https://chengyu.eu/tags/ai/Recent content in AI on Chengyu WangHugo -- gohugo.ioen© 2026 ChengyuTue, 18 Aug 2026 18:00:12 +0000https://chengyu.eu/posts/podcast-nas-ai-transcription-system/Tue, 18 Aug 2026 18:00:12 +0000https://chengyu.eu/posts/podcast-nas-ai-transcription-system/Rather than trying to run Whisper locally on an underpowered NAS, I offloaded transcription to Cloudflare Workers AI and article cleanup to LiteLLM, turning the NAS into an automation hub with its own web player for listening, searching, and reading.https://chengyu.eu/posts/ontology-reasoning-explained/Fri, 14 Aug 2026 16:00:24 +0000https://chengyu.eu/posts/ontology-reasoning-explained/Databases only know what you explicitly store; ontological reasoning lets a machine derive thousands of unstated facts from a handful of concept definitions — powerful, expensive to build, and increasingly paired with LLMs to keep each other’s weaknesses in check.https://chengyu.eu/posts/dropbox-to-seafile-nas-knowledge-base/Mon, 15 Jun 2026 08:00:48 +0000https://chengyu.eu/posts/dropbox-to-seafile-nas-knowledge-base/After bringing photos home to Immich, files were next — Seafile replaced Dropbox for syncing, then got paired with Open WebUI, RAG, and LiteLLM to turn a plain file store into a searchable, AI-queryable knowledge base.https://chengyu.eu/posts/hacker-news-coding-model-reputation/Mon, 04 May 2026 22:00:09 +0000https://chengyu.eu/posts/hacker-news-coding-model-reputation/A site that runs AI sentiment analysis on real Hacker News developer comments, rather than benchmark scores — Claude comes out on top for engineering sense, GPT close behind, with DeepSeek a notable dark horse.https://chengyu.eu/posts/trendradar-information-feed-hub/Mon, 04 May 2026 21:00:13 +0000https://chengyu.eu/posts/trendradar-information-feed-hub/Too many English finance and tech RSS feeds to skim by hand — TrendRadar, an open-source aggregator, filters, translates, and summarizes them via AI, then pushes the results to Telegram.https://chengyu.eu/posts/hermes-domain-isolation-optimization/Wed, 29 Apr 2026 10:00:06 +0000https://chengyu.eu/posts/hermes-domain-isolation-optimization/Investing talk kept leaking into work notes and vice versa — the fix was splitting Hermes into a router plus dedicated stock, work, diary, and life sub-assistants, each with its own knowledge base partition and a strict no-cross-referencing prompt.https://chengyu.eu/posts/chatgpt-images-2-first-look/Thu, 23 Apr 2026 22:00:19 +0000https://chengyu.eu/posts/chatgpt-images-2-first-look/Four quick tests of the new ChatGPT Images 2.0: visualizing my current life as the AI understands it, an imagined retirement, a comic-strip take on a 2013 journal entry about orchids, and a PowerPoint-generation test.https://chengyu.eu/posts/trinity-ai-system-weekend-build/Mon, 20 Apr 2026 02:01:20 +0000https://chengyu.eu/posts/trinity-ai-system-weekend-build/A personal assistant on the home NAS, a home assistant tied into an investment system, and an intelligence-gathering agent on a VPS — three modules starting to work together as a real perception-cognition-action loop, not just a pile of tools.https://chengyu.eu/posts/litellm-model-routing-config/Sun, 19 Apr 2026 23:00:11 +0000https://chengyu.eu/posts/litellm-model-routing-config/A full writeup of the LiteLLM gateway config behind my AI stack — free NVIDIA NIM and OpenRouter models as the safety net, cheap paid models like DeepSeek V3 for daily use, and Claude Sonnet as an on-demand top-tier fallback.https://chengyu.eu/posts/hermes-synology-nas-deployment/Sun, 19 Apr 2026 22:00:12 +0000https://chengyu.eu/posts/hermes-synology-nas-deployment/After getting Hermes running on a VPS, I moved it to my Synology NAS instead so it could gradually gain access to household journal and investment data as an internal AI hub — and redesigned the model routing along the way.https://chengyu.eu/posts/hermes-new-ai-assistant-setup/Sun, 