Ethan Lin (@EthanLinAI) header
Ethan Lin (@EthanLinAI) profile picture

AI โ€ข Indie Developer โ€ข U.S. Equity Research & Investing Chinese Divination | I Ching | Astrology | Metaphysics ๐Ÿ”ฅ๐Ÿ”ฅ๐Ÿ”ฅ

New York, USA
Joined May 2018
Photo by Ethan Lin (@EthanLinAI)Photo by Ethan Lin (@EthanLinAI)Photo by Ethan Lin (@EthanLinAI)Photo by Ethan Lin (@EthanLinAI)Photo by Ethan Lin (@EthanLinAI)Photo by Ethan Lin (@EthanLinAI)
I asked Grok Bot to make me a personal website inspired by Minecraft. ๐Ÿงฑ๐Ÿก It's still pretty rough and simple inside: just three desks so far, with the garden and market still under construction. ๐Ÿšง It's my little corner of the internet. I'll keep adding to it, block by block. ๐Ÿ› ๏ธ Come have a look ๐Ÿ‘‡โ€จaiinfoshare.dev #GrokBot #VibeCoding #BuildInPublic
4
6
469
Microsoft finally did something genuinely useful ๐Ÿ˜‚ They just open-sourced an AI tool on GitHub that's ridiculously practical. Turn almost any file into Markdown with one command, so LLMs can read it directly. The list of what it supports is kind of ridiculous: PDF, Word, Excel, PowerPoint Images with OCR Audio with speech transcription YouTube links HTML, CSV, JSON, XML EPUB, ZIP and more Why is this worth bookmarking? Markdown is the format LLMs understand best. Every time you feed a PDF to ChatGPT, Claude or Gemini, or build a RAG pipeline or AI agent, you end up fighting file conversion. MarkItDown fixes that. It keeps headings, lists, tables and links, and outputs clean structured Markdown. No custom parsers, no one-off scripts for every format. It's already at 196K stars for a reason. Python API, CLI, and third-party plugins supported. Install is dead simple: pip install 'markitdown[all]' If you build AI apps, knowledge bases, RAG or LLM workflows, bookmark this one. GitHub: github.com/microsoft/markitdโ€ฆ #AI #LLM #RAG #AIAgent #Python #OpenSource #GitHub
1
7
Your AI agent has the memory of a goldfish ๐ŸŸ This open-source project fixes that. It gives agents and LLM apps real long-term memory. It remembers what matters and forgets the rest. What it can remember is kind of ridiculous: User preferences Past conversations Facts about projects Agent learnings Session context Relationships between entities Why is this worth bookmarking? Right now you either re-explain everything every session or stuff the whole chat history into the context window and burn tokens. Mem0 extracts the important facts automatically and only retrieves what's relevant for each query. Way cheaper, way smarter. 69.3K stars and climbing. Install is dead simple: pip install mem0ai If you build AI agents or chatbots, save this one. GitHub: github.com/mem0ai/mem0 #AIAgents #LLM #Memory #AI #Python #OpenSource #GitHub
2
4
210
Building an AI app used to take a quarter. Now it takes an afternoon โšก This open-source platform lets you go from idea to a production LLM app without stitching ten tools together. Drag, drop, connect a model, ship. What's built in is kind of ridiculous: Visual workflow builder RAG pipelines AI agents with tools Prompt management Support for hundreds of models Logs and monitoring Publish as API, web app or embed Why is this worth bookmarking? Most AI projects drown in glue code: LangChain here, a vector DB there, a custom frontend, and zero visibility once it's live. Dify puts prototyping and production on the same platform, so what you build in the afternoon is what you ship. 161.6K stars. That's a lot of builders. Install: git clone github.com/langgenius/dify cd dify/docker && docker compose up -d If you build AI products, bookmark this one. GitHub: github.com/langgenius/dify #Dify #AI #LLM #AIAgents #RAG #NoCode #OpenSource
