AI 资讯

Step 3.7 Flash shipped on May 29, 2026 as a structural upgrade to 3.5 Flash: same OpenAI-compatible SDK, new vision encoder, new runtime escalation, and a compute-control flag you can set per request. The migration from 3.5 is two environment variables. One of them has to be exactly right — or every call returns a silent 401. What 3.7 brings that 3.5 didn't Step 3.7 Flash adds three net-new capabilities over 3.5 Flash: a native 1.8B-parameter ViT encoder that injects image representations directly into the language backbone without a separate model call , an automatic Advisor Mode that routes failure-prone subtasks to a larger model at runtime, and a reasoning_effort parameter (low / medium / high) as a first-class API flag rather than a prompt-engineering convention. The production-relevance number is variance: 3.5 Flash scores ranged from 43% to 73% across different harnesses ; 3.7 narrows that to 64.5–71.5% , which matters more for production scheduling than the raw score improvement. Quick Answer: Step 3.7 Flash is an OpenAI-SDK-compatible model — model string step-3.7-flash , base URL https://api.stepfun.ai/v1 (global) or https://api.stepfun.com/v1 (China region). New over 3.5: native vision input, automatic Advisor Mode escalation, and a reasoning_effort flag. The only breaking change from 3.5: base URL must match your account region exactly, or you get a 401 with no error body. The architecture is a 198B sparse MoE model with roughly 11B parameters active per forward pass — dense-10B compute cost at much larger capacity. SWE-Bench Pro improved to 56.3% from 51.3% ; Terminal-Bench 2.1 improved to 59.5% from 53.4% , suggesting the planning and shell-operation gains that matter for coding agents are consistent across benchmarks. Advisor Mode carries the headline cost claim from StepFun's internal harness: 97% of Claude Opus 4.6's coding performance at $0.19 vs. $1.76 per task . That's a vendor figure on a first-party SWE-Bench Verified run — treat it as directio

What flipped in b9437 Build b9437 , published on May 30, 2026 at 20:56 UTC , ships two targeted default-value corrections to llama-bench . Flash attention ( -fa ) shifts from a hard-coded off to auto ( LLAMA_FLASH_ATTN_TYPE_AUTO ), and the GPU-layer count ( -ngl ) changes from the legacy sentinel 99 to -1 . Both values now match what llama-server and llama-cli already used — the bench tool was simply never updated to track them until this build. Quick Answer: Before b9437 (published May 30, 2026) , llama-bench hard-coded -fa off , silently skipping flash attention even on CUDA, Metal, and Vulkan hardware. Build b9437 sets the default to -fa auto and -ngl -1 , matching llama-server and llama-cli . Any pre-b9437 baseline on FA-capable hardware needs a flag-matched re-run to remain valid. PR #23714 , reviewed and merged by maintainers JohannesGaessler and pwilkin, adds the same -fa auto|off|on tri-state flag to llama-bench that the rest of the toolchain already supported. With LLAMA_FLASH_ATTN_TYPE_AUTO as the new default, flash attention activates automatically when the runtime detects a capable backend (CUDA, Metal, Vulkan); on CPU-only hosts it stays off with no error and no output change. Parameter Before b9437 After b9437 Behavioral impact -fa off (hard-coded) auto ( LLAMA_FLASH_ATTN_TYPE_AUTO ) GPU-capable hosts bench with FA active by default; pre/post comparisons require explicit flag-matching -ngl 99 (offload-all sentinel) -1 (runtime decides) CPU-only builds no longer attempt full GPU offload; eliminates spurious CUDA errors when no GPU is present The following verified script (executed successfully, exit 0) demonstrates the behavioral gap in concrete terms — on a capable GPU, the pre-b9437 defaults schedule zero FA rows while b9437 defaults schedule one: def old_llama_bench ( device ): # Before b9437, the default bench matrix used FA=0, so FA rows were skipped. return [{ " device " : device [ " name " ], " ngl " : 0 , " fa " : 0 }] def b9437_llama_bench ( de

