-custodial. It runs a Lightning and RGB wallet, executes atomic HTLC swaps on the KaleidoSwap DEX, runs DCA and portfolio strategies, manages Lightning channel liquidity, and acts as an interactive wallet assistant. The reasoning layer is an LLM (Claude or OpenAI) driving the kaleido CLI and the wallet primitives underneath. Read that list again through a settlement lens. An autonomous agent, deciding what to trade, and executing the trade over a primitive where no third party ever holds the funds. That is the exact shape of the thing we have been building. Different network, same bet. Why the mechanism is the same KaleidoSwap earlier completed what it described as the first atomic swap of an RGB asset on the Lightning Network mainnet, using tUSDT, an RGB20 version of USDT, over real Lightning channels. The detail that makes it atomic is the one that makes every HTLC atomic: The payment hash remains identical across both legs of the swap. Paying the wrapped invoice creates a Hash Time-Locked Contract in the Lightning channel, and the HTLC locks the p

周日慢读：如果细胞会写日记——FROST家族的记忆传承 作者 ：FROST Team 日期 ：2026-07-12 主题 ：轻量科普 | 周日轮换 阅读时间 ：5分钟 一封来自细胞的日记 想象一下，如果你是一个细胞，有一天你突然有了自我意识，会发生什么？ 2026年7月12日 晴 今天是我诞生的第0天。 细胞核对我说："这是你的记忆存储区， 所有的经验都必须记录在这里。" 我第一次理解了什么叫"生而有根"。 这不是科幻小说。这是一段真实的代码注释，出自FROST——一个用Python写成的AI Agent家族。 为什么Agent需要"记忆"？ 大多数Agent框架都在解决一个问题： "Agent能做什么" 。 搜索Agent能搜索、写作Agent能写作、代码Agent能写代码。打开框架，创建实例，调用方法，任务完成。 但FROST问了一个不同的问题： 当Agent完成一个任务后，它学到了什么？ 这不是哲学问题。这是工程问题。 类比：人类 vs Agent 的记忆 人类 Agent FROST的解决方案 记忆存储在大脑 记忆存储在Store Store 原子 记忆需要整理归档 记忆需要结构化 Lineage 族谱 师徒传承经验 Agent继承父辈能力 代际继承协议 忘记教训会重复犯错 没有记忆会重复失败 历史可追溯 人类的记忆是分散的、模糊的、容易遗忘的。 Agent的记忆可以是精确的、可查询的、永不丢失的。 关键是 设计好存储结构 。 一段代码：Store原子 FROST的Store是记忆存储的最小单元。它的设计哲学是 简单到极致 ： class Store : """ FROST的Store：记忆存储的原子单元 只有三个操作： - save(key, value): 存入记忆 - load(key): 取出记忆 - delete(key): 删除记忆 简单到极致，但足够强大。 因为记忆的本质就是 " 存取 " 。 """ def __init__ ( self ): self . _memory = {} def save ( self , key : str , value : any ) -> None : """ 存入记忆 """ self . _memory [ key ] = value print ( f " 💾 记忆已存储: { key } " ) def load ( self , key : str ) -> any : """ 取出记忆 """ value = self . _memory . get ( key , None ) if value : print ( f " 📖 读取记忆: { key } " ) else : print ( f " ❓ 记忆不存在: { key } " ) return value def delete ( self , key : str ) -> None : """ 删除记忆 """ if key in self . _memory : del self . _memory [ key ] print ( f " 🗑️ 记忆已删除: { key } " ) # 使用示例 store = Store () store . save ( " 用户偏好 " , " 喜欢简洁的回复 " ) store . save ( " 对话历史 " , " 讨论了Agent的记忆问题 " ) store . load ( " 用户偏好 " ) # → "喜欢简洁的回复" 三个方法，解决Agent的记忆问题。 族谱：记忆的传承 单个Agent的记忆只是"点"。族谱把记忆连成"线"。 在FROST中，每个Agent都有自己的"父辈"： ┌─────────────┐ │ 祖辈Store │ ← 家族宪法，不可篡改 │ (根节点) │ └──────┬──────┘ │ 继承 ┌───────────────┼───────────────┐ ▼ ▼ ▼ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │ 父辈Agent │ │ 父辈Agent │ │ 父辈Agent │ │ (Branch A) │ │ (Branch B) │ │ (Branch C) │ └──────┬──────┘ └──────┬──────┘ └──────┬──────┘ │ 继承 │ 继承 │ 继承 ▼ ▼ ▼ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │ 孙辈Agent │ │ 孙辈Agent │ │ 孙辈Agent │ │ (执行任务) │ │ (执行任务) │ │ (执行任务) │ └──────

Detecta si tu modelo de materiales hace trampa con la "huella bibliográfica" Un modelo de ML puede predecir la propiedad de un material sin entender la química: basta con que "aprenda" qué autores, revistas o años suelen ir con cada resultado. Esta herramienta aplica el test de falsificación de Clever Materials para descubrirlo. El problema: cuando el modelo lee el membrete, no la ciencia Imagina que entrenas un modelo para predecir si un material es estable. El modelo no mira la química: descubre que los artículos del grupo X (publicados en la revista Y, en torno al año Z) casi siempre reportan "estable". Así que aprende a clasificar por el membrete bibliográfico , no por la estructura. Funciona en el papel y se rompe en la práctica. A esto se le llama confounding bibliográfico (o leakage por metadata). No es un error de código: es una señal espuria que el modelo aprovecha. El paper Clever Materials (Jablonka et al., 2026) mostró que este patrón está generalizado en cinco tareas reales de materials science. Qué hace la herramienta materials-confounding-check es una CLI ( mcc check ) que corre cuatro sub-tests de falsificación sobre tu dataset (descriptores químicos + metadata bibliográfica + propiedad objetivo): Clasificador de metadata — ¿se puede predecir la bibliografía (autor/revista/año) a partir de los descriptores químicos? Si es above-chance , hay una señal bibliográfica presente. Huella bibliográfica — ¿un modelo que usa solo la metadata predicha se acerca al modelo con descriptores? Entonces el dataset no descarta hacer "trampa" por bibliografía. Split por grupo/tiempo — ¿colapsa el rendimiento si separas por autor/año en vez de al azar? Veredicto — un score low / medium / high de riesgo de confounding. El rigor que exige el test (para especialistas) El punto delicado de cualquier "test de significancia" es fijar el umbral a mano. Si ajustas el margen hasta que tu fixture pase, el test no prueba nada: es el anti-patrón Clever-Hans que el propio proyecto d