Jerry Zhang

fine-tuning my jumper | @princeton

[Jerry Zhang @usbastats]

People irl: hey man, how’s it going

[gaurav @gaxrav]

celebrating what? what have you achieved with your fucking life?

[Jerry Zhang @usbastats]

@bitacolyte Relying on trips to make life feel more meaningful doesn’t work. Change and fulfillment come from within

[Jerry Zhang @usbastats]

the asia trip after graduating undergrad is the signal that your life is over

[Jerry Zhang @usbastats]

@xwyul It cancels out if you never graduated high school

[alphaXiv @askalphaxiv]

Introducing autoresearch for arXiv papers

Change 'arxiv' to 'autoarxiv' in any paper URL

An agent deploys to resolve setup issues on the codebase, run a minimal reproduction, and estimate full replication cost. Read more below

[Jerry Zhang @usbastats]

@rivermillion @ysu_nlp "a few hundred Ks"

[Jerry Zhang @usbastats]

@MangQiuyang don't sleep on fudan

[Jerry Zhang @usbastats]

@juliali touristmaxxing

[Jerry Zhang @usbastats]

i win in the end

[banteg @banteg]

it all makes sense now. dario was still at openai in 2019. he left next year and took his marketing playbook with him. hasn't changed a thing since.

[Jerry Zhang @usbastats]

Some lectures you just can’t fade

[Jerry Zhang @usbastats]

@gabriel1 

[Jerry Zhang @usbastats]

@aleabitoreddit Are you as confident about $AAOI as you are with $AXTI or $SIVE?

[Abiral Shakya @abiralxshakya]

Wrote some thoughts on the deployment gap between scaling robot foundation models and the inference serving infrastructure that .. doesn't exist yet!
https://t.co/RTfFpGyxZd

[vik @vikhyatk]

ML interview question:

Here are the weights for Llama 3.1 70B. Generate a token by executing the forward pass manually using pen and paper. You have 30 minutes.

[Jerry Zhang @usbastats]

alphaxiv is so useful when it works, but the search agent is legit the worst i've ever used. 8/10 search queries crash

[Jerry Zhang @usbastats]

The optimal PPO setup from our training also generalizes to more difficult tasks like Walker2D, HalfCheetah, and Ant (with slight changes to account for task-specific constraints).

Notes: PPO as the algorithm is held constant. In the future, the agent should be able to iterate on the choice of algo as well. Many environments require longer rollouts to achieve optimal performance. 12-minute rollouts suffer from short-horizon bias and may not generalize to longer training runs.

[Jerry Zhang @usbastats]

I adapted @karpathy's autoresearch for RL.

Given a locomotion RL environment, the agent continuously optimizes a PPO algorithm to solve the task.

I ran it for 6 hours, 29 experiments overnight, nearly solved BipedalWalker-v3 on a single MacBook CPU.

https://t.co/o2k8jN8En5

[Jerry Zhang @usbastats]

Ran into this guy in nyc a couple weeks ago, first thing he said was “do you know who Greg Yang is, big things coming.”

Congrats on the new job bro

[Rich @richzou]

Very excited to share the news I’ll be joining @xai, focusing on talent.

I grew up chasing the American Dream the only way I knew how - watching my immigrant mom work nonstop so I could have a chance to try. From Chinatown to SF, the path hasn’t been linear, but it’s been life changing.

I wrote more about that journey below.

Grateful to @TheGregYang and @barisakis for the opportunity and trust.

As my mom always says: be kind, be honest, take care of people.

We’re building something special here.

Would love to meet thoughtful builders and kind humans.

Time to make @grok 🚀🚀🚀

[Jerry Zhang @usbastats]

in college you meet people who remind you of yourself, and your interactions with them make you realize you're becoming more similar to your friends that you used to disagree with

[Jerry Zhang @usbastats]

@saurishk skill issue