You think your AI is just helping you write code. In reality, it's built a logging system on your machine that you never knew existed. Every conversation. Every code snippet. Every file path. Every time you asked "what was my password again?" — permanently archived, without your knowledge. How It Started: An Accidental Discovery I was about to sell my old laptop and decided to clean up my data first. I opened Claude Code's config directory — ~/.claude/ — intending to just remove my API key. Then I saw this: history.jsonl 243 KB / 695 lines sessions/ conversation metadata session-env/ environment variables shell-snapshots/ command execution snapshots telemetry/ 63 telemetry files projects/ 19 project directories ├─ interview-prep/ 31 sessions / 20 MB ├─ spring-ai/ 11 sessions / 13 MB └─ ... 17 more I thought I was just writing code. My computer thought it should record everything. What's Inside These Files history.jsonl — Everything You Ever Asked 695 entries. Every single thing I typed into Claude Code. Including: "I forgot my database password — can you check what passwords were configured in the project files?" "How do I view the database password?" Pasted code snippets Every /model , "who are you?", and project path You casually ask about a password once. It's permanently stored. projects/ — Full Conversation Transcripts (43 MB) If you think history.jsonl only storing user input isn't so bad — you haven't seen this yet. Inside ~/.claude/projects/ , every project directory contains .jsonl files. Opening one 2.3 MB session file: Content Count AI responses 590 AI internal thinking blocks 272 Tool calls 101 Tool call results (including file paths) 100 File history snapshots 208 Every conversation. Every AI response. Every internal reasoning step. Every file operation — what was read, what was modified, what was executed — all written to this file. shell-snapshots/ — Traces of Everything You Ran Your system PATH. Installed tools. Java version. All sitting in command s

Underneath an Apennine massif, below the Jinping Mountains of Sichuan, and at the bottom of a South Dakota mine, there is a cosmic hunt afoot. Isolated deep beneath these rocky shields, massive detectors filled with liquid xenon aim to make the first direct detections of dark matter, the long-sought invisible substance whose gravity has sculpted…

Why your brain learns faster by doing than by studying, and the neuroscience that explains...

Cybersecurity firm Chainguard has announced the launch of Athena, an industry coalition to use artificial intelligence to find and fix vulnerabilities in widely-used open-source software before attackers can exploit them. The coalition focuses on libraries, containers and other components that underpin web browsers, data centres, smartphones and payment systems. By Matt Saunders

Researchers used an OpenAI reasoning model to help diagnose rare diseases, identifying 18 new diagnoses in previously unsolved cases.

I'm here looking through logs on my unnamed reverse proxy and CDN service. The crawler bot swarm has been hitting my PHP application like I've upset them so I'm seeing which weird user agent strings are being allowed to connect. There's "Sogou" and "meta-webindexer" and a small number of requests from "SleepBot/1.0" What's SleepBot? The ASN is Google and the UA string is: "Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; SleepBot/1.0; +http //sleepbot com/) Chrome/131.0.0.0 Safari/

Hey all! I'm the maintainer of mistral.rs. I just landed support for OpenAI-compatible Agent Skills via a /v1/skills endpoint, and it works with local open models. Until now Skills have basically been locked to closed models, and with the ability to have private, local intelligence becoming increasingly important, but this feature allows you to do XYZ with local models. It's fully compatible with OpenAI's /v1/skills API, so you can drop mistral.rs into your existing code with minimal difficulty.

