Haresh Karnan

Introducing STERLING 💫: "Self-Supervised Terrain Representation Learning from Unconstrained Robot Experience" 💫 to be presented at CoRL 2023 in Atlanta! 😀 Project Page 🌎: hareshkarnan.github.io/sterling/ Paper 🗞️: openreview.net/pdf?id=VLihM67… A thread 🧵

@hallerite @emaadmanzoor Curious, what in verl do you find is painful ?

@emaadmanzoor verl is a pain to work with and verifiers is an environments framework first and foremost

I see the need for 2 different kinds of LLM-RL frameworks. One should be optimized for post-training of LLMs by labs that want to build frontier models. It doesn't have to be feature-complete, but should follow best practices and consolidate what we learn over time.

@ThereseMaggie @_arohan_ I meant the IdlyExpress / Mylapore chain of restaurants (you’ll find them in South Bay). I remember reading somewhere they have an idlyexpress branch open in SF recently

@KarnanHaresh @_arohan_ who is they? Copra?

Number of VCs liking this from SF makes me think … there is space to raise money for a startup to make frontier roast dosas.

@ThereseMaggie @_arohan_ Didn’t they open an idly express in SF?

@_arohan_ the white space for delicious South Indian food in SF massive. It’s a shame that whole of South Indian food is reduced to just the humble dosa. The only other restaurant that serves lesser known South Indian food is Copra, but that’s mostly experimental, although delicious.

The Cost of Ambiguity: There’s a massive, painful gap between knowing a mental model and applying it exactly when the conversation needs it most. Every Founder, PM, and Exec we spoke to knows this pain. And the result: Meetings dissolving into opinion, or worse, ending with a vague "let’s revisit this next week". 😩 Introducing 🥁 Vana AI. It's not a note-taker. It's not an action-item generator. It’s your proactive AI strategic thinker for real-time insights in meetings.

@Devvrit_Khatri @louvishh @rish2k1 @rach_it_ @dvsaisurya @brandfonbrener @agarwl_ nice work! By any chance, did you ablate the sampling temperature in GRPO/DAPO across tasks?

Wish to build scaling laws for RL but not sure how to scale? Or what scales? Or would RL even scale predictably? We introduce: The Art of Scaling Reinforcement Learning Compute for LLMs

@jbohnslav Time to bring back inperson interviews

just wasted an hour of my time interviewing some obvious cluely-using motherfucker. do I have to conduct all interviews blindfolded? mirrored glasses?

@gowthami_s arxiv.org/pdf/2509.04419 relevant work on interleaving SFT and RL

Why is there a weird distinction in post-training, like SFT first and then RL? Why can't they be done together? IMO, I see merit in interleaving these paradigms. Is there any research pointing to the contrary or supporting this sequential process?

@ArmenAgha How does this compare to mds ?

We've internally built a streaming solution to stream 1PB of multi-modal data over hundreds of GPU's for weeks without ever touching NFS, with no train performance regression. Enough interest for a blog post?

@iScienceLuvr Tesla is an appreciating asset ?

Did they beat Tesla to the Tesla strategy? Lol Basically, they are deploying "self-driving" humanoids to consumers, it might not initially work that well and needs a lot of human supervision, but the data collected will continue to improve the robot and get it closer to "full self driving". I would note that humanoids in a home is a much more challenging environment than cars on the road: everyone lives differently in a variety of different floorplans and environments while roads have more consistency. So I do wonder how long it's gonna take for humanoids to reach near full autonomy. That said, like Teslas, Neo could be an appreciating asset if it continues to get better and more autonomous. Anyway that would be the hope, right? Let's see what happens in practice. I'm cautiously optimistic! I wonder when Tesla Optimus will be available 🤔

@jaredpalmer @github @pimoroni waaaant!

Your @GitHub Universe badge comes with a hackable Raspberry Pi built into it with full color display, 5 buttons, USB-C, Bluetooth, and WiFi. Built by @pimoroni

@natolambert Congrats! 🎉

Life update, she said yes. 🤩👩‍❤️‍👨🐕‍🦺

@natolambert 💯

Retiring my original airpods pro after over 5 years of use. Ridiculous amount of value really when we're used to 2 year product cycles.

@sivareddyg @YejinChoinka Thanks for sharing! Is the full recorded talk published on YouTube?

Lot of insights in @YejinChoinka's talk on RL training. Rip for next token prediction training (NTP) and welcome to Reinforcement Learning Pretraining (RLP). #COLM2025 No place to even stand in the room.

@willccbb @trungthvu verl does support all of that. The default params aren’t the most efficient

the sampling is fine, the biggest red flag is that they need 12 GPUs for backprop and don’t do any microbatch packing / grad accum to accommodate for async/server on 16 GPUs i’d want 2-4 for training + 12-14 for inference you can do a loooot of backward passes in the time it takes to generate 512 rollouts

"veRL is the best RL framework it's super efficient" really. are you sure about that. are you sure that you need 16 GPUs to tune a 7B model at 8k context. do you think that it's reasonable each step takes 19 minutes for this

@public_steve @kohjingyu have you heard of "uvx nvitop" ?

@kohjingyu We ride!

My toxic trait is thinking I can train a frontier model with 8 H100s

@Alibaba_Qwen Can your guy see images and videos?

Ready to meet the biggest, brainiest guy in the Qwen3 family?

@willccbb That sounds a lot like o3

wow. thanks

@KyleMorgenstein I miss Austin summers! I’m hearing this summer is milder than past years

despite living here 5 years this is somehow my first summer in Austin. I was warned it would be brutal but honestly it hasn’t been that bad. the key, imo, is to spend as much time outside as possible. I run or row most days. ofc if you spend all day with AC blasting you’ll melt!

@jbohnslav 💀

> be me, training vlms > use hf transformers because who wants to reimplement models if they don't have to > low MFU > `pytorch_profiler.py` > ViT takes 40X longer than the LLM > sus > .item() in the forward pass makes cudaStreamSynchronize every attention layer > MFW

@g_k_swamy ziebart et al. :’)

This story makes me smile every time I get to tell it. Back when I was an undergrad, a bunch of my friends who were much smarter than me were getting into this thing called "research." So, I also cold-emailed a few professors and one got back to me, asking me to read some paper: