标签：#AI

Over the past year I have been fine-tuning open vision-language models - 9B dense up to a 35B mixture-of-experts - with supervised fine-tuning and GRPO-style reinforcement learning on verifiable rewards. Most of what I learned was not about algorithms. It was about the ways a training run can look healthy while doing nothing, or crash for reasons that have nothing to do with your code. Three failures, in increasing order of how long they fooled me. Failure 1: the metric that measured the wrong thing (18 hours) I ran an 18-hour supervised fine-tune that reported token accuracy climbing steadily to 99%. Looked like a textbook run. The real evaluation metric - accuracy on multiple-choice questions - never moved. The cause was a mismatch between what I supervised and what I evaluated. The training loss was over free-text reasoning traces; the evaluation scored a single extracted answer letter. The model got extremely good at reproducing the shape of the training text - hence 99% token accuracy - without that transferring to the decision I actually cared about. Token accuracy is a proxy, and proxies drift from the target exactly when you stop checking. The fix was structural, not a hyperparameter: supervise the thing you evaluate. If the deliverable is a constrained answer, the training signal has to reach that answer, not just the prose around it. The general rule I took: any training metric that is not your evaluation metric is a hypothesis about correlation, and you should check that correlation before you spend GPU-days on it. Failure 2: the crash that was two libraries disagreeing about position ids The GRPO trainer for the 9B vision model crashed in the forward pass, deep inside rotary position embedding code. Nothing in my training code had changed. The diagnosis took a while because the bug lived at the boundary between components: the text sequence length was derived from token-type ids, while the vision sequence length came from the image grid - and image-pad t

Cloudflare recently open-sourced Cloudflare OS. It allows enterprise teams to output work artifacts grounded in enterprise knowledge, know-how, and provisioned connectors, automate repetitive workflows with optimized token cost (with AI assistance only where needed), and build personal, shareable, customizable work software that caters to specific, complex use cases within a secure sandboxed model By Bruno Couriol

Small language models can now run directly on phones. But most of them stop learning the moment they ship. For personal AI, that feels like a strange stopping point. Some of the most useful signals arrive only after the model acts: Did the user dismiss the notification? Did they open it later? Did they rewrite the suggestion? Did they ask for it again? These interactions contain useful information about the user, but they are delayed, private, and ambiguous. They are not clean labels, and they are not reliable scalar rewards. To explore this problem, we built Online-SDFT , an open-source prototype that continually fine-tunes a small language model from delayed interactions while keeping the learning loop on the device. The prototype uses: LiquidAI/LFM2.5-230M A rank-4 LoRA adapter ONNX Runtime Training A bounded on-device replay buffer An Android notification-routing testbed Once the model has been provisioned, inference, interaction storage, replay, and adapter updates all happen locally. Why standard fine-tuning is awkward here Suppose the model receives a notification and chooses one of three actions: Show it now Save it for later Archive it Supervised fine-tuning would require a correct action for every notification. But the phone never observes what the ideal action was. Reinforcement learning replaces the correct answer with a reward, but that reward is also difficult to define. Opening a notification does not necessarily mean it arrived at the right time. Ignoring it does not necessarily mean it was unimportant. The user may simply have been busy. There is another complication: the model only observes the result of the action it actually took. If it archives a notification, it cannot know what would have happened had it shown the notification immediately. What the phone receives is not a label or reward. It receives hindsight . Using the same model as student and teacher The core idea is simple: let the model reconsider its decision after seeing what happened

Those 200 × 200 and 400 × 400 numbers tell you where the mounting holes line up, but you still need to check the mount's weight limit and hardware.

