AI 资讯

Your ads, managed by AI that gets smarter daily. Discussion | Link

A process that took 2–3 weeks and produced a 40-page document is the kind of thing everyone wants to "AI-ify." Here's what most people get wrong about it: they reach for the LLM first. In regulated lending, the LLM is the last and smallest part of the pipeline. The hard parts are: ingesting messy documents reliably, extracting data into a canonical model, validating and cross-reconciling it, and doing the actual financial calculations - and those calculations must be done in deterministic code, never by a language model that might quietly round a number wrong. The LLM drafts prose. Code does the math. A human owns the risk decision. That ordering isn't a limitation - under OSFI E-21 it's the only design that's allowed to exist. The teams that win with AI in banking aren't the ones using the biggest model. They're the ones who know which 10% of the problem the model should touch. Where have you seen an LLM used for something that should have been deterministic code? FinancialServices #AI #Lending

EDRSR — the Unified State Register of Court Decisions — is effectively all of Ukraine's judicial practice in open access. Today Qdrant holds **44M+ vectors : criminal (19M), civil (14.3M), commercial (5.1M), misdemeanors (5.6M). Vectorization of civil cases (CPC, justice_kind=1) — the largest cohort at 33.7M documents — runs on a dedicated EC2 instance (r6a.xlarge, 32 GB RAM, 2 TB gp3). Here's what's under the hood: models, pipeline, cost, rakes, and current status. Why Vectorize Courts When a lawyer searches "is there case law on recovering bank prepayment fees" — they don't want to open 40 decisions and read them through. They want the system to surface the top 5 most relevant ones, pull out key paragraphs, and show how courts reasoned. Full-text search (FTS) over keywords doesn't give that — it returns every document containing the word "fee", and there are thousands. For this semantic task you need vector representations of text. The model turns a paragraph from a decision into a point in a 1024-dimensional space; semantically similar paragraphs sit near each other. A kNN search in Qdrant returns the top K nearest, and an LLM composes the answer from exactly those relevant fragments. The only problem: the register is big. Very big. Scale Our prod database holds full texts of decisions starting from 2006. Breakdown by procedural type: Civil (CPC) — 33.7M documents. The largest category. Consumer, housing, labor, family. Criminal (CrPC) — 12M+ Administrative (CAS) — 14M+ Commercial (CC) — 6M+ Misdemeanors (CUaP) — 6M+ The Qdrant collection edrsr_decisions on a dedicated EC2 currently holds 44M+ vectors (122 segments, on_disk=true): | Proceeding type | justice_kind | Vectors | |—|—|—| | Criminal (CrPC) | 2 | 19,036,347 | | Civil (CPC) | 1 | 14,328,427 | | Misdemeanors (CUaP) | 5 | 5,579,432 | | Commercial (CC) | 3 | 5,098,662 | | Total | | 44,042,868 | Civil cases processed: 14.3M out of 33.7M — that's 42%. After CPC completes there will be roughly 63M+ vectors in

In production we have ~1.5 TB of full-text court decisions and their vector embeddings, plus another ~550 GB of other legal data: registries, legislation, business entities, a Spanish case law corpus, EU-Lex. If we take this corpus and train an MoE model the size of DeepSeek V3, scaled to 860B parameters, on GCP — what comes out? We break down the dataset, architecture, compute cost, and the properties such a model would have on Ukrainian law. What's in the Dataset The entire corpus is what's already running in SecondLayer's production. No extra scrapes, no Common Crawl, no noise. EDRSR — the dataset core, ~1.5 TB. The Unified State Register of Court Decisions of Ukraine. 96.2 million full-text decisions (1,079 GB in PostgreSQL TOAST), 471 GB of vectors in Qdrant (voyage-3.5, 1024-dim), 28 GB of metadata (court, judge, date, case category, proceeding type, statute code). Breakdown by jurisdiction: civil 33.7M, administrative 14M+, criminal 12M+, commercial 6M+, misdemeanors 6M+. Largest annual cohort — 2024 (115 GB of TOAST text). OpenReyestr — 43 GB. Ukrainian public registries: 16.7M legal entities (EDR), ownership structures (beneficiaries, shareholders), debtors (State Enforcement Service), NAIS registries. This is the foundation for SneakyPiper — our due-diligence platform — but here it serves as raw corpus for the model. Legislation — ~40 GB. The Constitution, major codes (Civil, Criminal, Criminal Procedure, Civil Procedure, Commercial Procedure, Administrative Procedure, Labor, Tax, Customs), laws, and secondary legislation. All structurally annotated: articles, parts, clauses, revision dates with effective-date tracking. This isn't flat text: we know that Article 124 of the Constitution took effect on a specific date, carries particular references, and is cited in a precise number of decisions. Supreme Court review practices + lu_court_decisions — ~25 GB. SC plenary decisions, practice overviews, Grand Chamber rulings. This is the most valuable slice — the

