karthik vetrivel

building something new
prev. @stanford ai/ml, @tesla ML, @nvidia infra

[karthik vetrivel @karthikvetrive3]

on the topic of tastemaking & choosing problems, this video by 3blue1brown on the role of AI in mathematics is wonderful

https://t.co/Tf0nEbfT59

[karthik vetrivel @karthikvetrive3]

even though i could probably find the same information in a few seconds, this quick retrieval and embedded knowledge contributes to the tastemaking required to pick valuable research & engineering problems

[karthik vetrivel @karthikvetrive3]

even though information is abundantly and quickly accessible, the best engineers I’ve worked with have incredible O(1) retrieval of highly technical systems, built from years in a given domain

[karthik vetrivel @karthikvetrive3]

i've spent the last few weeks chatting with dozens of PhDs at Stanford and have felt this deep, quick expertise in many of my chats

[karthik vetrivel @karthikvetrive3]

@dreamlabs_ai are there strong concurrency limits with a single API key? running some sims at scale :)

[karthik vetrivel @karthikvetrive3]

@praneethsamine @dreamlabs_ai Looks great! I'm having trouble getting the API to work on basic simulation tasks w/ YAM & MuJoCo. What sim setup did you use?

[karthik vetrivel @karthikvetrive3]

@itsvedantgg tell me more, what are you thinkin here

[karthik vetrivel @karthikvetrive3]

@mattgapp A naive Claude query shows Ubicloud is a bit cheaper. Any reason they're not on your graph? Or are they in a different category of CI runners?

[karthik vetrivel @karthikvetrive3]

@pentestduck @MunariAI @aravraval05 What is the high level inference cost associated with using Dreamscale?

[karthik vetrivel @karthikvetrive3]

@xdotli how did you structure the compute for running 2500+ trials?

[karthik vetrivel @karthikvetrive3]

has anyone noticed that if you ask claude/codex to find demand for a product (not recommended ofc), it goes straight to job postings. It assumes that if a company is hiring someone to do X manually, they’re already paying for the problem.

[karthik vetrivel @karthikvetrive3]

@boson2photon how would you get that out of them

[karthik vetrivel @karthikvetrive3]

@madsf88 high protein snacks in non-rigid containers

recently been drinking (?) a lot of miils

[karthik vetrivel @karthikvetrive3]

@_varunnair as in you're looking to deploy robots into factories or run the factory?

[karthik vetrivel @karthikvetrive3]

install in 2 minutes: https://t.co/5qh5L2uSQx

chrome extension, MIT license. jev picks each category for about half a cent a week (so expensive...)

[karthik vetrivel @karthikvetrive3]

I had fomo from everyone building with jev, so I built a Google Calendar that color-codes itself

Every event gets its color the moment I type the title, before Google even finishes saving

open source, link below!

[karthik vetrivel @karthikvetrive3]

I think so. There are two different kinds of secrets, right?

1) App secrets: DB URLs and other third-party keys the repo needs. The sandbox can hold these once as env var since every harness reads them the same way.

2) Harness creds: your Claude/ChatGPT/other login. These can be injected per session via a proxy per user.

[karthik vetrivel @karthikvetrive3]

Why is every multi-harness coding agent tool (Orca, Conductor, etc.) BYO compute?

I'm looking for a tool that has
1) a chat interface like Codex / Conductor
2) managed persistent sandboxes
3) multi-harness

[karthik vetrivel @karthikvetrive3]

@jonmayxyz missed this, this is good to know! wondering if orca integrates well with vercel sandboxes

[karthik vetrivel @karthikvetrive3]

ideally, this tool can one-click install new coding harnesses and set them up for me :^)