--- title: "Tech FAQ" description: "Questions and answers from hands-on engineering posts: computer vision, applied machine learning, agentic deployment, and the site's own static-plus-Workers infrastructure." updated: 2026-10-02 author: "Philipp D. Dubach" canonical_url: "https://philippdubach.com/faq/tech/" source_url: "https://philippdubach.com/faq/tech/index.md" --- # Tech FAQ *Philipp D. Dubach ยท Updated October 2, 2026* --- From posts in the Tech category, newest first. Each answer reads as a citable claim and links back to the source post for the code, the deploy notes, or the failure mode that motivated the project. The angle: code that runs, not opinion pieces about the industry. Project work covers computer vision (a card counter on OpenCV, an audit of F3ED, the NeurIPS 2024 tennis-shot detector), applied ML (visualizing PyTorch gradients, modeling postprandial glucose response with XGBoost), health-data engineering (a Python library for the Dexcom CGM), and agentic deployment (Cursor on a fresh Azure environment). The site itself is a recurring case study: Hugo plus GitHub Pages plus five Cloudflare Workers run social automation on Llama 4 Scout, content negotiation for AI crawlers, security headers, URL shortening on KV plus D1, and analytics aggregation. Answers lead with the version number, the benchmark, or the failure mode. Most questions come from engineers who want to know what actually shipped, so the answers say what shipped. ## Frequently Asked Questions ### How can I turn my reading list into a personal AI podcast? Research agents read a chosen set of links, group the themes, and look for related and unexpected sources. They develop one question into a script, which ElevenLabs narrates in short sections. The audio is assembled and checked against the script. From [How I turn my linkblog into a personalized podcast](https://philippdubach.com/posts/saved-links-to-personal-podcast/). ### How do I save links with an iOS Shortcut? An iOS Shortcut sends a URL and short description to a Cloudflare Worker. D1 stores the entry, and a public page lists the saved links by date. From [How I turn my linkblog into a personalized podcast](https://philippdubach.com/posts/saved-links-to-personal-podcast/). ### Which ElevenLabs voice narrates this podcast? The podcast uses the Christopher voice with ElevenLabs v4. The episode discloses synthetic narration, and the production workflow includes pronunciation and listening checks. From [How I turn my linkblog into a personalized podcast](https://philippdubach.com/posts/saved-links-to-personal-podcast/). ### Why rebuild a YouTube feed instead of using the app? I used the app only to see new videos from 10 to 15 channels. I did not use comments, recommendations, or Shorts. My replacement keeps only the subscription feed. From [Degoogling cost me my YouTube feed, so I made my own](https://philippdubach.com/posts/degoogling-cost-me-my-youtube-feed-so-i-made-my-own/). ### How do I use a YouTube RSS feed without an API key? Read each channel's public feeds/videos.xml endpoint. It publishes recent uploads in Atom, a standard XML feed format. The Worker reads that feed directly, so it needs no Google project, quota, or key. From [Degoogling cost me my YouTube feed, so I made my own](https://philippdubach.com/posts/degoogling-cost-me-my-youtube-feed-so-i-made-my-own/). ### Can I follow YouTube subscriptions without a Google account? Yes. Public YouTube channel feeds need no sign-in. You keep your own list of channels, and Google stores no subscription list for this feed. From [Degoogling cost me my YouTube feed, so I made my own](https://philippdubach.com/posts/degoogling-cost-me-my-youtube-feed-so-i-made-my-own/). ### How does it find a channel's videos without the API? You add a URL or @handle. The Worker reads the channel page once and finds its stable channel ID. It saves that ID and uses it for later feed requests. From [Degoogling cost me my YouTube feed, so I made my own](https://philippdubach.com/posts/degoogling-cost-me-my-youtube-feed-so-i-made-my-own/). ### How are YouTube Shorts filtered out? The Worker first checks the title for #shorts. It then tests how the video URL redirects, because Shorts and full videos resolve differently. A video appears only if it passes both checks. From [Degoogling cost me my YouTube feed, so I made my own](https://philippdubach.com/posts/degoogling-cost-me-my-youtube-feed-so-i-made-my-own/). ### Does watching a video track you? Not before you press play. The page shows static thumbnails. A click loads YouTube's privacy-enhanced player, which can then receive your IP address and store a local identifier. On iOS, the video starts muted because Apple blocks one-tap sound. From [Degoogling cost me my YouTube feed, so I made my own](https://philippdubach.com/posts/degoogling-cost-me-my-youtube-feed-so-i-made-my-own/). ### How are channels stored and the feed kept fast? Cloudflare KV stores the channel list. The Cache API keeps the merged feed near users for 15 minutes. Feed requests run in parallel and have timeouts, so one channel cannot stall the page. From [Degoogling cost me my YouTube feed, so I made my own](https://philippdubach.com/posts/degoogling-cost-me-my-youtube-feed-so-i-made-my-own/). ### What is Software 3.0 according to Karpathy? Karpathy frames three eras of software. Software 1.0 is humans writing explicit code. Software 2.0 is humans curating datasets and training neural networks, where the weights are the program. Software 3.0 is humans writing prompts, where the LLM