Kishore Kumar

23, SWE @ Databricks, ICPC AWF, Codeforces Master.
Databases & LLMs. Ex-HPC & Compilers // Prev cs @iiit_hyderabad, swe @uber

[Kishore Kumar @theakcube]

New blog post!

Debugging distributed storage systems is hard. Over the last few months @databricks, we’ve been building a system that helps our engineers diagnose incidents faster using specialized agents and collaborative reasoning inspired by @DSPyOSS and other open literature.

https://t.co/i1dduvHift

PS: I dug into some of the lit survey & related ideas on my personal blog as well (links below).

[Kishore Kumar @theakcube]

@owen_venter Try out competitive programming

[Michael J. Miraflor @michaelmiraflor]

Getting paid to think is peak everything. If you’re getting paid well to think, don’t take it for granted. That’s 1% of civilization type stuff. Your ancestors are proud and jealous that you’ve gotten this far.

[Kishore Kumar @theakcube]

Really puts into perspective how impressive it is that @Google was able to scale search without any major hiccups.

All the deep research & AI mode queries fan out to anywhere from 10-100x the original load.

[Maggie Appleton @Mappletons]

I don't work on reliability & scaling at GitHub, but the people who do aren't bad at their jobs. They're dealing with unprecedented scale from agents.

It's easy to shit on GitHub from the outside if you're not in charge of 30X-ing capacity within a few months. Have some grace.

[Kishore Kumar @theakcube]

We're all here for the day too. If you're around feel free to come say hi / dm on X and we can meet! @_bhuvanesh_ @AnimeshSinha13 @adyanshkakran

[Jared Friedman @snowmaker]

We had room for 2,000 people at Startup School India.

More than 25,000 applied.

No Startup School anywhere in the world has ever had this many people apply. Not SF, not NYC, not London. India blew them all away.

[Kishore Kumar @theakcube]

Great read: https://t.co/bcVlrCyPx0

tl;dr:
"The value of human communication is distillation. When you write something, you can defend it. Saying nothing would have been better. A blank document at least doesn't pretend to contain decisions. A generated document creates the illusion of thought without the substance."

[Kishore Kumar @theakcube]

@NimaNaderi2004 https://t.co/NWZ62Ne2Tn

[Kishore Kumar @theakcube]

IIIT-H Programming Club is finally conducting a CodeCraft again after a 2 year hiatus! As a tester & alumni, I highly recommend giving the contest :) Good luck!
https://t.co/5XQWVDLcue

[Vinayak Kamath @vin_kamath]

If you're going to be in Bangalore either for #ycombinator #startupschool or #vibecon next weekend (April 18 - 19), here's a calendar of all the tech events happening in the city in one place. Included registration links, venues and dates.

#ycsus
https://t.co/am20YFlWsv

[Kishore Kumar @theakcube]

Great blog. That matrix transpose idea is one of my favorite "tricks" you can do with SIMD!

But I do think it's pretty hard to understand without some visualization, and I happened to remember where I dumped some screenshots from when I was first playing around with this. Hope it helps make it easier to understand for others too!

[Dr. Julie Gurner @drgurner]

What people don't tell you:

At the highest levels...discipline is the exception, not the rule.

[Kishore Kumar @theakcube]

I don't think simply sharing works well in practice. I've seen such internal "marketplaces" and the number of users using these shared artifacts is pretty low.

I believe most developers have custom battle hardened "ways" of doing things and small nuances in dev workflows that is very hard to generalize. The cost of building these artifacts (skills, hooks, etc.) is so low that it's almost always better to take inspiration from others and build it for yourself.

[Kishore Kumar @theakcube]

If I'm in a deep work zone I've managed to bump my comfortable limit to 5 now. But I believe the solution to parallelize more is ultimately just going to be better tooling, more tokens and prep.

How much tooling and tokens can I burn to close the loop on testing? If my agents are able to do e2e deployment / perf testing, and provide me a report of the tests done, that greatly helps bump the quality of the PR and also my confidence in reviewing it.

Similarly involving debate and simplifying the interfaces that will be built in a longer "plan" phase helps pre-aligning the agent on building changes in an incremental manner that's easier to review.

[Kishore Kumar @theakcube]

@kayareyouokay_ @fardeentwt Lol this is (and has been) pretty standard practice for most developers now for like the last 1 year+ or so. Both big tech and startups included :D

[Kishore Kumar @theakcube]

@daddy__broccoli @schematikio Cool!

[Kishore Kumar @theakcube]

@bindra_dhruv @zamana_hq Congrats on the launch and good luck Dhruv! 🚀

[Kishore Kumar @theakcube]

Thank you for sharing this, it is so incredibly well articulated!

How do you usually go about writing content & iterating with LLMs? I tried working with them a bit and quickly realised that I lost a lot of my personal understanding gains & creative ability if I started using them for structure and articulating arguments.

Wondering how you deal with this and how your workflow looks like.

[Kishore Kumar @theakcube]

@AudiPhotography So not the day I wanted to be getting on a plane...

[Kishore Kumar @theakcube]

@aadityansha_06 https://t.co/4WFgw1tk4O

[Kishore Kumar @theakcube]

@arpit_bhayani Great read. Related https://t.co/YSLEg5QlrF :)