I Built AgentCheck Because “The Coding Agent Said Done” Wasn’t Enough AI coding agents are getting surprisingly good at writing code. I use them regularly, and they can handle increasingly large tasks: refactoring code, adding features, updating dependencies, modifying configuration, creating migrations, and touching files across an entire repository. But I kept running into the same problem after the agent finished: How do I independently verify what it actually changed? The agent usually gives me a perfectly reasonable summary. Something like: Done. Implemented the requested changes, updated the tests, and cleaned up the affected code. Useful? Absolutely. Enough for me to commit without checking? Not really. So I built AgentCheck . The Problem Happens After “Done” After a coding agent finishes a task, I still find myself manually checking things like: Which files actually changed? Were any files deleted? Did configuration change? Were dependencies added or updated? Was a database migration introduced? Did anything that looks like a secret appear? Were related tests changed? Is the overall change set larger or riskier than expected? Of course, Git already gives us the raw information. I can run: git status git diff git diff --stat Then inspect individual files. And I still do that. But once coding agents become part of your normal workflow, repeating the same verification process after every task starts to feel like something that should be structured. That was the idea behind AgentCheck. What AgentCheck Does AgentCheck creates a trusted checkpoint before your coding agent starts working. Then, after the agent finishes, it compares the current Git-visible repository state with that checkpoint. The basic workflow is deliberately small: agentcheck start Then let your coding agent work. That can be: Codex Claude Code Cursor another AI-assisted coding tool or technically even a human When the work is finished: agentcheck AgentCheck then produces four sections: Changes

Quick one: if PyRIT (Microsoft's Python Risk Identification Tool) is on your shortlist for LLM red-teaming, check the repo first. Azure/PyRIT was archived on GitHub on March 27, 2026. It's read-only now: no commits, no releases, no issue triage, nothing. Whatever version you pip-installed is the last version you'll ever get. That matters more for PyRIT than it would for most tools, because PyRIT was never a turnkey scanner. It's a framework for scripting multi-turn attack orchestration, the kind of thing a red team builds custom attack sequences on top of. A framework that's stopped shipping fixes is a worse foundation to build on than a finished tool that's stopped shipping features, because you were relying on it staying flexible to your needs, and now it can't. So what do you use instead? Depends on what you were actually using PyRIT for: You wanted a broad, actively maintained app-layer scanner -> promptfoo . Zero-install via npx promptfoo , 50+ red-team plugins, OWASP/NIST/MITRE ATLAS report mappings, and it's still getting regular releases. You wanted model-layer testing (jailbreaks, encoding tricks, data leakage on the base model itself, not your app) -> garak . NVIDIA-maintained, pip installable, 8k+ stars, actively developed. You wanted OWASP-mapped detectors and don't mind a paid tier for continuous scanning -> Giskard . The open source scanner is real and current; the always-on Hub is commercial. You wanted a fast, zero-setup smoke test before reaching for any of the above -> that's the gap we built sentinel-scan-cli for. Dependency-free CLI (Python and npm ports, identical output), 15 attack patterns each tagged to its OWASP LLM Top 10 category, --demo runs with no config and no API keys in under a minute. None of these replace PyRIT's specific multi-turn orchestration model one-for-one, if that's genuinely what you need, Microsoft's PyRIT Community fork discussion or building your own harness on top of a maintained model API is probably the honest answe

When Satoshi Nakamoto designed Bitcoin's Proof-of-Work consensus, the goal was simple: one CPU, one vote. What actually happened was very different. ASIC farms centralized mining into industrial warehouses, and the "one CPU" vision became "one warehouse, one vote." Proof-of-Stake was supposed to fix this by replacing energy expenditure with economic stake. Instead, it created a different problem: the rich get richer, forever. RustChain's Proof-of-Antiquity (PoA) takes a radically different approach. Instead of rewarding who has the most money or the newest hardware, it rewards who has kept the oldest hardware running the longest. The core insight is elegant: time is the one resource that can't be bought, faked, or manufactured. Either your hardware has been alive for twenty years, or it hasn't. This article does a deep technical comparison of Proof-of-Antiquity and Proof-of-Stake, drawing on the actual RustChain source code to explain how each consensus mechanism handles decentralization, Sybil resistance, economic fairness, and network security. The Fundamental Philosophies Proof-of-Stake: Wealth as Security Proof-of-Stake systems — Ethereum 2.0, Cardano, Algorand, Solana (with its Delegated PoS variant) — all share a common assumption: the more tokens you stake, the more committed you are to network security. If you act maliciously, your stake gets slashed. The economic logic is straightforward: attackers would need to acquire a majority of the token supply, which would be prohibitively expensive. The problem is what happens after someone acquires that stake. In PoS, staking rewards compound. A validator with 10x the stake of a small holder earns 10x the rewards, which they can reinvest into more stake. Over time, validator concentration increases. On Ethereum, Lido + Coinbase + Binance + Kraken collectively control over 50% of staked ETH. The "rich get richer" dynamic isn't a bug — it's a mathematical inevitability of proportional rewards based on capital. Proof-

In June a reviewer on DEV who goes by @anp2network told me to stop storing a conclusion. I had a...