is the interpreter and the context window is the program. The unit of programming shifts from a function to a paragraph. From [Karpathy's Software 3.0 Playbook](https://philippdubach.com/posts/karpathys-software-3.0-playbook/). ### When did agentic coding actually start working? Karpathy points to December 2024 as the inflection point. Before then, agentic tools were 'kind of helpful' but required constant correction. Over the December break, the latest models crossed a line where Karpathy stopped correcting them and started trusting the system. He flagged this on the record, warning that anyone whose mental model of AI was set by ChatGPT was already a generation stale. From [Karpathy's Software 3.0 Playbook](https://philippdubach.com/posts/karpathys-software-3.0-playbook/). ### What is the difference between vibe coding and agentic engineering? Vibe coding raises the floor: it lets non-engineers build software they could not build before. Agentic engineering raises the ceiling: it lets professional engineers preserve the existing quality bar while moving much faster. Karpathy thinks the productivity gap for the best users now exceeds the old 10x engineer benchmark by a wide margin. From [Karpathy's Software 3.0 Playbook](https://philippdubach.com/posts/karpathys-software-3.0-playbook/). ### Why are LLMs good at code and math but bad at common-sense tasks? Frontier labs train models with reinforcement learning, which requires verifiable rewards. Verifiable domains attract environments and signal, so they get the steepest gains. Everything outside the verifiable distribution stays jagged. Karpathy's takeaway for founders is that building a verifiable environment in your domain is real leverage. For workers, the more useful question than 'is my job safe' is 'is my job verifiable.' From [Karpathy's Software 3.0 Playbook](https://philippdubach.com/posts/karpathys-software-3.0-playbook/). ### What does Karpathy mean by 'outsource your thinking, not your understanding'? As agents do more execution, the bottleneck moves into the human's head. You still have to know what is worth building, why, and how to direct the work. Your value sits upstream of execution. Karpathy keeps building knowledge bases out of his own reading because the constraint of the next decade is less about compute than about how fast humans can deepen comprehension to keep directing systems that out-execute them. From [Karpathy's Software 3.0 Playbook](https://philippdubach.com/posts/karpathys-software-3.0-playbook/). ### Why does Karpathy say MenuGen 'shouldn't exist'? Karpathy built MenuGen as a full-stack app: photo a restaurant menu, OCR it, generate dish images, render a new menu. Then he saw the Software 3.0 version: hand the photo to Gemini, say 'use NanoBanana to overlay the dishes,' and a single model call returns the rendered menu. The lesson is that a lot of what gets built today is scaffolding around a capability the model could perform end-to-end. Before writing the next CRUD app, ask whether the model is the app. From [Karpathy's Software 3.0 Playbook](https://philippdubach.com/posts/karpathys-software-3.0-playbook/). ### How should hiring change in the agentic era? Karpathy argues whiteboard puzzles measure the wrong thing. Hiring should look like giving someone a really big project, having them implement it, and then trying to break it. His example: build a Twitter clone for agents, make it secure, simulate activity, then have ten Codex 5.4-X-high instances try to break the website. If your interview loop has not changed since 2022, you are selecting for the previous era. From [Karpathy's Software 3.0 Playbook](https://philippdubach.com/posts/karpathys-software-3.0-playbook/). ### Why have smartphone upgrade cycles slowed down? The average global smartphone replacement cycle has stretched to 3.5 years. Cameras, screens, and processors have reached a quality plateau where year-over-year improvements are small. Battery life has overtaken price as the top purchase driver for the first time, suggesting hardware differentiation has stalled. From [On-Device AI Models Will Be The New Reason to Upgrade Your Phone](https://philippdubach.com/posts/on-device-ai-models-will-be-the-new-reason-to-upgrade-your-phone/). ### How does Apple use Google Gemini for on-device AI? Google gave Apple complete access to the Gemini model in Apple's own data centers. Apple uses a process called distillation, where smaller models learn from Gemini's reasoning outputs to produce efficient models with Gemini-like performance at a fraction of the compute. These distilled models can run on-device without an internet connection. From [On-Device AI Models Will Be The New Reason to Upgrade Your Phone](https://philippdubach.com/posts/on-device-ai-models-will-be-the-new-reason-to-upgrade-your-phone/). ### What is the Apple Foundation Model? Apple's on-device Foundation Model is a roughly 3 billion parameter language model optimized for Apple Silicon through innovations like KV-cache sharing and 2-bit quantization. It runs at 30 tokens per second on iPhone 15 Pro and powers Apple Intelligence features including summarization, writing tools, and Siri enhancements. From [On-Device AI Models Will Be The New Reason to Upgrade Your Phone](https://philippdubach.com/posts/on-device-ai-models-will-be-the-new-reason-to-upgrade-your-phone/). --- Canonical: https://philippdubach.com/faq/tech/ This file is the canonical machine-readable variant of https://philippdubach.com/faq/tech/. Author: Philipp D. Dubach (https://philippdubach.com/).