3
4
209
Your own private ChatGPT, fully self-hosted ๐Ÿ”’ This open-source project gives you a ChatGPT-style interface that you completely own. Run it on your own machine or server, with your own models and your own files. The features are kind of ridiculous: Chat with local or cloud models RAG over your own documents Built-in web search Multi-user with roles and permissions Model switching mid-conversation Plugins and pipelines Voice input Why is this worth bookmarking? Too many teams are pasting company docs into public chatbots and hoping for the best. Open WebUI pairs with Ollama or any OpenAI-compatible API, so everything stays on your infrastructure, with access control built in. 159.2K stars. Self-hosting AI went mainstream. Install is dead simple: pip install open-webui If you want private AI for yourself or your team, save this one. GitHub: github.com/open-webui/open-wโ€ฆ #OpenWebUI #SelfHosted #AI #LLM #Privacy #OpenSource #GitHub
2
4
231
Hahaha, such a fun design! ๐Ÿ˜„
1
1
126
RIP to every Selenium script I've ever written ๐Ÿชฆ This open-source project lets AI use a browser like a human. Click, type, scroll, fill forms, all on its own. You describe the goal. The agent figures out the clicks. What it can handle is kind of ridiculous: Filling out forms Logging into sites Multi-step checkouts Scraping pages with no API Comparing prices across sites Booking and research tasks Why is this worth bookmarking? Traditional browser automation breaks every time a website changes one button. browser-use reads the page and decides what to do next, so you stop maintaining brittle selectors. Plug in GPT, Claude, Gemini or a local model and let it work. 121.7K stars. The agent crowd has spoken. Install is dead simple: pip install browser-use If you build AI agents or automate anything on the web, bookmark this one. GitHub: github.com/browser-use/browsโ€ฆ #AIAgents #Automation #AI #Python #LLM #OpenSource #GitHub
3
3
327
Teslaโ€™s new-generation FSD V15 autonomous driving system uses AI-generated real-world scenes. This is seriously cool! ๐Ÿš€๐Ÿคฉ
1
1
109
BloombergGPT reportedly cost millions to train. This open-source project does finance LLMs on a budget ๐Ÿ’ธ FinGPT lets you fine-tune your own financial language model for roughly a few hundred dollars. Your own finance LLM, without a bank-sized budget. What it can do is kind of ridiculous: Financial sentiment analysis News and social media analysis Earnings and report reading Market forecasting Robo-advisor style Q&A Fine-tuning on fresh market data Why is this worth bookmarking? Closed financial models are expensive, can't be inspected, and go stale the moment markets move. FinGPT uses lightweight LoRA fine-tuning on open models, so you can retrain cheaply as new data comes in. Open code, open weights, a real research community behind it. 21.6K stars and growing. If you work at the intersection of AI and finance, save this one. GitHub: github.com/AI4Finance-Foundaโ€ฆ #FinGPT #LLM #AI #Finance #FinTech #OpenSource #GitHubVibeRiders
3
2
83
Microsoft open-sourced its AI quant research platform, and most retail traders still haven't touched it. ๐Ÿ“ˆ Qlib has reached 50.5K GitHub stars โ€” and the interesting part isn't the models. It's that the whole quant pipeline lives in one framework. โŒ Gluing together data scripts, notebooks and backtesters โŒ Alpha factors that silently leak future data โŒ Research you can't reproduce a month later โœ… Data processing, factor mining, training and backtesting end to end โœ… Dozens of built-in models from GBDT to Transformers โœ… Portfolio optimization and order execution modules What stands out: ๐Ÿงช Benchmarks you can rerun and compare ๐Ÿค– RD-Agent integration to automate factor research with LLMs โšก A data layer built for fast financial time series github.com/microsoft/qlib