There are two kinds of articles about coding with AI. The ones that generate a sorting function and...

A mysterious new AI model called Ox Alpha has driven certain corners of the internet into a frenzy of speculation.

It’s been over a decade and two console generations since GTA V came out, and its sequel is still a work in progress. GTA VI has faced multiple delays, with developer Rockstar Games bumping back its predicted release date twice, first to May 26th, 2026, then again to November 19th, 2026, a full year after […]

Linkdaze's smart digital calendar stands out for not putting its features behind a paywall, including an AI meal planner tool.

Originally published on michael-kaminski.io . The Genome of Games publishes the same 1,180 records four different ways, and one command writes all four: node build.js , 0.39 seconds, zero npm dependencies. Out come 1,245 static HTML pages for crawlers, an interactive canvas graph for humans, a 129,037-byte search index for the site's own search box, and a Model Context Protocol server exposing 8 tools to agents. The decision worth copying is the one that sounds like a downgrade. The MCP server does not query the site and does not read the source data. It statically imports a 1.9 MB index that the build wrote. There is exactly one place where slugs, lineage, and adoption edges get joined, so an agent and a crawler cannot come back with different answers. The dataset is an ontology of video game mechanics — 168 mechanics, 618 games, 394 companies, 4,366 recorded links, 1962 to 2025. What the records are about does not matter here. The shape of the problem shows up anywhere a structured knowledge base has to serve both a search engine and a model. Four surfaces, one build, a twelve-fold expansion Six hand-edited JSON files under data/ are the source of truth: the feature ontology, the graph, the prose, the company registry, the site copy, and the verified outbound links. Together they are 1,312,577 bytes. The build turns that into 16,644,215 bytes of generated read surface. A 12.7× expansion, and every byte of it is disposable. Surface Consumer Bytes Per entity 1,245 static HTML pages Crawlers, humans 14,613,203 11,728 / page mcp-index.json → MCP server Agents 1,901,975 1,612 search-index.json The site's own search box 129,037 109 /graph/ canvas Humans exploring lineage data injected at build — The build also emits sitemap.xml with 1,245 entries, llms.txt , robots.txt , and a 404 page. The same run reports 96,843 internal links across those pages. Nothing in that list is authored. Delete the whole output directory and the next build restores it in under half a second.

Building AI Agents with frameworks like CrewAI or LangChain often hits a bottleneck: heavy, slow web scraping that bloats context windows and increases LLM token costs. To solve this, I built Smart Scraper M2M — a lightweight, high-performance web scraper API designed specifically for machine-to-machine (M2M) communication. 🌟 Key Features ⚡ Ultra-fast: Returns clean structured JSON in ~30ms . 🧠 Context-optimized: Strips out useless HTML/CSS junk so your LLMs process only relevant data. 🤖 Agent-friendly: Built to integrate seamlessly into CrewAI, LangChain, or custom Node.js agents. 🚀 Quick Start You can test the API or check the full source code directly on GitHub: 🔗 GitHub Repository: https://github.com/MRIGL/smart-scraper-m2m 💬 Feedback & Community I’m actively improving the API and would love to hear your thoughts, feature requests, or contributions! Feel free to star the repo or leave a comment below.

The model was never the problem, and that is exactly why the bug took three days to find. My ticket-classification service started returning the fallback label for long, non-English messages shortly after I moved the inference path to a cheaper endpoint, and every instinct pointed at the new model. The real culprit was a token-counting mismatch that silently truncated the prompt before the model ever saw the classification instruction. The Symptom The failure was remarkably consistent, which made it even more misleading. Messages under roughly two thousand characters classified correctly, while longer ones, especially in German and Japanese, fell through to a generic "other" bucket with a perfectly valid JSON response. The parser was not the issue, the prompt had not changed in weeks, and the retry logic never fired because the endpoint returned a normal 200 status. My first assumption was that the cheaper model was simply weaker at long-context reasoning, so I ran a controlled comparison using the same fifty tickets against the previous endpoint. The old path classified all fifty correctly, the new one failed on nineteen, and that result seemed to confirm the model-quality theory. What bothered me was the distribution: the failures clustered exactly where the input length crossed a threshold, and no ticket under that threshold ever failed. The Reproduction To isolate the variable, I needed a clean environment where I could swap endpoints without touching the production deployment, and MonkeyCode's free server option turned out to be a practical debugging tool. The project is open source, and its free model access let me replay the failing tickets without spending my own quota, so I spun up a disposable instance and pointed the same harness at the same prompt. Disclosure: This article was prepared as part of MonkeyCode's product outreach. The reproduction took about twenty minutes, and the result was identical on every retry: long inputs failed, short inputs passed.

