kyle

witness me

Get out of that spaceship and fight me like a man

@suribp25 If you don't smile at the thing you're thinking about doing then it's not meant for you, something else is out there

@kylecompute So, we’ve moved on from the “mom’s test”? What’s the smile test?

coining a new term called the smile test

Okay we're back, it's time to go through 100 papers this weekend but RIGHT after my non-negotiable, scheduled 1 hour long melodramatic walk

@bebeal_ I was curious and tried this with a $25 1080ti years ago

@bebeal_ They send you literal rocks inside the box

someone explain this scam to me — what happens if you order this

being in san francisco has genuinely reduced cycle time to failure / success by like 80x for me

@growing_daniel Don't forget head of growth

The two future jobs are anthropic employee and sex worker servicing anthropic employees

@Zakariea_sh Have you guys explored sim data

Bad data is not allowed

@xandercogan It always can

@kylecompute What if everything goes right

probably one of the most important thing i've learned over the last two years is how to gather data market quickly through questions you should have a set of questions to ask that go from 1. does the general landscape i think exist, actually appear in reality? example: do you know of anyone at a smaller robotics lab that deals with purchasing training data? 2. more precise market questions on things you discovered from questions 1 (you should have rough hypotheses written prior and then update them based on data you get) example i've used: when a policy fails in eval today, what actually happens? is it (a) collect more real data, (b) hand-author edge cases in sim, (c) just accept the failure rate 3. final questions that end up tying to dollar amount companies would be willing to pay example: if a company could get x% lift on y benchmark over the standard pipeline. would you pay (market number) for said dataset this gives you a much better understanding of the landscape you're trying to capitalize in because you can magnify a bunch of different subcategories while getting better lay of the land when combined the responses form answers to questions you weren't are even existed

@vussyviz @distributedkv I'm strongly of the opinion that the movie is characteristically flawed and the book has a malformed copy paste hero's journey. 5/10

For context kyle is very passionate about this

Model company raises a lot of money and gives them to the data company

I think we're going to have 100% recyclable materials before incredibly cheap space freight, but I'm of the opinion that we should launch trash into the sun for novelty and narrative reasons at least once

@vussyviz @distributedkv @finiteopencover i will be trying out the speed boost irl by rolling

@vussyviz @distributedkv @finiteopencover highkey relevant x.com/kylecompute/st…

Just one tonight

i was born in the right timeline i can wake up, run thousands of robot rollouts with mjx on one gpu. actively running a thousand little minions to help me find where a policy fails specifically robotics is going to get weird so fast

@wavefnx There's a ton of work to do before getting a technical co founder, market validation is an art lowkey

I find funny when startups look for "technical co-founders" So like, you haven't been doing shit until now

@FeduniakS do you know if getting more volume in general was helpful or targeted spots where the policy specifically was failing?

Spent last week benchmarking policy speedup methods. Then we just collected faster data and it beat all baselines... Although obvious, but turns out first step to speed up your policy is … collect faster data.

@gpusteve Your name is gpusteve how do you not have a Ron

@kylecompute can i befriend ron

amex black card should actually just come with a concierge who finds b200s on demand for u