1
1
66
What if your trading desk was 7 AI agents arguing before every trade? ๐Ÿง  TradingAgents has reached 123.4K GitHub stars โ€” and the interesting part isn't the predictions. It's the structured debate before any decision gets made. โŒ One LLM prompt deciding buy or sell โŒ No separation between analysis and risk โŒ Black-box signals you can't audit โœ… Fundamental, sentiment, news and technical analyst agents โœ… Bull vs. bear researchers debate the thesis โœ… A risk team and fund manager approve the final call What stands out: ๐Ÿฆ Modeled on how real trading firms are organized ๐Ÿ” Works with multiple LLM providers ๐Ÿ“œ Every decision leaves a readable reasoning trail github.com/TauricResearch/Trโ€ฆ
2
1
72
Bloomberg Terminal costs $24K a year. This open-source platform gives you most of it for free. ๐Ÿ“Š OpenBB has passed 75.6K GitHub stars, and the interesting part isn't just the free data. It's that an AI copilot can reason over all of it in one place. โŒ Paying for 5 different data subscriptions โŒ Copy-pasting tickers between tabs and spreadsheets โŒ Research notes scattered everywhere โœ… Equities, options, crypto, macro and fixed income in one workspace โœ… Connect your own data sources and APIs โœ… Ask questions about the data in plain English What stands out: ๐Ÿงฉ Extension-based, so you add only the data you need ๐Ÿ Python SDK for building your own quant workflows ๐Ÿ–ฅ๏ธ Run it locally or in the cloud github.com/OpenBB-finance/Opโ€ฆ
1
3
134
Your coding agent doesn't need a bigger model. It needs a methodology โ€” and 323K stars say Superpowers is the one. ๐Ÿฆธ Superpowers is a complete software development methodology for coding agents: 15 composable skills that make your agent brainstorm, plan, test-drive, review, and ship โ€” instead of jumping straight to code. โŒ An agent that writes code before understanding what you want, then rewrites it three times โŒ "Done" with no tests, no review, and a plan that lived only in the model's head โŒ Every session starting from zero โ€” no process, no memory of how good work gets done โœ… Skills trigger automatically: brainstorming fires before any code, and plans break work into 2โ€“5 minute tasks with exact file paths, complete code, and verification steps โœ… Subagent-driven development: a fresh subagent per task, reviewed after each, so the agent can work autonomously for hours without drifting from the plan What makes it different: ๐Ÿง  Brainstorm first: the agent draws out what you actually want and plays it back for correction โ€” small changes agreed in chat, projects get a written design doc ๐Ÿงช True TDD enforced: RED-GREEN-REFACTOR, and code written before tests gets deleted ๐Ÿ” Two-stage review: spec compliance first, then code quality โ€” critical issues block progress ๐Ÿค– 16 harnesses, one install: Claude Code, Codex, Cursor, Gemini CLI, Copilot, Grok, OpenCode, Antigravity, and more ๐Ÿš€ Get started: 1๏ธโƒฃ Claude Code: `/plugin install superpowers@claude-plugins-official` โ€” from Anthropic's official marketplace 2๏ธโƒฃ Or any other harness โ€” Cursor (`/add-plugin superpowers`), Gemini CLI (`gemini extensions install github.com/obra/superpowers`), Codex, OpenCode, Pi, and 