When a free endpoint returns 429, most agents do the most expensive thing possible: retry. Retrying looks harmless. A 200-millisecond request becomes a 2-second wait, then another attempt. But under peak load, that loop becomes a 30-second stall while your agent clicks refresh on an empty response. If the quota window resets during the stall, every retry burns tokens you could have spent on actual work. The retry loop assumes the failure is temporary. For rate limits, that assumption is usually wrong. Quota counters reset on a fixed schedule, not on your convenience. You are not just waiting; you are burning wall-clock time that could have gone elsewhere. The Cascade Pattern A cascade router is the alternative. It sends requests to the free endpoint, backs off on rate-limit signals, then degrades gracefully to a backup endpoint. The free tier carries the load; the backup exists only when needed. You get the cost advantage of the free tier and the reliability of the paid tier. The design has three parts: an endpoint abstraction layer, a rate-limit detector, and a circuit breaker that trips when the free endpoint fails repeatedly. Here is the core code: # cascade_router.py — free tier first, paid/self-hosted as fallback. import json import os import time import urllib.error import urllib.request from dataclasses import dataclass @dataclass class Endpoint : name : str url : str api_key : str model : str cooldown_until : float = 0.0 consecutive_failures : int = 0 def available ( self ) -> bool : return time . time () >= self . cooldown_until class CascadeRouter : def __init__ ( self , endpoints : list [ Endpoint ]): self . endpoints = endpoints def _call_one ( self , ep : Endpoint , messages : list [ dict ]) -> tuple [ int , dict ]: body = json . dumps ({ " model " : ep . model , " messages " : messages , " max_tokens " : 256 }). encode () req = urllib . request . Request ( ep . url , data = body , headers = { " Content-Type " : " application/json " , " Authorization "

A developer usually meets Mac power management through a simple need. A build, local server, download, or agent is still running, and idle sleep would interrupt it. The caffeinate command can be enough for that open lid case. Lid close is a different boundary. An idle sleep assertion does not mean the same thing as a closed display session, and a product should not blur the distinction. I built Afterlid around three explicit states. Sleepy follows normal sleep. Awake prevents idle system and display sleep while the lid is open. Always On is the lid closed mode, with the display off. The important engineering work begins after activation. What happens if the app crashes? What happens when the battery is falling or the machine is under thermal pressure? What state is restored after a helper failure? For Afterlid, Always On ends at 30 percent battery while unplugged, under serious or critical thermal pressure, when the app heartbeat disappears, or after eight hours. When a limit fires, the app drops its wake assertion and returns the Mac to normal sleep behaviour. The current implementation uses a small privileged helper and an undocumented macOS sleep setting for the lid closed path. That makes broad hardware testing and honest release notes essential. It is not something I want to hide behind a friendly menu bar character. A useful principle emerged from the work: activation is a feature, but recovery is the product. If you are building a system utility, test the path back to the operating system defaults with the same seriousness as the path into your special mode. Founder disclosure: I built Afterlid. The full product and current boundaries are here: AfterLid

The class action suit claims that Amazon never obtained consent from Twitch streamers to be used to train its AI models.