19 Apr 2026 01:00:07 +0000https://chengyu.eu/posts/hermes-new-ai-assistant-setup/OpenClaw’s buzz is fading and everyone’s talking about Hermes — its token compression alone (50%+, reportedly up to 70% in some cases) was reason enough to spin up a VPS and try it.https://chengyu.eu/posts/testing-pt-pt-language-models-real-scenario/Sat, 18 Apr 2026 17:00:16 +0000https://chengyu.eu/posts/testing-pt-pt-language-models-real-scenario/Blind-testing two Ollama models against real CIPLE A2 Portuguese-exam questions, the well-documented, benchmark-topping model lost badly to an obscure community fine-tune — because the benchmarks were measuring Brazilian Portuguese, not European.https://chengyu.eu/posts/audit-rules-ai-existing-assets-first/Sat, 18 Apr 2026 10:01:41 +0000https://chengyu.eu/posts/audit-rules-ai-existing-assets-first/Writing the plan for an AI-generated audit-rules project, we chose to reverse-engineer rules from years of proven production scripts rather than rebuilding a rule system from raw data and business logic.https://chengyu.eu/posts/vibevoice-asr-tts-setup-diary/Sat, 18 Apr 2026 10:01:18 +0000https://chengyu.eu/posts/vibevoice-asr-tts-setup-diary/Setting up Microsoft’s VibeVoice to test audio-to-subtitle generation and text-to-speech: the ASR side (multi-speaker SRT generation) turned out production-ready, while TTS — especially Chinese — wasn’t there yet.https://chengyu.eu/posts/is-openclaw-worth-it/Thu, 12 Mar 2026 09:00:22 +0000https://chengyu.eu/posts/is-openclaw-worth-it/An honest review after months of daily use, from when it was still called ClawdBot: genuinely useful for memory, reminders, and lightweight investing research — but with real drift and security caveats.https://chengyu.eu/posts/claude-finance-analysis-nvda-demo/Thu, 26 Feb 2026 18:00:24 +0000https://chengyu.eu/posts/claude-finance-analysis-nvda-demo/Setting up Claude’s Finance Analysis tool through Cowork, then having it produce a full comps-based valuation report on NVIDIA against AMD, Broadcom, Qualcomm, Marvell, and Intel.https://chengyu.eu/posts/nvidia-blackwell-architecture/Thu, 26 Feb 2026 06:00:06 +0000https://chengyu.eu/posts/nvidia-blackwell-architecture/Two full-reticle dies fused into one GPU, FP4 inference, and a fifth-generation NVLink that connects up to 576 GPUs — the architecture behind NVIDIA’s ‘AI factory’ strategy.https://chengyu.eu/posts/openclaw-macos-idea-to-video-pipeline/Thu, 26 Feb 2026 05:00:35 +0000https://chengyu.eu/posts/openclaw-macos-idea-to-video-pipeline/Moving OpenClaw off a Linux VPS and onto a hackintosh, then adding NotebookLM’s CLI to turn a finished blog post into a narrated slide video automatically.https://chengyu.eu/posts/ai-era-wealth-redistribution/Wed, 25 Feb 2026 11:41:00 +0000https://chengyu.eu/posts/ai-era-wealth-redistribution/Reacting to Citrini Research’s viral 2028 doom scenario: technological revolutions have never shrunk the economy, but the gap between people who can afford frontier AI and those who can’t is opening faster than any previous divide.https://chengyu.eu/posts/configuring-blogwatcher/Thu, 12 Feb 2026 15:00:10 +0000https://chengyu.eu/posts/configuring-blogwatcher/Watching an AI-curated information feed surface a YC ’ex technical founder’ listing made one thing click: the real skill now is building systems with AI, not just using it.https://chengyu.eu/posts/clawdbot-obsidian-apple-notes-integration/Thu, 12 Feb 2026 11:00:12 +0000https://chengyu.eu/posts/clawdbot-obsidian-apple-notes-integration/Voice in, an Obsidian note out, a blog post and short video out the other end automatically — plus Apple Notes, Google Sheets, and an email pipeline running hands-free.https://chengyu.eu/posts/three-models-working-for-me/Wed, 11 Feb 2026 09:00:07 +0000https://chengyu.eu/posts/three-models-working-for-me/Codex writes the code fast, Claude Code audits it, Gemini plugs into OpenClaw as a