Yash Singhal 🔥

1:03 AM. PM texts: “Quick change-can we just add dark mode before release?” I turned off notifications. Because my soul already is in dark mode. The feature is complete.

@kirat_tw @jarredsumner Jarred speedran ‘build startup → get acquired’ faster than Bun installs dependencies

Congrats @jarredsumner and the BunJS team

@kirat_tw Reading it on Twitter hits different than experiencing it at 3AM in production. 😭🔥

Move fast, break things. Momentum is everything. Failing is probably better than not trying. Unfortunately these things should be experienced in hindsight and not read on twitter.

@mehulmpt Bro didn’t just warn us, he basically posted the spoiler alert for the whole outage arc. 😂

Well this didn't take long

@arpit_bhayani Wild ! how an amoeba explains range partitioning better than half the system design textbooks 😄

Range-based partitioning is like an amoeba. You split when a partition gets hot, you merge when things cool down, and you move the partition across nodes (and transitively the data, if needed) when you want to grow beyond a single machine. The best part of range-based partitioning is that it sits in the middle between hash and static approaches. It avoids the randomness of hash ownership and the heavy metadata burden of static ownership. That's why, for stateful workloads, so many systems, operating at scale, prefer range-based ownership.

@archiexzzz Absolutely. JSON is a token tax - every { , : " } is money

JSON is a token-hungry parasitic dinosaur. If you look at openai/tiktoken, the pre-trained BPE vocabulary has most JSON structural characters - brackets, commas, quotes - as individual tokens in the vocabulary. This means even a small JSON can emit >1000 tokens, since every {, }, ", :, and , burns a separate token. Now imagine your use case: you're ingesting some massive JSON response from an MCP server tool, and you need to shove it all into your LLM for inference. Your cost just went through the roof. I've been there. We were constantly hitting into context windows longer than 1M tokens, just inferring over these giant JSON strings. We ended up with all these hacky % operations on the token length just to figure out how many LLM calls we needed to digest the whole thing. It's a nightmare. Your options are trash: > either you throw away data, which is a hard no because you don't want data loss, or > you accept that you're just going to burn more VC money because every provider charges you per token. TOON is a decent alternative, but it's not perfectly accurate. The thing is, LLMs were pre-trained on a lot of JSON. You switch everything to TOON, and your accuracy takes a hit. It's a trade-off. If you're ranking formats for accuracy, it's roughly: Markdown-KV > XML > YAML > HTML > JSON > TOON TOON is getting better, but as an AI engineer, the most optimized way to solve this right now is to figure out your personal trade-off between latency <> cost <> model accuracy. If your task is simple and doesn't live or die by accuracy, try TOON -> Run evals on a small dataset -> If the numbers look good, stick with it -> If they're bad, fall back to JSON or XML. But really, the smart move is to think before the LLM call: > Can you filter that JSON down to just the fields the LLM actually needs? > Can you summarize big text blobs with a cheaper model first? > Can you break the task into a conversation instead of one giant context dump? That's where you win. The format is just one part of the puzzle.

@LowLevelTweets Relax 😆 that’s just bots knocking on every door on the internet hoping someone left .env outside. Ours is locked.

guys how dumb do you think I am? come on. nice IP addresses btw.

The takeaway: For global businesses, depending on a robust CDN isn't optional—it's survival. 🥷 A single point of failure at this scale creates a cascading crisis for whole ecosystem. 🌋 Are we too centralized? What's solution? #Tech #Cloudflare #CDN #Cybersecurity #DevOps

2/ The 3 Critical Functions, CDN provides: 🚀 Speed: Caching + instant TLS = ⚡global delivery. 🛡️ Security: Uses Anycast to diffuse massive DDoS attacks across its entire network capacity. 🟢 Resilience: High fault tolerance. Origin servers struggle-> CDN serves cached copy.

/1 𝘊𝘭𝘰𝘶𝘥𝘧𝘭𝘢𝘳𝘦 is world’s largest 𝐂𝐃𝐍 . What it does: - 𝐂𝐥𝐨𝐬𝐞𝐫 𝐜𝐨𝐧𝐭𝐞𝐧𝐭: Uses global PoPs & edge servers to cut latency. - Faster connections: Terminates TLS handshakes instantly at edge. - 𝐂𝐚𝐜𝐡𝐢𝐧𝐠: Protecting your origin server from massive load.

When 𝐗, 𝐎𝐩𝐞𝐧𝐀𝐈, 𝐒𝐩𝐨𝐭𝐢𝐟𝐲, & 𝐂𝐚𝐧𝐯𝐚 all vanish at once, it's not a simple server crash—it's an INFRASTRUCTURE FAILURE. 🚨 The recent 𝐂𝐥𝐨𝐮𝐝𝐟𝐥𝐚𝐫𝐞 outage was a $100𝘔 𝘳𝘦𝘮𝘪𝘯𝘥𝘦𝘳 of how 1 company powers internet. Why? A deep dive into the tech 👇

6. That’s the first half of the story. 𝘚𝘵𝘢𝘵𝘪𝘤 → 𝘔𝘝𝘊 → 𝘚𝘗𝘈 → 𝘉𝘍𝘍 Next up in 𝐏𝐚𝐫𝐭 𝟐: How 𝐒𝐒𝐑, 𝐒𝐒𝐆 & 𝐑𝐞𝐚𝐜𝐭 𝐒𝐞𝐫𝐯𝐞𝐫 𝐂𝐨𝐦𝐩𝐨𝐧𝐞𝐧𝐭𝐬 changed everything ⚡️ #frontend #webdev #nextjs #react #softwarearchitecture

5. SPAs solved interactivity but not complexity. So next: the 𝐁𝐚𝐜𝐤𝐞𝐧𝐝 𝐟𝐨𝐫 𝐅𝐫𝐨𝐧𝐭𝐞𝐧𝐝 (𝐁𝐅𝐅) pattern. Each frontend (web, mobile) gets its own API. ✅ 𝘊𝘭𝘦𝘢𝘯𝘦𝘳 𝘥𝘢𝘵𝘢 ✅ 𝘍𝘢𝘴𝘵𝘦𝘳 𝘪𝘵𝘦𝘳𝘢𝘵𝘪𝘰𝘯 ❌ 𝘔𝘰𝘳𝘦 𝘪𝘯𝘧𝘳𝘢 𝘵𝘰 𝘮𝘢𝘯𝘢𝘨𝘦

𝐓𝐡𝐞 𝐄𝐯𝐨𝐥𝐮𝐭𝐢𝐨𝐧 𝐨𝐟 𝐅𝐫𝐨𝐧𝐭-𝐄𝐧𝐝 𝐀𝐫𝐜𝐡𝐢𝐭𝐞𝐜𝐭𝐮𝐫𝐞 (𝐏𝐚𝐫𝐭-𝟏)🚀 wanna sound 𝘴𝘦𝘯𝘪𝘰𝘳 in your next 𝘪𝘯𝘵𝘦𝘳𝘷𝘪𝘦𝘸 ... Let's go...👇🧵

Big news 🇮🇳 - OpenAI is giving 1-year free access to ChatGPT Go starting Nov 4! More power → GPT-5, uploads, memory, faster protos. Big W for devs 🚀 #AI #DevTools #ChatGPT

@pbteja1998 @SiteGPT Haha the SiteGPT chat flow

@yash_25log @SiteGPT What looks super smooth?

Made my website interactive using @SiteGPT