@Kristof_Poland Although I have great respect for macroeconomics and agree 100% with what you write, unfortunately, elections are won by securing the votes of those who follow the rules of "nanoeconomics".
@catalinmpit Don’t look around so much. Follow your dreams and read omens/gut feelings. There are many car makers doing more less the same thing. And so what.
💪🏻
I'm wondering if designing the solution using AI models would cause the same issue. And even further, if AI would have been able to optimize it to the level achieved by the @recallai Team.
#vibecoding#vibedesign
A WebSocket connection to localhost cost one team $1M a year on AWS.
https://t.co/TnLqsytlfe runs meeting bots that capture millions of meetings per month. Every bot needed 4 CPU cores, so the team set a goal to cut that in half and cut the compute bill with it.
When they profiled the bots, they expected video encoding to dominate. Instead, most of the CPU time went to two functions from the C standard library, 𝗺𝗲𝗺𝗺𝗼𝘃𝗲 𝗮𝗻𝗱 𝗺𝗲𝗺𝗰𝗽𝘆. Their only job is to copy memory around. The bots were spending their cores on moving bytes, not on encoding video.
The reason was hiding in an unexpected place. To move raw video out of Chromium and into their encoder, they used 𝗪𝗲𝗯𝗦𝗼𝗰𝗸𝗲𝘁𝘀 𝗼𝘃𝗲𝗿 𝗹𝗼𝗼𝗽𝗯𝗮𝗰𝗸. Raw 1080p video is about 𝟵𝟯 𝗠𝗕/𝘀 𝗽𝗲𝗿 𝘀𝘁𝗿𝗲𝗮𝗺, and their busiest bots received 𝟭𝟱𝟬 𝗠𝗕/𝘀. At that volume, two properties of the WebSocket spec become expensive:
1. 𝗙𝗿𝗮𝗴𝗺𝗲𝗻𝘁𝗮𝘁𝗶𝗼𝗻. Chromium splits messages larger than 131KB. A single 3MB video frame turns into 24 separate WebSocket frames, and each one carries its own overhead.
2. 𝗠𝗮𝘀𝗸𝗶𝗻𝗴. The spec requires XOR-ing every byte sent from client to server. That is not a problem for normal web traffic, but it is a full extra pass over 150 MB/s of video.
So they searched for a cheaper transport and considered three options. Raw TCP/IP still fragments packets and copies data into kernel space. Unix domain sockets skip the network stack. But they still copy data between user space and kernel space. Shared memory copies nothing at all, and that was the way to go.
They built a custom lock-free ring buffer in Rust. Chromium writes video frames into 𝘀𝗵𝗮𝗿𝗲𝗱 𝗺𝗲𝗺𝗼𝗿𝘆. The media pipeline reads them directly, with no copies at all.
The result was up to 𝟱𝟬% 𝗹𝗲𝘀𝘀 𝗖𝗣𝗨 per bot and over $𝟭𝗠 𝗮 𝘆𝗲𝗮𝗿 𝗼𝗳𝗳 𝘁𝗵𝗲 𝗔𝗪𝗦 𝗯𝗶𝗹𝗹.
I've signed enough cloud bills to know that the expensive part is almost never where you expect it. At scale, defaults have price tags, so profile before you trust your assumptions.
Full write-up in the comments.
Image: https://t.co/TnLqsytlfe
This is a very interesting topic - such situations always remind me the Google Maps significant pricing change that made many companies and apps which relied on Google Maps services collapse.
As it turned out, it provided a vast majority of developers a proper education on how to use and not abuse cloud services.
Drinking tap water from steel or ceramic mugs/bootles/cups is the best way to avoid stuffing yourself with particles of plastic!
#water#drinkwater#hydration#plastic
@MarjoNiemenmaa@YourAnonOne Thank you, Marjo! I think this is a very important recommendation that should be widely spread. People think drinking mineral water is good for their health and it turns out they are stuffing themselves with particles of plastic!
@RyanEls4 it does not matter - my idea is better, because it is mine, and I can grow and adjust it.
Empowered solopreneurs and small team of passionate specialists can be unbelievably flexible!
@milan_milanovic Due to the flexibility of small teams, it is easier to enhance them by a wise use of AI tools.
Superpowered small teams of passionates are UNSTOPPABLE!
Tak jest. Gdyby mówił w dobrą stronę to byłby dziennikarzem. A tak to jest zwykłym pismakiem.
Gdzieś to już słyszałem.
Nawet księża ewangelizują na YouTube, a nie na ambonie, a ich sposób ewangelizacji jest dostosowany do aktualnych czasów (aggiornamento) - ale czy to powoduje, że przestają być księżmi?
Proponuję obejrzeć trylogię "U Pana Boga ..." - tam jest adekwatna kwestia: "tera są inne czasy, tera inne czasy są"
I think appropriate education and reaching high levels need tutoring.
Further self-education after getting a college degree is the way to go. But first of all you must know how to set proper goals.
Please change my mind