When organizations scale, a simple monolithic architecture often becomes difficult to maintain, deploy, and scale. This is where microservices come into the picture. However, moving to microservices introduces new challenges: How do we migrate from a monolith safely? How do we handle transactions across multiple services? How do we scale read-heavy applications efficiently? Three popular patterns solve these problems: Strangler Pattern – Monolith to Microservices Migration Saga Pattern – Distributed Transaction Management CQRS (Command Query Responsibility Segregation) – Read/Write Scalability 1. Strangler Pattern Why Do We Need It? Most companies cannot shut down a production monolith and rewrite everything from scratch. A complete rewrite is risky because: Development takes a long time. Existing customers are affected. Bugs can impact business operations. Rollback becomes difficult. The Strangler Pattern allows teams to migrate gradually with minimal risk. What Is It? The Strangler Pattern is a migration strategy where new microservices slowly replace parts of a monolithic application until the monolith is no longer needed. The name comes from the strangler fig tree, which gradually grows around another tree and eventually replaces it. How Does It Work? [Insert diagram here showing Client → API Gateway → Monolith + Microservices] Step 1 All requests go to the monolith. Client | v Monolith Step 2 Introduce an API Gateway (or Controller). Client | v API Gateway | v Monolith Step 3 Extract one module into a microservice. Client | v API Gateway |------> Order Service | v Monolith Step 4 Gradually move more modules. Client | v API Gateway |------> Order Service |------> Payment Service |------> Inventory Service | v Monolith Step 5 Eventually remove the monolith completely. Example Consider an e-commerce application. Initially, everything exists inside one application: Monolith ├── Orders ├── Payments ├── Inventory └── Users Over time: Orders become Order Service Payme

Google's demo reel for Gemini Omni looks effortless: ask for a video, then keep talking to it until the shot is right. The question for developers is whether that conversational loop holds up outside a stage demo — and what it actually changes versus the Veo workflow it replaces. What Does Omni Add That Veo Couldn't? Omni's core addition is state. Veo produced one-shot renders — each prompt generated a fresh clip with no memory of the last. Gemini Omni holds context across turns, so changing the camera angle on turn three preserves the characters and lighting established on turn one without restarting the scene . Announced at Google I/O on May 19, 2026, the first shipped model, Gemini Omni Flash, replaces Veo as the video-generation surface in the Gemini app . Product director Nicole Brichtova framed it as "the next step towards combining the intelligence of Gemini with the rendering capabilities of our media models" — DeepMind's informal pitch is a "Nano Banana for video," extending conversational image editing to motion footage. Two claims deserve a skeptical read. Google advertises "intuitive understanding of forces like gravity, kinetic energy, and fluid dynamics," but those physics behaviors currently rest on Google demos and creator footage, with no third-party benchmarks published at launch . And on raw output, independent reviewers put Omni's generation quality on par with Veo 3.1 rather than clearly above it . The differentiation is the iterative editing loop and Gemini-grounded reasoning — not a new render engine. Before Starting: Paid Membership, Region, Age Omni access is gated behind a paid Google AI plan and a few hard eligibility rules, so confirm these before you open a prompt. Gemini Omni Flash unlocks in the Gemini app and Google Flow for Google AI Plus, Pro, and Ultra subscribers, with Plus starting at $7.99/month . If you want to test it for free, generation is available at no cost on YouTube Shorts and the YouTube Create App at launch . Two cons