Local LLM Deployment, Agent Handbook, & LLM Cost Reduction: Applied AI Workflows Today's Highlights This week's highlights cover practical guides for running state-of-the-art LLMs locally and building AI agents, alongside an innovative technique to significantly cut LLM API costs for code processing. These resources focus on actionable insights and frameworks for real-world AI application development. Jamesob's guide to running SOTA LLMs locally (Hacker News) Source: https://github.com/jamesob/local-llm This GitHub repository provides a comprehensive, hands-on guide for setting up and running state-of-the-art Large Language Models (LLMs) on local hardware. It meticulously covers the necessary tooling, dependencies, and configuration steps required to get various open-source LLMs operational without relying on cloud APIs. The guide emphasizes practical considerations for local inference, including hardware requirements, model quantization techniques, and performance optimization for different architectures, directly addressing production deployment patterns. It serves as an invaluable resource for developers and researchers looking to experiment with LLMs, develop applications offline, or reduce costs associated with cloud-based inference by leveraging local compute. The guide offers concrete details and actionable steps, making it an essential resource for anyone aiming to implement LLMs in a controlled, private, or cost-effective environment. Comment: This guide is fantastic for anyone wanting to get serious about local LLM development. It covers the nitty-gritty details of setting up your environment and getting models like Llama-3 running efficiently on consumer hardware, which is crucial for privacy and cost savings. 60% Fable cost cut by converting code to images and having the model OCR it (Hacker News) Source: https://github.com/teamchong/pxpipe The pxpipe project introduces an innovative technique to drastically reduce API costs when processing code with Lar

One quiet handwritten page a day. No feed, no AI. Discussion | Link

I'm building Guildly, a Slack-like interface where you can run a company of AI employees. I recently used Fable to revamp the whole website into a cool retro pixel style. Comments URL: https://news.ycombinator.com/item?id=48780222 Points: 5 # Comments: 4

I build and run one of the tools on this list (AGenO — full disclosure below), and I use every other tool here regularly. This is what "free" actually gets you on each one, including the catches. The AI tool landscape has a dirty secret: almost nothing labeled "free" is free. Most tools give you a taste — ten messages, three images, one song — and then the paywall lands. So instead of another list of forty tools nobody has tried, here are nine that give you real value at $0, organized by what you're trying to make, with the actual limits spelled out. Quick comparison Tool Best for What's actually free The catch ChatGPT General chat & writing ~10 msgs/5h on the flagship model Silently switches you to a weaker model after the limit Claude Long documents, nuanced writing 10–25 msgs/5h, varies with demand Limits shrink when servers are busy Gemini Image generation & editing Generous with a Google account Best features drift to the paid tier Perplexity Research with citations Unlimited basic searches Pro searches are capped Suno AI music ~10 songs/day No commercial use on free; failed generations can eat credits Leonardo AI Stylized art & game assets Daily token allowance Confusing token system; images are public on free Character.AI Roleplay & AI characters Unlimited chat Heavy filters; your chats train their models AGenO All of it in one place Images, songs with vocals, chat, characters, stories, coding problems — daily free allowance One-person project — busy hours can mean a short queue Canva Magic tools Quick social graphics 50 text-to-image uses Design-tool add-on, not a real generator Chat and writing ChatGPT is still the default for a reason — the free tier includes the flagship model and it's good at nearly everything. The catch nobody tells you about: after roughly ten messages in five hours, it quietly downgrades you to a mini model without making it obvious. If your answers suddenly get dumber mid-conversation, that's why. Claude writes the most natural prose

