![[images/The Ridiculous Engineering Of YouTube Videos.webp]] **Creator:** Enrico Tartarotti · **Published:** 2026-05-07 · **Length:** 14:57 · [Watch on YouTube](https://www.youtube.com/watch?v=giP1iIbSVh0) > *The most-looked-at interface in the world isn't Netflix, TikTok, or Instagram, it's YouTube. And almost nothing about it is an accident.* (from the video description) Part of Enrico's "Ridiculous Engineering" series on the design, engineering and psychology behind everyday tech. Sponsored by **Gamma**. The video ends by pointing to the next episode, [The Ridiculous Engineering of Loading Screens](https://www.youtube.com/watch?v=XqhduyTnEM8). *Summary written from the auto-generated captions and sampled frames, so a few spoken figures may be slightly garbled; those are flagged below.* ## 1) Detailed outline ### [0:00] Cold open: the most-watched interface in the world - People spend more time watching YouTube than Netflix, Disney+, TikTok, Instagram, or even TV, and Enrico says that's not by chance. - **Hover previews:** the old hover preview showed a title, thumbnail and a few seconds of video. Since **2021** the preview is a full mini player with a **seek bar**, **captions on by default** (which you can switch off) and the option to **turn the audio on**, so you can watch a whole video without committing to it. - Choosing what to watch is one of the highest-friction points in the experience. A few seconds of watching in the preview **counts as a view** and lands in your **watch history**. - Every button and "seemingly inconsequential decision" is engineered to keep you up at 3 a.m. watching a documentary about shipping containers. - Thesis: three layers make YouTube videos work, and only the first one is visible. ### [1:23] Layer one: the interface and friction engineering - **You can't really close a video on mobile.** Tapping the X first drops it into **picture-in-picture** while you browse; only a second tap closes it. Mobile navigation is **gesture-based**: swipe down for picture-in-picture, swipe up for full screen, and neither closes the video. There's no close button other than leaving the app. - **TV redesign:** pressing any key brings up controls with **recommendations** peeking underneath. Even while you scroll recommendations that take over most of the screen, the main video keeps playing. - **The principle is minimizing non-playing time.** YouTube isn't optimizing raw time in the app. Like spending 20 minutes browsing Netflix, time spent choosing hurts your perception of the product ("there's nothing to watch here"). YouTube optimizes **playing time**, by growing watching and by shrinking browsing. That's why thumbnails play and the video stays visible while you browse. - **Two forces in tension:** "OG YouTube," a simple player for creative videos built when the platform had about 40 videos, versus changed user behavior. Short-form platforms (**TikTok**, **Instagram**, **YouTube Shorts**) have no non-playing time because videos are simply fed to you. YouTube has to keep what makes it loved while designing for today's habits. - **A/B testing at scale:** a 2024 clip posted by a user on X showed a test build where scrolling up and down in full screen doom-scrolled through long-form videos. It never shipped, but YouTube is likely running dozens of such tests at any time. - **PewDiePie's de-algorithmed YouTube:** he recently used browser extensions to strip Shorts, the algorithmic home page and fancy features to get back to old YouTube. Part of Enrico wants that too (he shows his old account creation date), but as someone who builds tech products, he cites the "golden rule of product management": don't listen to what users say, observe what they do. These features exist because they work. - **Autoplay countdown:** introduced in **2014** at 10 seconds, now **5 seconds**. People write custom scripts to change it, but it works. - **The quality picker** switched from explicit resolutions to generic "higher quality" and "lower quality" options. Enrico finds this infuriating but understands it. - The constant experimentation balances getting people to watch against turning YouTube into a "doom-scrolling slopfest." ### [5:12] Layer two: content engineering - Because YouTube is the most-watched medium, the **content itself** is hyper-engineered too. Enrico reveals that his own opening (the hook, the reveals, making you look at the picture-in-picture button before explaining it) was all planned. Many "organic" creator videos have hours of work, sometimes whole teams, behind them. - YouTube's superpower is in-depth content **you choose** to watch: tutorials for anything, a 12-hour Star Wars video essay. But choice creates the click problem. - **The thumbnail effect:** one of Enrico's videos got **3,776 views in 55 days**, essentially dead, then reached **3.5 million views** after he changed only the thumbnail. He tried **20 different thumbnails** for that one video, and shows similar sets for others. - **Thumbnail rules he uses:** - Since most people use **dark mode**, lighter backgrounds stand out more. - **Red** attracts attention, especially in contrast with other colors. - Realism doesn't matter; the goal is to be **recognizable in 0.1 seconds**, which is why the UI