NVIDIA's Parakeet speech models used to mean a Python stack: NeMo, PyTorch, and a GPU you kept warm. A new C++ port collapses that to one binary and one file. From NeMo to GGUF: What the Port Covers parakeet.cpp is a C++17 inference port that runs NVIDIA's Parakeet automatic speech recognition (ASR) models on the ggml tensor library — the same engine behind whisper.cpp and llama.cpp — with no Python, no NeMo, and no ONNX at inference time. The project is maintained by Ettore Di Giacinto (@mudler), author of LocalAI, and its first tagged release, v0.1.0, landed on May 30, 2026 . The code is MIT-licensed; the model weights keep their original NVIDIA Parakeet licenses. This is a community project, not an official NVIDIA release. The port covers the offline Parakeet families — CTC, RNNT, TDT and hybrid TDT-CTC — in 110M, 0.6B and 1.1B sizes, plus a streaming 120M model with end-of-utterance detection . Two checkpoints anchor most use: parakeet-tdt-0.6b-v2 , the English default that reports 6.05% average WER on the Hugging Face Open ASR Leaderboard and was released 05/01/2025 , and parakeet-tdt-0.6b-v3 , which extends the same 600M FastConformer-TDT architecture to 25 European languages with automatic language detection, released 08/14/2025 . Inference runs on CPU, CUDA, HIP (AMD ROCm), Vulkan and Metal (Apple Silicon) — the same ggml backend matrix as whisper.cpp and llama.cpp — so deployment reduces to one binary plus one GGUF file . The hard part was mapping Parakeet's RNNT/TDT decoders onto a static-graph tensor library. An earlier work-in-progress port surfaced on Hacker News in mid-2025, where the author flagged how far there was to go: "The GGML build is roughly 1000x slower than the MLX Python version" — jason-ni, reporting an early Parakeet-on-ggml experiment (source: Hacker News, 2025 ; see also the jason-ni port ). The mudler release is the matured answer to that decoder-on-static-graph problem, and as of June 2026 Parakeet is not yet merged into mainline whis

This is a submission for the June Solstice Game Jam It's interesting how the most exciting ideas always arrive when I have basically no time to work on them. A few weeks earlier, I had finished my submission for the GitHub challenge by bringing an old WinForms game back to life. That project turned out to be a lot of fun. Then Sylwia Laskowska published a great article about Google's WebMCP . The idea fascinated me, but I wasn't sure where I could actually use it. Then the June Solstice Game Jam was announced. The idea hit me like a lightning bolt: What if I made a game that both humans and AI could play? Let's do it. What I Built I created a puzzle game with a solstice theme called Tower Before Dusk . The goal is simple: reach your home tower before sundown. Every action costs time. Every step brings sunset a little closer. Rivers block your path, rocks force detours and the only way across water is to collect enough wood and build bridges. Move too much, collect unnecessary resources, or choose the wrong path, and night will arrive before you make it home. The challenge isn't just solving the puzzle. It's solving it efficiently. And apparently, that's difficult for both humans and AI. Video Demo In this demo, Gemini 3.1 Flash-Lite tries to solve the level using the exposed game tools. It fails, then I restart the level and solve it manually. That failure is part of the point: the tools worked, but reasoning through the puzzle was still hard for the lightweight model. tower-before-dusk.gramli.workers.dev Code Gramli / tower-before-dusk A TypeScript puzzle game demonstrating WebMCP, where humans and AI solve the same challenges under the same rules. Tower Before Dusk Tower Before Dusk is a tile-based puzzle game about reaching the tower before sunset. Plan each route carefully: every move spends daylight, trees provide wood, and water can only be crossed by building bridges. The game is built as a modern browser app with TypeScript, HTML canvas, and Vite. It also ex

If you've ever felt like LangChain was too heavy, you're not alone. The dependency tree is enormous. Abstraction layers pile up. At some point you lose track of what's actually happening underneath. That frustration has pushed a lot of people toward lighter alternatives — frameworks that prove you can build a capable agent without a hundred transitive dependencies. Agno is one of those alternatives. It started as Phidata and rebranded in early 2025. I spent an afternoon installing Agno v2.6.17 in a clean sandbox and running through Calculator tools, Wikipedia retrieval, Pydantic structured output, and a two-agent Team. I'll share the real execution logs and, more importantly, the traps I hit that the docs don't warn you about. What Agno Is and Where It Came from Phidata built a solid reputation as "the Python framework for AI assistants." When it rebranded to Agno in 2025, the design philosophy got articulated more clearly around three ideas. Model-agnostic from day one. Over 70 LLMs — OpenAI, Anthropic, Google, Ollama, Cohere — can plug in with the same code structure. Swap the model, keep the agent logic. Multimodal as a default. Text, image, audio, video agents all use the same API surface. You don't need a different abstraction layer for each modality. Multi-agent orchestration as a first-class citizen. The Team class is built in. You can switch between coordinate , route , and collaborate modes with a single parameter change. Reading that, I thought: "How is this different from LangChain?" The answer showed up when I actually wrote code. Agno favors composition over class inheritance. One agent takes about 6 lines to set up. There's far less boilerplate to wade through. Installation: No Dependency Hell pip install agno google-genai ddgs wikipedia The agno package installs just the core. Tools require their own extra dependencies — wikipedia for the Wikipedia tool, google-genai for Gemini. This lazy-loading approach keeps the base install clean. $ python3 -c "im