tool, NotebookLM organizes notes, and Nanobanana handles visuals — a small AI production line.https://chengyu.eu/posts/programming-is-just-typing/Sun, 08 Feb 2026 14:00:29 +0000https://chengyu.eu/posts/programming-is-just-typing/My notes on Jensen Huang’s unusually candid remarks at a 2026 AI summit: syntax stopped being the moat, and domain expertise became the whole game.https://chengyu.eu/posts/ai-benchmark-opus-gpt-gemini/Sun, 08 Feb 2026 14:00:22 +0000https://chengyu.eu/posts/ai-benchmark-opus-gpt-gemini/A benchmark chart making the rounds shows large models shifting from chatbots to agents — with Opus 4.6 dominating agentic computer use, GPT-5.2 still king of pure reasoning, and Gemini 3 Pro owning vision and multilingual work.https://chengyu.eu/posts/solo-ai-built-platform-and-app/Sun, 08 Feb 2026 12:00:11 +0000https://chengyu.eu/posts/solo-ai-built-platform-and-app/Codex framed the app, Cursor did the surgical edits, and a working SIP call-center app got built by one person — the rough UI is the actual lesson here.https://chengyu.eu/posts/deleting-fine-tuning-notes/Sun, 08 Feb 2026 03:00:07 +0000https://chengyu.eu/posts/deleting-fine-tuning-notes/Clearing out a year-old Obsidian folder on model fine-tuning made the shift obvious: from chasing model capability to chasing business results through workflow and RAG.https://chengyu.eu/posts/ai-content-workflow-outline-to-slides/Wed, 04 Feb 2026 23:00:38 +0000https://chengyu.eu/posts/ai-content-workflow-outline-to-slides/A three-step pipeline — outline with Gemini, flesh out a transcript, then let NotebookLM turn both into a presentation — that cut a week of course prep down to a day or two.https://chengyu.eu/posts/moltbot-two-days-deploy-configure-use/Fri, 30 Jan 2026 11:00:21 +0000https://chengyu.eu/posts/moltbot-two-days-deploy-configure-use/Watching an agent diagnose its own missed schedule, switch itself from passive to active wake mode, and confirm the fix — without me tracing a single step of it myself.https://chengyu.eu/posts/google-account-suspension-appeal/Thu, 29 Jan 2026 15:00:09 +0000https://chengyu.eu/posts/google-account-suspension-appeal/Registered a new Google account to pay for steadier API access after hitting rate limits on the free tier — and got flagged as bot-created within a day of paying.https://chengyu.eu/posts/clawdbot-2026-vps-deployment/Wed, 28 Jan 2026 21:59:54 +0000https://chengyu.eu/posts/clawdbot-2026-vps-deployment/Working around a broken systemd installer, teaching a model to fetch stock data itself when the plugin store isn’t reachable, and locking the firewall down to SSH-only.https://chengyu.eu/posts/todays-ops-log-codex-network-podsync/Fri, 23 Jan 2026 15:00:18 +0000https://chengyu.eu/posts/todays-ops-log-codex-network-podsync/Moving from Cursor to OpenAI Codex, two days spent simplifying the network layout, a fix for Podsync’s YouTube rate-limiting, and a disposable-container trick for benchmarking a NAS.https://chengyu.eu/posts/claude-skills-vs-ai-workflows/Wed, 21 Jan 2026 10:00:12 +0000https://chengyu.eu/posts/claude-skills-vs-ai-workflows/Why describing what you want in plain language and letting Claude auto-apply a Skill beats another afternoon spent wiring up nodes in Dify, n8n, or Coze.https://chengyu.eu/posts/tpu-vs-gpu-deep-dive/Sun, 18 Jan 2026 17:00:08 +0000https://chengyu.eu/posts/tpu-vs-gpu-deep-dive/GPUs are the flexible all-rounder with a mature ecosystem; TPUs are the specialist that’s brutally efficient at large-scale matrix math — plus a rundown of who actually trains on TPUs.https://chengyu.eu/posts/tesla-accident-ai-physics-analysis/Sun, 18 Jan 2026 15:01:38 +0000https://chengyu.eu/posts/tesla-accident-ai-physics-analysis/Asking Gemini Pro and ChatGPT 5.2 to do an engineering-style read on whether emergency braking was slow to react — and what the physics of a ‘ghost probe’ collision actually allow.https://chengyu.eu/posts/ai-models-traffic-accident-liability/Thu, 