in this video's thumbnail is exaggerated. - Doodles and drawings don't mean anything literal; they signal that something hidden is waiting to be discovered. - **MrBeast** tested across his videos and found a **closed mouth** performed slightly better than an open one, so he stopped doing the open-mouth face. - Creators make 10 to 30 variations, changing shirt colors and framing, to make thumbnails more clickable. - **The "OG" counterfactual:** Enrico shows a plain, descriptive version of this video's thumbnail and title next to the real one and asks which you'd actually have clicked. As the platform optimizes for attention, the videos become even more engineered than the platform. ### [8:02] Inside YouTube Studio - **Retention curves** show, for every sentence, whether viewers drop or stay. **YouTube Studio** shows a huge range of metrics, and Enrico adds extensions for even more, "more powerful than many other data analytics softwares." - YouTube knows this is overwhelming and is managing it. The Studio home page shows a feed of recent comments across all videos. - **Negativity bias:** in a past video Enrico measured his channel's comments as **21.4% positive, 17.1% neutral, 61.5% negative**, like people being more likely to leave a one-star restaurant review after a bad meal than a five-star one after a good meal. - **Comment filtering:** for a few months, the Studio home feed has shown him only positive comments, while the full list still includes negative ones. He can't confirm this officially (YouTube hasn't announced it) and asks fellow creators to weigh in, but it looks like YouTube is shielding creators from negativity. - **Why:** if creators stop creating, or new creators can't succeed, the platform dies. - **Hype feature:** viewers can "hype" a video they like to boost it in the algorithm, and the boost is **inversely proportional to subscriber count**, so smaller channels get bigger boosts. ### [9:55] Sponsor: Gamma - Enrico needed to share channel metrics with the Gamma team. Instead of building slides by hand, he pasted raw YouTube Studio screenshots into **Gamma**, described what he wanted, picked a style, and got a full presentation in seconds. - **Gamma Imagine** extends this to infographics, posters and social media content without design skills. He cites 70 million Gamma users. ### [10:59] Layer three: a YouTube video isn't a video - **The request:** clicking a video makes your browser send a **GET request** to YouTube's servers with your user ID, browser info, cookies and the video ID. - **One upload, hundreds of files:** a 20 GB 4K upload can't be streamed as is, so YouTube makes versions at **4K, 1080p, 720p, 480p and 360p** (even the 4K version is lower quality than his original) and chops each into **2 to 10 second segments**. - **The DASH manifest** is "a restaurant's menu" listing links to every version and resolution. - **Adaptive streaming:** the browser checks screen resolution, current internet speed and device, then requests the first chunk (say, the first two seconds at 1080p). While that plays, the player fetches the next chunk and rechecks your connection, over and over. The **gray buffer bar** next to the playhead is chunks already downloaded. When quality drops during a heavy download, the player simply grabbed the next chunk at a lower resolution. - **Codecs by popularity:** new uploads start in **H.264** (cheapest, lowest quality). Past roughly **3,000 views** a video is re-encoded in **VP9** (mid-tier), and past **1 million views** in **AV1** (highest tier). "More popular videos scientifically look better." He jokes that you should hype this video so it reaches the AV1 tier. - This skips **CDNs**, codec details and more; the point is that a video arriving seamlessly in 0.1 seconds is anything but simple, at a scale of about **106,000 years of video watched every day**. ### [13:23] Why YouTube almost didn't happen - The founders once worried they only had 40 videos; a year later they had the opposite problem. Enrico says they were spending around **$1 billion a year** on infrastructure to deliver video while revenue was tiny (the captions render the revenue figure as "$30 a year," which is likely garbled), with each view costing roughly **a penny** to deliver. - In **2007**, YouTube alone used as much bandwidth as the entire internet did in **2000**. By **2014** it was responsible for about **20% of all bits** moved across the internet. - **Google bought YouTube in 2006 for $1.65 billion** (the captions momentarily flip this around). It wasn't just a business deal: in 2006 there was no cloud, so a video platform needed its own data centers, cables and servers, and Google was basically the only company with that infrastructure at scale. - Managing YouTube's scale is still one of the hardest problems in tech. Google even built a **custom chip** purpose-built for video encoding, compressing videos into segments so you can watch on a bus in a remote mountain in Peru without the spinning wheel. ### [14:39] Outro - Teases the next episode on **loading screens**, another everyday thing that hides more design, engineering and psychology than you'd think. ## 2) Things mentioned ### YouTube interface features | Item | Role in the video | |---|---| | **Playable