Google Trends has no official API, and most wrapper libraries rot within months. But the Trends site itself runs on a keyless JSON API that anyone can call, and it serves the exact numbers you see in the UI. Here is the full recipe, including one gotcha where Google quietly labels your session a scraper. The two step flow Trends works in two steps. First you call explore , which returns a list of widgets, one per chart on the page, each with a signed token: GET https://trends.google.com/trends/api/explore ?hl=en-US&tz=0 &req={"comparisonItem":[{"keyword":"web scraping","geo":"US","time":"today 12-m"}],"category":0,"property":""} Then you call a widget data endpoint with that widget's request and token : GET https://trends.google.com/trends/api/widgetdata/multiline?hl=en-US&tz=0&req=<widget.request>&token=<widget.token> The widget kinds map to endpoints: TIMESERIES uses multiline , GEO_MAP uses comparedgeo , and both RELATED_QUERIES and RELATED_TOPICS use relatedsearches . The cookie trick Call explore cold and you get a 429. The API wants a NID cookie, and here is the counterintuitive part: you get it by requesting the public explore page first, and that page may itself respond 429 while still setting the cookie you need. const res = await fetch ( ' https://trends.google.com/trends/explore?geo=US&q=test ' ); // res.status may be 429. The Set-Cookie header is still there. const cookies = res . headers . getSetCookie (); Grab the cookie from the 429 response, retry explore , and everything works. Strip the anti JSON prefix Every Trends response starts with a junk line like )]}' to break naive JSON.parse calls. Drop everything up to the first newline: const body = await res . text (); const data = JSON . parse ( body . slice ( body . indexOf ( ' \n ' ) + 1 )); The scraper flag Here is the part I have not seen documented. Look inside the widget request object that explore returns to a keyless session: "userConfig" : { "userType" : "USER_TYPE_SCRAPER" } Google knows. And

It happened during a quiet, solemn moment in a community prayer hall. I was sitting in the third row, reflecting, when suddenly, a high-pitched ringtone shattered the silence. It wasn't my phone, but the ripple effect of embarrassment was immediate. Everyone looked around, shifting uncomfortably. That collective tension is something we have all felt—a moment of human error that technology should have intercepted. I looked at my own device, feeling the familiar anxiety of whether I had remembered to flip the physical silent switch. It was then that I decided to stop relying on my own memory. We live in an era of hyper-connected devices, yet the most basic context-awareness—knowing where we are and how our phone should behave—remains manual. I found myself constantly toggling between 'Normal' and 'Silent' modes at the library, the office, and the gym. If I forgot, I was the person disrupting a meeting. If I remembered to mute it, I inevitably forgot to unmute it, missing urgent calls from family for hours. The existing solutions were either too bloated, requiring invasive cloud permissions, or they simply failed to trigger reliably when the screen was off. I needed a solution that was local, predictable, and battery-conscious. Building Muffle started with the realization that I had to master the GeofencingClient API without draining the user's battery. The primary challenge wasn't just triggering an event; it was doing so while the device was in a deep sleep state. I initially experimented with a standard LocationManager approach, polling GPS coordinates at set intervals. That was a disaster. It kept the radio active, pinged the GPS satellites constantly, and decimated the battery life in under four hours. It was an immediate non-starter for a production-ready application. I pivoted to the Geofencing API provided by Google Play Services, which leverages the fused location provider. This is significantly more efficient because the system handles the batching and hardwa

Hi everyone, I want to share textparser, a high-performance, lightweight text parsing and AST generation library written in pure C. 💡 The Core Idea & Why It's DifferentTraditional parser generators (like Flex/Bison) come with a steep learning curve and rigid code generation. On the other end, hand-writing recursive descent parsers or state machines becomes an unmaintainable mess as your language grows.textparser bridges this gap using a hybrid JSON + Python + C workflow:You define your tokens, syntax patterns, and color styles in a clean, human-readable JSON grammar file. A lightweight Python compiler tool processes the JSON, runs optimizations, and emits a dense, static C header array.The C runtime engine loads these pre-compiled arrays instantly. It processes raw text strings into an abstract token tree (textparser_token_item) using a highly optimized regular expression engine (crpe2).By offloading the grammar overhead and heavy state-machine parsing logic to the Python build step, the actual runtime C library stays incredibly lean, memory-efficient, and fast. 🚀 Features At A Glance30+ Languages Out-of-the-Box: Includes ready-to-use JSON grammars for C, C++, Rust, Python, JavaScript, HTML, SQL, and dozens more.Rich Token Metadata: Every parsed token tracks exact code coordinates, structural flags, and custom syntax styling options.Zero Bloat: Ideal for terminal text editors, syntax highlighters, custom linters, and lightweight static analysis tools where bringing in a massive compiler front-end is overkill.

The rise of AI has brought an avalanche of new terms and slang. Here is a glossary with definitions of some of the most important words and phrases you might encounter.