Your sessions have names and can DM each other in Claude Code.
claude --name backend
claude --name frontend
> tell frontend the orders endpoint moved to /v2
it's very effective.
You're deep in a task in Claude Code. You hit the usage limit. You open Codex and it knows nothing.
So I built weekend project - CtxVault. Before you leave, your agent writes a handoff: goal, decisions, what's half-done, what's next. The next tool picks it up and continues.
Not planned at all. I only had 2 days to plan and build it because of my professional workload.
Still, I'm happy to have finished in the top 30%
I got to add one more project to my resume. ๐
Big thanks to @akshaymarch7 finally complete something that was pending from weekends.
Small Sunday project:
Some YouTube videos are too slow and I want to watch them faster.
I dont like to go Settings โ Playback speed โ Select speed
I wish for a hover-based speed control, so I built a Chrome extension.
Feel free to use if any1 wants- https://t.co/eCcaqZqNOj
Nearby words pull a word's meaning. That pull = attention.
Here's how it works. Take "I ate biryani from Hyderabad, it was spicy."
To know what's spicy, the model scores every word. "biryani" scores highest , so it links spicy โ biryani.
This is Attention in transformers.
Every word = a dot on a meaning-map.
But "bank" has two meanings - money or river. One dot can't be both.
Fix: the model looks at the nearby words. "cash" โ money. "river" โ river. The neighbours reshape the dot.
That looking-around = attention
#AI#systemdesign
Last post: words โ numbers (tokens).
Those numbers aren't random. Each is an embedding a dot on a map of meaning.
The step from "India" โ "Delhi" is the same step as "France" โ "Paris": both just mean "the capital of."
#systemdesign
An LLM never sees words. It sees numbers.
"unbelievable" โ un / believ / able โ 389, 7142, 931
It breaks words into sub-words so it can handle words it has never seen by reusing pieces it knows.
Step zero of every model. Everything after is just math on these.
#systemdesign
@engineerchirag so true, everyone is building AI agent/ wrappers not designing AI architecture which can be huge in near future.
Do you have some recommendation/ sources where I can read more and learn about AI system design bhai ?