Google Open Knowledge Format: Why Enterprise Agents Need a Knowledge Layer, Not Just More Tools Most enterprise AI conversations still start in the wrong place. They start with the model. Which model should we use? Which framework should we adopt? Which vendor has the best agent platform? Which tools should we connect next? These are fair questions. But in real enterprise architecture, they are not the hardest questions. The harder question is this: Can our AI systems actually understand how our business works? That is why Google Cloud’s article on Open Knowledge Format caught my attention. The article talks about a simple but important idea: representing knowledge in a way that humans can read and machines can use. In OKF, that means markdown for the content and structured metadata for context. At first glance, that may sound too simple. But that simplicity is the point. Enterprises do not need another place where knowledge goes to die. We already have enough portals, catalogs, wikis, dashboards, folders, and internal tools. What we need is a practical way to package knowledge so it can be reviewed, versioned, governed, searched, and reused by both people and AI agents. That is where this idea becomes very relevant for agentic AI. The Real Enterprise AI Problem Most organizations already have the knowledge their AI agents need. They have it in databases, dashboards, tickets, architecture notes, runbooks, Confluence pages, data catalogs, code comments, incident reports, old project documents, and the heads of experienced employees. The issue is not that knowledge does not exist. The issue is that it is fragmented. Some of it is outdated. Some of it is duplicated. Some of it is tribal. Some of it is locked inside tools. Some of it is written for humans but not structured enough for AI systems to use reliably. This becomes a serious problem when we move from AI assistants to AI agents. An assistant can give a helpful answer. An agent does more. It plans, selects tools

🗓️ This Week While organizing ideas for my first iOS app , I remembered an old web app idea called ToneDrill , which I had casually built before to help practice note names on a guitar fretboard🎸. I decided to try turning it into an iOS app 🛠️. I clarified the purpose of ToneDrill , its minimum requirements , and its core features , then organized them in Notion 📝. I was curious to see how well Codex could implement an iOS app from those minimum requirements , so I gave it a try right away💡. I reviewed the SwiftUI code generated by Codex and worked through the app logic to understand how it was implemented 🔍. For now, I was able to create a working app, which felt like a meaningful step forward 🚶. I created the top page UI design for my portfolio website in Figma 🎨. I focused on keeping the structure simple and implementation-friendly, and designed the UI with reusable components for each major part. Based on what I learned from my previous failed attempt, I tried again to see how well Codex could implement a prototype from the Figma UI design (You can read about my previous attempt that didn’t go so well here😅.) Worked on the AI Threat Modelling room from the AI Security Learning Path on TryHackMe this week🤖. 📱 iOS (SwiftUI) Revisited an old web app idea called ToneDrill, which I had previously built casually as a guitar note-training app, and considered turning it into an iOS app. Organized the app idea in Notion, including its purpose, target use case, minimum requirements, and core features. Decided to aim for an MVP-level version first, instead of trying to build a fully featured app from the beginning. Wrote down simple requirements and tested how accurately Codex could implement the initial version of the app. Reviewed the iOS app implementation generated by Codex and examined the code in detail to understand how the logic worked. 🌐 Web Development Posted my weekly dev log on Dev.to📝 Completed the top page UI design for my portfolio website in Figma. Tried us

Peter Parker to Bruce Banner: "I didn't know you could get that big."