15 Jan 2026 18:51:39 +0000https://chengyu.eu/posts/ai-models-traffic-accident-liability/After my own Tesla was hit by a car pulling out from under a bridge, I ran the accident photo past ChatGPT, Gemini, Grok, Copilot, Claude, and a few domestic models to see which ones actually got liability right.https://chengyu.eu/posts/ai-blog-redesign-model-comparison/Wed, 07 Jan 2026 12:00:14 +0000https://chengyu.eu/posts/ai-blog-redesign-model-comparison/Three models, one vague brief to redesign a blog’s UX, visuals, and SEO — and a clear split between models that plan first and models that just ship a result.https://chengyu.eu/posts/nvidia-ces-2026-token-economics/Wed, 07 Jan 2026 12:00:00 +0000https://chengyu.eu/posts/nvidia-ces-2026-token-economics/Jensen Huang’s keynote wasn’t really about FLOPS — it was about token cost, and NVIDIA quietly shifting from selling chips to selling the entire stack.https://chengyu.eu/posts/veo-3-1-vs-jimeng-video-test/Thu, 18 Dec 2025 18:02:01 +0000https://chengyu.eu/posts/veo-3-1-vs-jimeng-video-test/The same prompt, two video models — one produced a slightly rough transition clip, the other gave a character a 180-degree head spin.https://chengyu.eu/posts/gpt-5-first-day-impressions/Fri, 08 Aug 2025 17:00:09 +0000https://chengyu.eu/posts/gpt-5-first-day-impressions/From debugging code to language practice to research, ChatGPT has quietly become a daily-use tool rather than a novelty.https://chengyu.eu/posts/agents-vs-workflows/Fri, 13 Jun 2025 16:00:16 +0000https://chengyu.eu/posts/agents-vs-workflows/A drag-and-drop OCR-plus-LLM tool got me thinking about where a workflow ends and an agent begins — with a simple three-question test.https://chengyu.eu/posts/qwen-agent-smart-customer-service/Wed, 11 Jun 2025 11:00:12 +0000https://chengyu.eu/posts/qwen-agent-smart-customer-service/A telecom customer-service example — using Tools for live data lookups and a knowledge base for static rules, with an LLM deciding which to call.https://chengyu.eu/posts/testing-google-video-model/Mon, 09 Jun 2025 18:02:39 +0000https://chengyu.eu/posts/testing-google-video-model/Eight-second clips, decent visual richness, and Chinese text prompts that either get dropped or come out garbled.https://chengyu.eu/posts/cursor-three-months-pro-free/Mon, 09 Jun 2025 14:00:52 +0000https://chengyu.eu/posts/cursor-three-months-pro-free/A quick look at why the AI coding editor everyone’s been talking about is worth the free trial.https://chengyu.eu/posts/subscribing-google-ai-pro/Mon, 09 Jun 2025 14:00:36 +0000https://chengyu.eu/posts/subscribing-google-ai-pro/Google I/O 2025 rebuilt every product around AI — enough to pull a committed OpenAI user over to a Pro subscription.https://chengyu.eu/posts/openai-clay-style-3d-art/Mon, 09 Jun 2025 14:00:24 +0000https://chengyu.eu/posts/openai-clay-style-3d-art/Comparing Gemini’s and OpenAI’s image generation on the same blurry wall painting, then pushing the same subject into a 3D clay-render style.https://chengyu.eu/posts/perplexity-pro-free-year-code/Mon, 09 Jun 2025 14:00:10 +0000https://chengyu.eu/posts/perplexity-pro-free-year-code/Perplexity has joined the AI subsidy wars — a full year of Pro for free with a discount code, plus the usual advice to turn off auto-renewal.https://chengyu.eu/posts/diy-voice-cloning-project/Sat, 10 May 2025 10:00:18 +0000https://chengyu.eu/posts/diy-voice-cloning-project/An open-source voice cloning project on a consumer GPU, then wired into my publishing pipeline so blog posts turn into short videos automatically.https://chengyu.eu/posts/qwen3-4b-unlimited-local-translation/Fri, 09 May 2025 18:00:21 +0000https://chengyu.eu/posts/qwen3-4b-unlimited-local-translation/Pairing a small local Qwen3 4B model with Immersive Translate for a private, uncapped translation setup — plus a prompt trick that speeds it up.https://chengyu.eu/posts/chatgpt-4o-photo-editing-fun/Wed, 