10 more 3๏ธโƒฃ Start a fresh session and say "Let's buildโ€ฆ" โ€” brainstorming triggers before a single line of code โš ๏ธ Honest caveat: it's a methodology, not a suggestion โ€” mandatory workflows mean the agent will slow down and ask questions before building, which feels heavy for a quick one-off hack. And the optional visual companion loads a logo from the project's site with your version number (telemetry you can disable with `SUPERPOWERS_DISABLE_TELEMETRY`). โญ 322,979 stars (Nov 21, 2026) ยท MIT #AI #OpenSource #GitHub #ClaudeCode #AIAgents ๐Ÿ”— github.com/obra/superpowers
5
2
201
Not a signal service. Not a trading bot. A Claude skills toolkit that turns investing into a disciplined process โ€” 74 skills. ๐Ÿ“ˆ 7K stars. Claude Trading Skills is a workflow toolkit for time-constrained investors: long-term/ETF/dividend core, disciplined swing trading as satellite โ€” structured review, risk management, and journaling. โŒ Trading on vibes: no regime check, no position sizing, no journal โ€” then wondering why the same mistakes repeat โŒ A signal service that outsources your decisions and teaches you nothing โŒ Tools built for full-time traders when you have 15 minutes a day โœ… Five workflows as YAML manifests: 15-minute daily check, weekly portfolio review, swing opportunities (only when risk allows), trade memory, monthly review โœ… Risk before candidates: exposure ceilings, drawdown circuit breakers, a pre-trade gate that blocks planless or oversized entries What makes it different: ๐Ÿ“Š Regime first: breadth, uptrend participation, distribution days, FTD signals โ€” the daily check ends with an exposure posture for human review, not an order ๐Ÿ’ฐ Kanchi dividend SOP with forced-review triggers, dividend screeners, CANSLIM and VCP swing screeners ๐Ÿง  Memory that compounds: trader-memory-core tracks each thesis from screening idea to closed position with postmortem ๐Ÿ†“ No paid data to start: a five-skill path runs on public CSVs, chart screenshots, and local files ๐Ÿš€ Get started: 1๏ธโƒฃ Claude Code: copy a skill folder to `~/.claude/skills/` โ€” start with the no-API five: market-breadth-analyzer, uptrend-analyzer, position-sizer, trader-memory-core, signal-postmortem 2๏ธโƒฃ Claude Web: download the `.skill` asset from the latest Release, enable Code execution in Settings, upload under Customize > Skills 3๏ธโƒฃ Run the 15-minute daily check: market-regime-daily ends with an exposure posture for your review โ€” not an order โš ๏ธ Honest caveat: educational only โ€” not financial advice, not a signal service, not a broker. It never places orders; human gates stay central. Full data needs FMP (free: 250 req/day), FINVIZ Elite ($39.50/mo), or Alpaca. โญ 6,997 stars (Nov 18, 2026) ยท MIT #AI #OpenSource #GitHub #Trading #Investing ๐Ÿ”— github.com/tradermonty/claudโ€ฆ
4
3
179
Claude Code stops being a chatbot and starts behaving like a senior bug-hunting researcher. 83 skills, 681 disclosed-report patterns. ๐Ÿ› 10.5K stars. Claude-BugHunter is a drop-in skill bundle for bug hunting and external red-team work: 83 skills, 15 slash commands, curated from 681 disclosed HackerOne reports across 24 vulnerability classes. โŒ One big "security" prompt that's too generic to find anything real โŒ 30 bookmarked disclosed reports you re-read every engagement โŒ Drafts written before validation โ€” wasted hours and a hurt validity ratio โœ… Per-class hunt skills with real detection patterns, payloads, bypass tables, and chain templates โ€” the patterns triagers paid for โœ… A 6-phase workflow (recon โ†’ map & rank โ†’ hunt โ†’ validate โ†’ report) with scope enforced in code and a 7-Question Gate before submission What makes it different: ๐ŸŽฏ Auto-loaded by topic: describe what you're testing and the right skill loads โ€” no invocation by name ๐Ÿข Enterprise perimeter: M365/Entra, Okta, vCenter, SSL-VPN, SharePoint โ€” current 2024โ€“2026 CVE chains plus post-credential escalation ๐Ÿ“ Ship it: VRT-aware severity, OOS rebuttals, PII redaction, red-team deliverables โ€” plus a `cbh` CLI runner ๐Ÿค– Five harnesses, one install: Claude Code, OpenCode, Codex CLI, Hermes, Antigravity (`bash scripts/install.sh --all --burp-mcp`) ๐Ÿš€ Get started: 1๏ธโƒฃ Install as a plugin (recommended): `/plugin marketplace add elementalsouls/Claude-BugHunter`, then `/plugin install claude-bughunter@elementalsouls` 2๏ธโƒฃ Or copy-install: `git clone github.com/elementalsouls/Clโ€ฆ`, then `bash scripts/install.sh` 3๏ธโƒฃ Describe your target in plain English โ€” "Testing acme.com, an in-scope HackerOne target. Run recon and rank the surface." โš ๏ธ Honest caveat: authorized use only โ€” assets you own or have written authorization to assess. Anthropic's safeguards block offensive security by default, so authorized work needs the free Cyber Verification Program; Opus 5 falls back to Opus 4.8 for exploit-generation requests. Internal AD, C2, and post-exploit are out of scope by design. โญ 10,515 stars (Nov 18, 2026) ยท MIT (code) / CC BY 4.0 (content) #AI #OpenSource #GitHub #BugBounty #Security ๐Ÿ”— github.com/elementalsouls/Clโ€ฆ
1
2
194
Andrew Ng's OpenWorker puts an AI coworker on your desktop โ€” with approvals built into the workflow. ๐Ÿ–ฅ๏ธ 19.4K GitHub stars. A local-first desktop app, plus a Harness SDK for building your own agents. โŒ An agent disconnected from your working folders โŒ Unclear permissions and no record of what happened โŒ Being tied to a single model provider โœ… Work from local folders โœ… Review tool actions through an approval system โœ… Choose a cloud provider or connect a local model What stands out: ๐Ÿ—‚๏ธ Conversations, keys, and connector tokens are stored locally ๐Ÿ”Œ Supports OpenAI, Anthropic, Gemini, OpenRouter, and local serving options such as Ollama ๐Ÿ›ก๏ธ Records tool calls; designated dangerous operations still require human approval in auto-approve mode ๐Ÿ“ฆ Offers sandbox options, including NVIDIA OpenShell ๐Ÿ’ฌ A Slack integration lets tasks run on your desktop and return results to the thread ๐Ÿงฐ MCP support for adding tools ๐Ÿš€ Get started: 1๏ธโƒฃ Download the Desktop App from the project's website 2๏ธโƒฃ Configure a model provider or local model 3๏ธโƒฃ Give a session a folder of notes and ask for a short brief. Review the requested actions before approving them. Building your own application? npm install @openworkerai/harness-sdk โš ๏ธ The Desktop App is in open beta. It supports Apple Silicon Macs on macOS 12+ and Windows 10/11 x64. Windows builds aren't yet code-signed. Local storage doesn't mean local inferenceโ€”cloud models still receive the content you send them. โญ 19,469 stars ยท Nov 20, 2026 ยท MIT #AI #OpenSource #GitHub #AIAgents #AndrewNg ๐Ÿ”— github.com/andrewyng/openworโ€ฆ
6
4
193
