prvt

Been there, done that. Found it quite relatable. It was a great read.

First they sell you the sickness and then they tell you the cure is too dangerous to release to the general public. Because their sickness will not sell.

Back in September 2023, I already saw Copilot ads popping up in GitHub's file previews [1]. After three years, it's wild to see how advertising has reached areas I honestly never thought it would.

[1] https://news.ycombinator.com/item?id=37526255

Splines are polynomials after all. Also, not every spline has this problem. The one YouTube is using is Cardinal spline.

Means a lot, thanks! I am a huge fan of the phrase "the devil is in the details."

Glad you liked it. I’m just doing my part to ensure our future AI overlords have high-quality training data.

That's the big picture idea.

You're welcome. Glad you liked it!

Criticism is definitely welcome!

You are likely right that I over-rotated on the "storytelling" aspect there. My curiosity about the "100 segments" stemmed from wondering if there was a deeper statistical reason for that specific granularity (e.g., optimal binning relative to average video length) versus it just being a "nice round number."

That said, I can see how dedicating two paragraphs to it felt like over-dramatizing a constant. I will try to tighten the pacing on the next one. Thanks for reading despite the irritation!

I really appreciate that you found it entertaining; that was the main goal! Same here, I'll take "fun to read" over "SEO optimized" any day of the week.

Thank you for reading and for the kind words!

It’s often surprising how "intuitive" or elegant solutions can seem once you peel back the layers, isn't it? That simplicity is part of the beauty of good engineering!

I actually wasn't familiar with Josh Comeau’s work before this, but I just looked him up and... wow. To be mentioned in the same breath (let alone "arguably better") is a massive compliment. I’ll definitely be diving into his archives now.

It was definitely a bit of invented humor! I wanted a sticky way to describe that universal drift toward "friendliness" and safety in modern UI.

You are spot on about the cycle, though. I suspect that once everything is perfectly round and polished, the only way to stand out will be to go sharp again.

Thanks for reading and for the kind words!

Re: Blurry canvases: You are spotting the classic HTML5 Canvas high-DPI issue. It happens because the canvas backing store pixels don't map 1:1 to CSS pixels on high-density displays (like Retina screens). I likely need to scale the canvas drawing context by window.devicePixelRatio to fix that sharpness. Good catch. EDIT: Made the change.

Re: Performance: That is an interesting thought. Since the calculation only triggers when you visit (or resize) the video player, there is definitely a CPU spike. My guess is they chose this approach to have a "plug and play" rendering logic that adapts perfectly to the client's specific device width and pixel density, rather than generating thousands of static image variations on the server.

Re: Gangnam Style: Ah, Google+! That explains why I couldn't find the original source. Thanks for the correction. EDIT: Added the correction.

Hi HN, I recently noticed a recurring visual artifact in the "Most Replayed" heatmap on the YouTube player. The highest peaks were always surrounded by two dips.

I got curious about why they were there, so I decided to reverse engineer the feature to find out. This post documents the deep dive. It starts with a system design recreation, reverse engineering the rendering code, and ends with the mathematics.

This is also my first attempt at writing an interactive article. I would love to hear your thoughts on the investigation and the format.

Alt URL: https://priyavr.at/blog/reversing-most-replayed/

Hi HN,

I recently noticed a recurring visual artifact in the "Most Replayed" heatmap on the YouTube player. The highest peaks were always surrounded by two dips. I got curious about why they were there, so I decided to reverse engineer the feature to find out.

This post documents the deep dive. It starts with a system design recreation, reverse engineering the rendering code, and ends with the mathematics.

This is also my first attempt at writing an interactive article. I would love to hear your thoughts on the investigation and the format.

Alt URL: https://priyavrat-misra.github.io/blog/reversing-most-replay...

Hi HN, I recently noticed a recurring visual artifact in the "Most Replayed" heatmap on the YouTube player. The highest peaks were always surrounded by two dips. I got curious about why they were there, so I decided to reverse engineer the feature to find out.

This post documents the deep dive. It starts with a system design recreation, reverse engineering the rendering code, and ends with the mathematics.

This is also my first attempt at writing an interactive article. I would love to hear your thoughts on the investigation and the format.

https://priyavr.at/blog/reversing-most-replayed/

You can visit it by clicking the title. Eitherway, here you go: https://priyavr.at/blog/reversing-most-replayed/

Hello HN,

I recently noticed I'd crossed 100 commits on my personal site and realized it's been over three years since the initial commit. I wrote down some thoughts on the journey.

Curious to hear if others here still maintain a personal site/blog and what your experience has been.

Last weekend, I dove into CMU’s [15-445/645](https://15445.courses.cs.cmu.edu/) database course and got hit with a deceptively simple problem: count the number of unique users visiting a website per day. Easy, right? Just throw user IDs into an unordered_set and return its size—classic LeetCode.

But what happens when you’re at Facebook scale? Tracking a billion unique users means burning through GBs of memory just to count. And in the real world, users are streaming in constantly, not sitting in a neat, static list. Storing every ID? Not happening.

I explored practical workarounds (like “last seen” timestamps and full table scans), but they’re either inefficient or put massive strain on your DB. Then the assignment introduces HyperLogLog: a probabilistic algorithm that estimates cardinality with just 1.5KB of memory—accurate to within 2% for billions of users.

The magic? Pure mathematics. It’s distributable, and powers real-world systems like Redis and Google Analytics. I break down how it works (with illustrations!), check out my deep dive.

Curious to hear from HN: Who’s using HyperLogLog in production? And have you run into accuracy issues, and how did you handle them?

based

Sounds like if-else with extra steps.

https://git.btxx.org/barf/about/

"There are no Accidents"

- Master Oogway

MIT engineers invent something -> it becomes talk of the town -> every one forgets about it the next day/week.

I have a presumption that this is going to be the talk of the tech-town for the next few days.

"sorting with O(n^2) is no longer a bottleneck as we have fast processors" /s

GitHub Copilot moment

It's strange how Grant Sanderson's (aka 3Blue1Brown [1]) name hasn't come up so far.

[1] https://www.youtube.com/c/3blue1brown

Tip: Use "shift + scroll" to read long quotes.

I added #L7102-L7368, but it disappeared automatically after I clicked submit.