26 Mar 2025 23:01:20 +0000https://chengyu.eu/posts/chatgpt-4o-photo-editing-fun/Turning family photos into cartoons, adding a warship and a dinosaur to travel pictures, and restyling a portrait — a weekend with ChatGPT-4o’s new image tools.https://chengyu.eu/posts/github-copilot-subscription/Fri, 15 Nov 2024 11:03:26 +0000https://chengyu.eu/posts/github-copilot-subscription/Ten dollars a month for autocomplete that actually understands what I’m writing.https://chengyu.eu/posts/chatgpt-o1-preview-o1-mini-first-look/Sat, 14 Sep 2024 07:17:01 +0000https://chengyu.eu/posts/chatgpt-o1-preview-o1-mini-first-look/A quiet rollout, a tight usage cap, and a version that’s clearly a big step up on reasoning.https://chengyu.eu/posts/whisper-fastapi-remote-service/Sun, 25 Aug 2024 19:00:02 +0000https://chengyu.eu/posts/whisper-fastapi-remote-service/Offloading transcription from a GPU-less NAS to a desktop with an actual GPU, exposed as a simple internal API.https://chengyu.eu/posts/whisper-on-nas-storage-fixes/Sat, 10 Aug 2024 21:00:17 +0000https://chengyu.eu/posts/whisper-on-nas-storage-fixes/A simple script to auto-transcribe Obsidian voice notes, derailed twice by the NAS running out of space in root and /tmp.https://chengyu.eu/posts/obsidian-smart-connections-plugin/Thu, 08 Aug 2024 11:00:18 +0000https://chengyu.eu/posts/obsidian-smart-connections-plugin/Chatting with my own notes via a local model, plus a gotcha when pointing the plugin at a self-hosted One API server.https://chengyu.eu/posts/comparing-llms-for-customer-service-summaries/Sat, 03 Aug 2024 12:00:04 +0000https://chengyu.eu/posts/comparing-llms-for-customer-service-summaries/Running ten locally-hosted models on consumer GPU hardware to see which ones best summarize customer-service calls after speech-to-text.https://chengyu.eu/posts/llm-agents-basics/Tue, 30 Jul 2024 00:00:09 +0000https://chengyu.eu/posts/llm-agents-basics/What an ‘agent’ actually means in the context of large language models, and a worked example in a customer-service setting.https://chengyu.eu/posts/free-gpu-finetuning-colab/Fri, 10 May 2024 17:00:18 +0000https://chengyu.eu/posts/free-gpu-finetuning-colab/After AMD’s ROCm ecosystem let me down for local fine-tuning, free Colab GPUs turned out to be the pragmatic way to fine-tune Llama 3 for free.https://chengyu.eu/posts/maxkb-personal-knowledge-base/Fri, 10 May 2024 09:00:23 +0000https://chengyu.eu/posts/maxkb-personal-knowledge-base/Self-hosting the open-source MaxKB knowledge-base tool at home, and pointing it at a Qwen-backed model for Chinese-language documents.https://chengyu.eu/posts/local-llama3-open-webui/Fri, 10 May 2024 09:00:11 +0000https://chengyu.eu/posts/local-llama3-open-webui/Pairing a consumer GPU running local Llama 3 with Open WebUI and One API to combine local and remote models behind one interface.https://chengyu.eu/posts/whisper-obsidian-voice-notes/Wed, 01 May 2024 13:00:11 +0000https://chengyu.eu/posts/whisper-obsidian-voice-notes/Wiring up Obsidian’s Whisper plugin to the free Cloudflare-backed OpenAI API from the previous post.https://chengyu.eu/posts/free-qwen-openai-compatible-api/Wed, 01 May 2024 06:48:10 +0000https://chengyu.eu/posts/free-qwen-openai-compatible-api/Wrapping Cloudflare’s free Workers AI models (including Qwen) behind an OpenAI-compatible endpoint, so existing front-ends don’t need to change a line of code.https://chengyu.eu/posts/local-llm-deployment-experience/Tue, 23 Apr 2024 15:30:00 +0000https://chengyu.eu/posts/local-llm-deployment-experience/Some hands-on experiments with un-tuned open models, and a few thoughts on where the real value in LLM products actually sits.https://chengyu.eu/posts/elmo-ai-browser-extension/Sun, 14 Apr 2024 18:09:00 +0000https://chengyu.eu/posts/elmo-ai-browser-extension/Elmo can summarize any page on the fly — and it turns out that’s genuinely useful for reading in a second language.