arunabh
@ArunabhMishra8
Followers
254
Following
1K
Media
86
Statuses
1K
CV @ Zoox. opinions my own. will post pictures of dog.
Boston, MA
Joined June 2019
Faster than Gemini 2.5pro. Smarter than codex.
cursor has won coding their data flywheel has turned into the most amazing coding model in the world. Composer-1 truly feels like another Claude 3.5 Sonnet sized jump in capability
0
0
3
Vibes are really really good. 👨🍳
So cool I can finally talk about our coding model. Guys this has been my go-to coding model for the last month or so. Its SO fast and accurate. Try it out, you wont regret it!
0
0
4
Simulation doesn’t replace physical testing; it complements it, refining hardware and software designs before they enter the real world. For @WilliamsRacing and Zoox, simulation plays a key role in transforming engineering ambition into performance. https://t.co/pqZsUpLdtn
4
17
79
First couple hours of coding with composer and I don't think I can go back.
Introducing Cursor 2.0. Our first coding model and the best way to code with agents.
0
0
3
Cheetah 2 has been RLed and unleashed. It's a good model, sir.
Meet Composer 1 and Cursor 2, have fun building ✌️
1
0
17
Very nice.
🧵 LoRA vs full fine-tuning: same performance ≠ same solution. Our NeurIPS ‘25 paper 🎉shows that LoRA and full fine-tuning, even when equally well fit, learn structurally different solutions and that LoRA forgets less and can be made even better (lesser forgetting) by a simple
0
0
0
@ArunabhMishra8 Sharing a training pass between multiple models is 🤯 I dont know the business case to do this, but its cool for sure
0
1
1
Good robot making waves.
Introducing our latest Las Vegas partner. They’re spheriously cool. @SphereVegas Head to our Journal for all the details: https://t.co/03zxdxbTI0
0
0
4
Updated with more juicy details!
1
0
2
Muon’s speed arguably comes from approximate orthogonalization: a few fast Newton–Schulz iterations instead of an expensive full SVD. But this makes the update inexact. So how should hyperparameters change as approximation quality varies?🧵 1/n
5
24
287
Jk- just posted on slack about getting more capacity. None of my shit got scheduled.
0
0
0
If you don't have alarms at 2:45 to debugging torch internals for your training you're ngmi
1
0
0
Condition with action then text then everything at once! Looks incredible.
Odyssey-2 is now live (and GPUs are scaling up)! It's instant, interactive AI video that works like magic. This isn't a blog post announcement. This is a model that's available to try (for free) right now. Have fun! https://t.co/OoL9oVtXkO
0
0
2
Guys, we converted vik.
0
0
4
🫡
At 10AM PT tomorrow, we launch Odyssey-2! It's a different exploration than our previous models. There's no WASD control. But it feels like something totally new, and like AI video for the rest of us. Instant, interactive, fun. Proud of the team & really hope you enjoy it 🥹
0
0
2
A new essay on the crazy, all or nothing approach to work happening in AI today, the looming human costs, and the lack of a finish line. I wouldn't say it's okay, but I'm not sure how to fix it.
17
52
601
Found an old note
A new essay on the crazy, all or nothing approach to work happening in AI today, the looming human costs, and the lack of a finish line. I wouldn't say it's okay, but I'm not sure how to fix it.
4
13
342
I am a little disappointed at how rarely dot-product attention mechanisms are explicitly described as differentiable lookup tables. It seems like it would just be an easier way to convey the intuition
35
23
749