Sathya (@sathya_ui_dev) header
Sathya (@sathya_ui_dev) profile picture

AI Architect building with GenAI & React ⚙️ Decoding tech, AI & Indian markets in plain English 📈 Chennai | Not investment advice

Chennai
Joined May 2021
Photo by Sathya (@sathya_ui_dev)Photo by Sathya (@sathya_ui_dev)Photo by Sathya (@sathya_ui_dev)Photo by Sathya (@sathya_ui_dev)Photo by Sathya (@sathya_ui_dev)Photo by Sathya (@sathya_ui_dev)Photo by Sathya (@sathya_ui_dev)Photo by Sathya (@sathya_ui_dev)Photo by Sathya (@sathya_ui_dev)Photo by Sathya (@sathya_ui_dev)
I'm an AI Architect who builds GenAI products for a living. Every day I break down: 📈 What's moving Indian markets & why 🤖 AI/tech news, from someone who actually builds with it 🛠️ Lessons from my own builds No tips, no hype. Just clear analysis. Follow if that's your thing 👇
4
2
38
1,846
A simple blinking cursor may be draining your battery. I tasked opus to debug an issue that the cpu was above my target of <0.1% cpu usage when an idle window of opencraft was open. And it showed that 1.5% of a cpu core was engaged just in blinking that cursor causing a window redraw.
6
2
1
25
1,208
Sathya retweeted
In 2016, a developer unpublished an 11-line JavaScript package called left-pad from npm after a trademark-related dispute over another package. within minutes, builds depending on babel, react-related tools, and thousands of other projects started failing because npm could no longer find the package. the disruption lasted 2.5 hours before npm restored the missing version. all because thousands of developers had come to depend on a tiny while loop that adds characters to the left of a string.
5
3
58
5,397
Sathya retweeted
This Anthropic tool use paper is f*cking gold Anthropic's own team took Opus from 49% to 74% accuracy and cut token use by 85%. And the model wasn't the upgrade. The upgrade: they stopped feeding Claude every tool and every raw result. Tool search → defer loading → code calls → parallel → examples The 6 parts of advanced tool use: Part 1 → Tool search 58 tools ate ~55,000 tokens before the agent read a word. Search loads ~500 tokens and finds the 3-5 it needs 77,000 → 8,700 tokens, and on their MCP evals Opus 4 went from 49% to 74% Part 2 → Defer loading Every tool marked defer_loading, only your 3-5 most used stay loaded Search "github" and only the GitHub tools show up Part 3 → Call tools from code One script reads 2,000+ expense lines in a sandbox The model only sees the 3 people over budget. 200 KB → 1 KB Part 4 → Run in parallel 20 tool calls in one code block, 19+ inference passes skipped 43,588 → 27,297 tokens Part 5 → Show examples The schema says "string". The example says "2024-11-06" Accuracy on complex parameters: 72% → 90% Part 6 → Fix the bottleneck Bloat → search. Big results → code. Bad params → examples Opus 4.5 went from 79.5% to 88.1% More tools in context made the agent dumber. Fewer tokens made it smarter. I packed all 6 parts into one prompt. Drop it into Claude Code and it sets the whole tool stack up for you. 6 parts. 1 prompt. 49% → 74%. Then read the article below ↓
Your agent passed every test The issue is still open A better prompt won't fix it, because the model isn't the broken part I broke down the harness, loop and graph around it: what each one owns, how they nest, and which one to fix first ↓
30
4
2
99
4,309
13 years of React taught me this: The hardest part of a feature is never the code. It's the "small change" that comes after you say it's done.
1
1
28
AI writes the code in 30 seconds. I spend 30 minutes reading it. Still faster. Still not "done".
1
22
2016: copy-paste from Stack Overflow 2026: copy-paste from AI The bug is still in the same place.
5
4
126
IND vs NZ, 1st Test, Day 4 in Wellington India declare at 720/9, a 352-run lead Rishabh Pant 128 off 153 (6 sixes) Dhruv Jurel 78 off 79 NZ 119/2 at stumps, still 233 behind Bumrah removed Latham for a duck in the first over Conway 59*, Ravindra 27* Can India bowl them out on Day 5?
1
1
89
Sathya retweeted
Tf 😂😂😂
106
752
85
8,746
582,838
Friday box office 🎬 🆕 OM: ₹17.70 Cr on Day 1, Dhanush's biggest opening ever 🏹 Ramayana: ₹16.30 Cr on its 3rd Friday. ₹880 Cr India net, ₹1,575 Cr worldwide A Day 1 film just out-earned the year's juggernaut. Which one are you watching this weekend?
1
1
88
Junior dev vs senior dev when production goes down
2
1
9
856
KL Rahul in New Zealand: 124* off 265 balls. 🏏 India 344/2 at tea on Day 3, just 24 behind NZ's 368. Calm, patient, no fuss. Test cricket the old-school way. #NZvIND
1
1
80
Clients don't pay for code anymore. They pay for judgment: what to build, what to skip, and what will break at 2 AM. AI made the first part cheap. The second part just got more valuable.
1
3
67
Wall Street lost $2.2 trillion in a single day this week on AI fears. Meanwhile Nifty and Sensex were up ~10% in November, at 9-month highs. Same AI story, very different reaction. India is being priced as a user of AI, not a seller of it. Not investment advice.
2
129
Apple shut down its user-to-user Support Community yesterday. Some of the best fixes I've ever found came from a random user's forum post, not official docs. If AI chatbots replace community forums, where does the next generation of answers come from?
1
68
Günün en tuhaf gifini buldum 😂😂
50
187
20
5,355
441,996
Me ending every standup when someone asks ‘so is it done?’
1
3
136
A fake Claude Code extension got into Cursor's marketplace. Here's what happened, and the 4 things I'm changing in my own AI coding setup. 🧵
4
5
132
Change 4: run agents that have terminal access in a container or a separate user account, not on your main machine.
1
8
AI coding tools are now a supply-chain risk, just like npm. I'm putting together a 1-page checklist for React teams. Reply "checklist" if you want it.
1
1
10