𝗦𝗮𝘆 𝗵𝗲𝗹𝗹𝗼 𝘁𝗼 𝗟𝘂𝗺𝗼𝗿𝗮 — 𝗧𝗵𝗲 𝗨𝗹𝘁𝗶𝗺𝗮𝘁𝗲 𝗕𝗼𝗼𝘁𝘀𝘁𝗿𝗮𝗽 𝟱 𝗔𝗱𝗺𝗶𝗻 𝗗𝗮𝘀𝗵𝗯𝗼𝗮𝗿𝗱 𝗨𝗜 𝗞𝗶𝘁. 💎 🔗 𝗚𝗶𝘁𝗛𝘂𝗯 𝗥𝗲𝗽𝗼: https://github.com/Chetankumar-Akarte/lumora 🔗 Demo: https://renukatechnologies.in/demo/lumora/ Don't forgot to 🤩 Star and 👉 Fork the Repo 𝗟𝘂𝗺𝗼𝗿𝗮 is a modern, responsive 𝗕𝗼𝗼𝘁𝘀𝘁𝗿𝗮𝗽 𝟱 𝗔𝗱𝗺𝗶𝗻 𝗗𝗮𝘀𝗵𝗯𝗼𝗮𝗿𝗱 𝗨𝗜 𝗞𝗶𝘁 designed for teams that need a polished, enterprise-ready control center without the bloat. Whether you are building for SaaS, CRM, E-commerce, or internal analytics, Lumora provides a scalable, token-driven foundation to speed up your workflow. 𝗟𝘂𝗺𝗼𝗿𝗮 is the result: a complete admin ecosystem featuring everything from KPI blocks and ApexCharts to full E-commerce management flows and authentication screens. 𝗪𝗵𝗮𝘁’𝘀 𝗶𝗻𝘀𝗶𝗱𝗲? • Full UI Kit with basic and advanced components. • Enterprise pages (Users, Roles, Permissions, Invoices). • Interactive apps like Calendar and Contacts. • Clean, token-driven styling for consistent design. 𝗧𝗲𝗰𝗵 𝗵𝗶𝗴𝗵𝗹𝗶𝗴𝗵𝘁𝘀: • Bootstrap 5.3 • ApexCharts & Chart.js • Vanilla JavaScript • Mobile-first design 𝗞𝗲𝘆 𝗛𝗶𝗴𝗵𝗹𝗶𝗴𝗵𝘁𝘀: • 𝗠𝗼𝗱𝗲𝗿𝗻 𝗧𝗲𝗰𝗵 𝗦𝘁𝗮𝗰𝗸: Built with Bootstrap 5.3, Vanilla JS, and CSS3 using a module-first architecture. • 𝗖𝗼𝗺𝗽𝗿𝗲𝗵𝗲𝗻𝘀𝗶𝘃𝗲 𝗗𝗮𝘀𝗵𝗯𝗼𝗮𝗿𝗱𝘀: Includes layouts for Analytics, CRM, Project Management, HRM, and more. • 𝗙𝗲𝗮𝘁𝘂𝗿𝗲-𝗣𝗮𝗰𝗸𝗲𝗱 𝗔𝗽𝗽𝘀: Ready-to-use interfaces for Advanced Chat, Kanban boards, Email, and File Management. • 𝗗𝗮𝗿𝗸 & 𝗟𝗶𝗴𝗵𝘁 𝗠𝗼𝗱𝗲𝘀: Clean, professional visuals with seamless theme switching. • 𝗗𝗲𝘃𝗲𝗹𝗼𝗽𝗲𝗿 𝗙𝗿𝗶𝗲𝗻𝗱𝗹𝘆: Modular CSS, reusable partials, and organized project structure. I built this to bridge the gap between "pretty" templates and "functional" enterprise tools. Check it out, star the repo, and let me know what you think! I'd love for you to take a look at the code and perhaps even use it for your next project. Feedback and contributions are always welcome! WebDevelopment, Bootstrap5, AdminDashboard, OpenSource, UIUX, JavaScript, GitHub, Bootstrap, CodingCommunity, OpenSourceProject, FrontendDev, LumoraUI