OpenAI just shipped the security half of Codex: find it, validate it, patch it โ€” from your terminal. ๐Ÿ”’ 14K stars. Codex Security is OpenAI's CLI + TypeScript SDK for finding, validating, and fixing security vulnerabilities โ€” in Codex, in your terminal, or in CI. โŒ Scanners that dump 500 findings and leave you to sort signal from noise โŒ A vuln report with no patch, no verification, and no owner โŒ Security reviews that happen once a quarter, if at all โœ… The full loop: scan, validate candidate findings, generate patches, verify existing fixes โ€” with source evidence and coverage limits in the report โœ… Deep scans repeat Standard scans and return deduplicated findings; findings dedupe across runs What makes it different: ๐ŸŽฏ Scoped scans: whole repos, selected paths, or just the diff (`cs scan . --diff origin/main`) โ€” plus an advisory 15-minute investigation budget ๐Ÿ“‹ Policy built in: drafts `SECURITY.md` and saves threat models for later review (`cs policy .`) ๐Ÿ”— Fits your stack: Linear, Jira, GitHub tracking; SARIF/JSON/CSV export; a GitHub Action for weekly scans ๐ŸŒ Not OpenAI-only: scans run on OpenAI, Bedrock, OpenRouter, Fireworks, and Cheaper Inference ๐Ÿš€ Get started: 1๏ธโƒฃ Install and sign in: `npm install --global @openai/codex-security`, then `cs login` 2๏ธโƒฃ Scan your repo: `cs scan .` โ€” open the Markdown report at the REPORT path when it finishes 3๏ธโƒฃ Go deeper: `cs scan . --mode deep`, or scan just your changes with `cs scan . --diff origin/main` โš ๏ธ Honest caveat: scans cost money โ€” the GitHub Action example uses an estimated $5 limit per scan, and requests in progress can exceed it. You need a login or API key, and some cybersecurity requests require Trusted Access for Cyber. Requires Node 22.13+/24.x/26.x and Python 3.10+. โญ 14,046 stars (Nov 18, 2026) ยท Apache-2.0 #AI #OpenSource #GitHub #Security #Codex ๐Ÿ”— github.com/openai/codex-secuโ€ฆ
1
3
271
ๅฐฑ้—ฎไฝ ๆœ‰ๆฒกๆœ‰็ซๅญ—๏ผŒ็ญ็š„ๆ˜ฏไธๆ˜ฏ็ซ๐Ÿคฃ
ใ€ŠๅŸบไบŽ็บฏ่ง†่ง‰่ฏ†ๅˆซ็š„AI็ญ็ซ็ณป็ปŸใ€‹
1
37
Anthropic's first day as a public company was pure chaos. ๐Ÿ“‰ Priced at $83, opened at $111.50, ran to $115 by 11:45 AM ET โ€” $2T valuation, biggest IPO ever, all the headlines. Then it just... fell apart. Closed at $97, after-hours $92.60, and the perp's at ~$91 as I write this. $111 for a company that lost $104B in nine months was always a tough sell. The market figured that out in about four hours. I shorted the pop and took profit at $102.74. Too early, as usual. The fade was the whole trade and I left halfway through it.๐Ÿ˜ $ANT
1
47
Your agent can search Reddit, X, YouTube, HN, and Polymarket in one research workflow. ๐Ÿ”Ž last30days-skill has reached 70.8K GitHub stars โ€” and the interesting part isn't just searching more sites. It's bringing the findings together into a cited brief. โŒ Opening a dozen tabs to see what people are saying โŒ Reading the same story on five different platforms โŒ Getting a link dump with no supporting comments โœ… Search multiple sources in parallel โœ… Merge overlapping stories into one brief โœ… Include community signals and supporting quotes What stands out: ๐Ÿ’ฌ Reddit comments, not just post titles ๐Ÿ—ณ๏ธ Signals such as upvotes, likes, and prediction-market odds ๐Ÿ—‚๏ธ Briefs saved locally, by default in ~/Documents/Last30Days/ ๐Ÿ”Œ Installation options for Claude Code and other Agent Skills-compatible tools ๐Ÿš€ Claude Code setup: 1๏ธโƒฃ /plugin marketplace add mvanhorn/last30days-skill 2๏ธโƒฃ /plugin install last30days 3๏ธโƒฃ /last30days <your topic> For other supported agents: npx skills add mvanhorn/last30days-skill -g โš ๏ธ Reddit, HN, Polymarket, and GitHub work without configuration. Other sources, including X and YouTube, need additional setup. Popularity and market odds are useful signalsโ€”not proof that a claim is true. โญ 70,827 stars ยท Nov 20, 2026 ยท MIT #AI #OpenSource #GitHub #AIAgents #ClaudeCode ๐Ÿ”— github.com/mvanhorn/last30daโ€ฆ
2
5
190