从 Demo 到生产：那些真正让 AI Agent 敢上线的护栏 开场钩子： 你在网上看到的多数「AI Agent」都是 demo。它们之所以上不了生产，原因往往 只有一个 —— 而下面这个开源的小脚手架，专门解决它。 我们已经过了「能调通大模型」就算赢的阶段。现在真正难的是那没人讲的 10%： 是什么阻止 Agent 做出伤害性的事？ 我在微软跑过一套约 25 个 Agent 的生产平台，现在也帮团队把 Agent 从笔记本推进到真实用户面前。两边的体会是一致的。 一个不太舒服的真相：能调 5 个工具的聊天机器人， 不是产品 。周末项目和你敢放到客户面前的 系统之间，差的只有三件事 —— 而且全都是不酷、不性感的工程： 你怎么给输出质量打分 （质量门）。 你怎么决定什么时候必须人签字 （审批门）。 你如何让整套东西模型无关 ，不被某个厂商锁死。 所以我写了一个很小的 harness，把这三件事摆在最显眼的位置。它故意做得很小 —— 一小时能 读完 —— 因为价值不在「框架」，在 模式 本身。 仓库： github.com/zhasun0818/ai-agent-scaffold 1. 质量门：别发布你无法打分的东西 Agent 的输出是「预测」不是「承诺」。上线前它必须过一道 检查 ：是否达到你的标准。脚手架里 这是一个可插拔的 QualityGate ，你可以换成 LLM 裁判或测试套件： # agent_harness/eval.py @dataclass class EvalReport : passed : bool score : float checks : List [ str ] class QualityGate : def grade ( self , proposal : str , context : str = "" ) -> EvalReport : return self . grader ( proposal , context ) 循环在门没过之前拒绝执行： result . report = self . quality . grade ( proposal , f " state= { state } " ) if not result . report . passed : self . approval . log ( " quality-gate " , " blocked " , result . report . __str__ ()) return result 注意它 把拦截记录下来了 。生产里你会想把这些被拦的尝试都进可观测性系统。「这周我们拦下 了 12% 的 Agent 提议」是个真实 KPI —— 它说明门在工作。 2. 审批门：所有人都忘掉的那一步 这才是让企业真正点头说「可以」的东西。当 Agent 想加急订单、取消订阅、或动钱的时候，它应该 停下来问人 。沉默不等于同意。 # agent_harness/approval.py class ApprovalGate : def request ( self , action : str , detail : str ) -> bool : # 生产里：推一条通知到 Teams / Slack / 邮件，然后等待。 decision = input ( f " Approve { action } ? [y/N] " ). strip (). lower () self . audit . append ( AuditEntry ( time . time (), action , " human-reviewer " , decision , detail )) return decision . startswith ( " y " ) 在脚手架里，标记 needs_approval=True 就够了： @tool ( " expedite_order " , " Mark an order as expedited. " , needs_approval = True ) def expedite_order ( order_id : str ) -> str : return f " PO { order_id } : marked expedited " 而且因为有 审计链 ，你永远能回答「谁改的、为什么」—— 这通常是合规团队问的第一个问题。 3. 模型无关的 provider：别跟一个厂商结婚 模型每几周就变，价格也是。你的 Agent 循环不该知道自己在对谁说话： # agent_harness/providers.py class ModelProvider ( Protocol ): def

From Demo to Production: The Guardrails That Make an AI Agent Safe to Ship Hook: Most "AI agents" you see on the internet are demos. Here's the single most common reason they never reach production — and a small, open-source harness that gets past it. We are past the phase where the hard part of building an AI agent was calling the model. The hard part now is the 10% nobody talks about: what stops the agent from doing something harmful? I've seen this from both sides — I built and ran a ~25-agent platform in production at Microsoft, and now I help teams take agent ideas from a notebook to real users. The uncomfortable truth: a chatbox that can call 5 tools is not a product. The difference between a weekend project and a system you can put in front of customers is three things — and they're all boring, non-glamorous engineering: How you grade output quality (the quality gate). How you decide when a human must sign off (the approval gate). How you make the whole thing model-agnostic so you're not locked into one vendor. So I wrote a tiny harness that keeps these front and center. It's intentionally small — small enough to read in an hour — because the value isn't in a framework, it's in the pattern . Repo: github.com/zhasun0818/ai-agent-scaffold 1. The quality gate: don't ship what you can't grade An agent's output is a prediction, not a promise. Before it ships, you need a check that it passes your bar. In the harness this is a pluggable QualityGate — a rule of thumb you swap with an LLM judge or a test suite: # agent_harness/eval.py @dataclass class EvalReport : passed : bool score : float checks : List [ str ] class QualityGate : def grade ( self , proposal : str , context : str = "" ) -> EvalReport : return self . grader ( proposal , context ) The loop refuses to execute if the gate fails: result . report = self . quality . grade ( proposal , f " state= { state } " ) if not result . report . passed : self . approval . log ( " quality-gate " , " blocked " , result