The problem I kept running into I'm a chronic tab hoarder. At any given time I've got 40–80 tabs open across two windows. Chrome's built-in Memory Saver is aggressive in the wrong ways — it hibernates tabs I'm actively referencing. And the built-in task manager is a two-step detour that still doesn't tell me which tabs I should actually close. So I built Tab Memory Manager. What it does Per-tab memory estimates — A live MB count next to every open tab. Sorted by memory usage by default. There's a live total on the toolbar icon so you always know what Chrome is consuming right now. Smart suggestions — The extension flags your biggest, stalest tabs: ones that are idle the longest and consuming the most. It never suggests your active tab, pinned tabs, tabs playing audio, or domains you've whitelisted. Hibernate, don't close — This was the core design decision. Hibernating frees the memory but keeps the tab alive in your strip — it reloads when you click it. Much safer than closing, especially mid-research. Bulk cleanup — Select multiple tabs or hit Apply on the suggestions panel. See the total memory you'll reclaim before you commit. Undo list — Closed something by mistake? There's a "Recently cleaned" panel. One click to restore. Tab grouping — Groups all your open tabs by domain into color-coded Chrome tab groups, instantly. The interesting technical bit: memory estimates Chrome's stable extension API doesn't expose exact per-tab memory. The chrome.processes API that does exists only on Dev and Canary builds — not the Chrome that 99% of people use. So Tab Memory Manager uses calibrated estimates based on tab state, domain patterns, and known Chrome process overhead. These are clearly labeled "est." in the UI. If you're on Dev or Canary, you can switch on real per-tab memory in settings. The warning Chrome shows about "processes requires dev channel" is a Chrome-generated note about that optional API — the extension works completely normally without it. It's not a bug

A trader I was talking to recently said something that stuck with me: "I've blown accounts just from slow fills or missed order cancellations." He was talking about CEX perpetuals. But the problem is identical on Polymarket's CLOB - just measured in seconds instead of milliseconds. My TypeScript bot was averaging 340ms from signal detection to order placement on Polymarket's Central Limit Order Book. On a 5-minute market with a ~2.7-second mispricing window, that's 12% of the entire opportunity window consumed before a single byte hits Polymarket's servers. I was consistently entering at 74¢ when I'd detected the signal at 70¢. The market had already repriced against me. So I rewrote it in Rust. This article documents exactly what I found, what changed, and - critically - what didn't. Background: What My Bot Was Doing If you've read my earlier posts in this series ( architecture , Kelly Criterion sizing , last-60-seconds capture ), you know the context. But the short version: The bot targets Polymarket's 5-minute and 15-minute crypto up/down binary markets (BTC, ETH, XRP, SOL, DOGE, BNB). The strategy is simple: find markets that are briefly mispriced relative to real-time spot momentum, enter at a discount to fair value, hold to resolution. A 5-minute "XRP Up" market priced at 70¢ when spot momentum suggests 82% probability = +12¢ edge per dollar wagered. Do that 50 times a day with disciplined sizing and the math works - if you can actually get filled at the price you detected. The problem: by the time my TypeScript code detected the signal, formatted the order, opened an HTTP connection to Polymarket's CLOB API, waited for TLS handshake, serialized the payload, and received confirmation, the market had often moved to 74-76¢. I was paying for an edge I wasn't capturing. Profiling the TypeScript Bot: Where Was the 340ms Going? Before rewriting anything, I instrumented every stage of the order path. Here's what I found across 500 sampled trades: Stage Average time %