AI researchers travel to South Korea for the International Conference on Machine Learning

@philhchen I struggle to believe that data and discrete structures, logic, and algorithms are an incorrect spike to try and get from interview panels. Leetcode may be broken, but fluency in the language of compsci is still of top tier importance. You will mishire without a DSA of some sort.

@DeadeyeRolf we still evaluate logic and fluency with problem-solving, just in solving harder problems with the help of agents.

omw to icml! 🇰🇷 기대된다~

@philhchen @pangram is this ai

@johnrtian @philhchen We believe that this document is fully human-written

https://www.pangram.com/history/0ec63209-cadd-43e5-952a-08ae0cb741ce

@philhchen wisdom from one of the best tech leaders in the industry

a great read indeed. two questions: 1. in what sense do you see humans as having durable differential capability at selecting meaningful problems for ASI? if general methods keep subsuming task-specific human structure, why wouldn't problem-selection eventually fall inside the graded region too? curious whether you think it's a permanent human edge or just the current frontier of "not yet gradable." 2. as a young founder, how do you actually know you're working on the most ambitious form of your problem, versus just believing you are?

@philhchen But for those like me trying to launch their career at the very start, it feels almost impossible. It’s a real rough cold start problem, and ATSs are a near impenetrable gate. How on earth are we supposed to get someone to take a chance on us, or prove soft skills?

@philhchen really good advice phil :)

@philhchen Thanks for sharing this, I learned a lot. My question is how do you find the right problem that helps best and helps maximize learning and helps get people to notice you. Second question is how can persons from a third world country be able to get access to opportunities and

@philhchen banger

@philhchen love this Phil, esp the point on last mile. With production in abundance, polish really shines through. Craft, intentionality and care can still separate you from some generic in-distribution output

@philhchen great advice!

@philhchen usually I don’t like articles, always felt like they long AI slop.

my friends always spoke about connecting w people who resonate your thought. Understood this is what it feels like. Shoot to kill is the idea, and I can’t love it enough

@philhchen The problem finding part matches what we see in hiring too. Give someone a messy repo and an agent, and the strong candidates spend their first minutes deciding what's actually worth fixing.

@philhchen Super unique and extremely helpful to parents

1. I think there's a mix of "not yet gradeable" and the (perhaps naive) belief that humans will continue to stay a step ahead.

Today models are trained on software tasks that take weeks, and those training runs also take weeks to months. Perhaps next year they will be trained on tasks taking months, but those training runs would also take months. We've never seen signs of AI short -> long time horizon generalization as strong as for humans, but then again it's hard to predict what's possible.

2. I'm super bullish in our own direction based on having seen what the labs are working on, but I don't claim to have any advice for founders! We're early in the journey too

@philhchen Great read!

@pangram @johnrtian @philhchen Nice

@philhchen funny isn’t it. I don’t think it’s even reductive to say that ‘people buy from people they like’ is the biggest signal… we’ve closed so many people against top labs following this idea