hover previews (2021)** | Full mini player in the thumbnail, with seek bar, captions and audio; counts as a view. | | **Mobile picture-in-picture and gestures** | Swipe down for mini player, up for full screen; the X needs two taps to really close. | | **YouTube TV app redesign** | Recommendations open over a video that keeps playing. | | **Autoplay countdown** | Launched in 2014 at 10 seconds, now 5. | | **Quality picker** | Changed from explicit resolutions to "higher" and "lower" quality labels. | | **Doom-scroll long-form test (2024)** | A/B test clip shared on X; never shipped. | | **YouTube Shorts** | YouTube's short-form feed, part of the shift toward auto-fed video. | | **Hype** | Viewer boost for videos, larger for channels with fewer subscribers. | ### Creator tools and content techniques | Item | Role in the video | |---|---| | **YouTube Studio** | Creator dashboard with retention curves, metrics and a comment feed that seems to filter out negativity. | | **Retention curve** | Shows sentence by sentence where viewers drop. | | **Third-party analytics extensions** | Enrico uses them for extra Studio metrics (not named). | | **Thumbnail iteration** | 20 versions for one video; 10 to 30 variations are common among creators. | | **Thumbnail design rules** | Light backgrounds for dark mode, red for attention, recognizable in 0.1 seconds, exaggerated UI, meaningful-looking doodles. | | **MrBeast's mouth test** | Closed mouth slightly outperformed open mouth. | | **PewDiePie's extension setup** | Removing Shorts and the algorithmic home page to recreate old YouTube. | ### Video delivery technology | Item | Role in the video | |---|---| | **HTTP GET request** | What your browser sends when you click a video. | | **DASH manifest** | The "menu" of every version and segment of a video. | | **Adaptive bitrate streaming** | Picking each 2 to 10 second chunk's quality based on connection and device. | | **Buffer bar** | The gray line showing chunks already downloaded. | | **H.264, VP9, AV1** | Codecs YouTube moves a video through at roughly 3,000 and 1 million views. | | **CDN** | Content delivery network, mentioned as skipped detail. | | **Custom video-encoding chip** | Google's purpose-built encoding hardware (Google has publicly described this as its Argos video coding unit; the video doesn't name it). | | **Data centers, cables, servers** | The infrastructure only Google had at scale in 2006. | ### Products, companies and platforms | Item | Role in the video | |---|---| | **YouTube** | The subject of the episode. | | **Google** | Bought YouTube in 2006 for $1.65 billion. | | **Netflix** | Comparison for browsing time that hurts the product's perception. | | **Disney+, TikTok, Instagram, TV** | Platforms YouTube out-watches. | | **X** | Where a user posted the 2024 test-build clip. | | **Gamma and Gamma Imagine** (gamma.app) | Sponsor; AI presentation and visual design tool. | | **Flask** (flask.do) | Enrico's video feedback tool, plugged in the description. | ### Concepts - **Friction engineering** and minimizing non-playing time. - **"Don't listen to users; observe what they do"** as a product-management rule. - **Content engineering**: hooks, reveal order and thumbnails as designed systems. - **Negativity bias** in comments and reviews. ## 3) Biographies ### Enrico Tartarotti (host, writer) Italian-born tech creator and designer behind the YouTube channel **Enrico Tartarotti**, where he goes behind the scenes of everyday tech products through designed video essays on UX, product design and tech psychology. He builds **Flask** (flask.do), a video feedback and review tool, offers one-hour calls through enricotartarotti.com, and handles sponsorships through **Rakugo Media**. Credits for this video: written by Enrico Tartarotti, produced by Enrico Tartarotti and **Bartek Malinowski**, edited by **Seequence**. He's on X as @EnriTarta and on Instagram as @enritarta. ### People referenced - **PewDiePie (Felix Kjellberg)**: Swedish YouTuber who was long the platform's most-subscribed individual creator. Cited for a recent video where he used extensions to strip YouTube back to its older, less algorithmic form. - **MrBeast (Jimmy Donaldson)**: American YouTuber known for large-scale challenge and giveaway videos and the most-subscribed channel on YouTube. Cited for testing thumbnail expressions and finding a closed mouth performs slightly better. - **YouTube's founders (Chad Hurley, Steve Chen and Jawed Karim)**: former PayPal employees who launched YouTube in 2005; Karim uploaded its first video, "Me at the zoo." The video refers to them as the founders worried about having only 40 videos and then about crushing bandwidth costs, without naming them individually. - **Bartek Malinowski**: Enrico's producer. - **Seequence**: the editing studio credited for the video. ### Organizations referenced - **YouTube**: Google-owned video platform founded in 2005, the most-watched video service in the world. - **Google**: acquired YouTube in October 2006 for $1.65 billion in stock and supplied the data-center infrastructure needed to keep it running. - **Gamma**: maker of the Gamma AI tool for presentations, documents and, with Gamma Imagine, other visuals; sponsor of this episode.