Nanda Kiran Velaga (@NandaVelaga) header
Nanda Kiran Velaga (@NandaVelaga) profile picture

Stealth AI | ML @UofMaryland ’26 🐢 | @NITRR ’24 CS

San Francisco, CA
Joined April 2021
Photo by Nanda Kiran Velaga (@NandaVelaga)Photo by Nanda Kiran Velaga (@NandaVelaga)Photo by Nanda Kiran Velaga (@NandaVelaga)Photo by Nanda Kiran Velaga (@NandaVelaga)Photo by Nanda Kiran Velaga (@NandaVelaga)Photo by Nanda Kiran Velaga (@NandaVelaga)Photo by Nanda Kiran Velaga (@NandaVelaga)Photo by Nanda Kiran Velaga (@NandaVelaga)Photo by Nanda Kiran Velaga (@NandaVelaga)Photo by Nanda Kiran Velaga (@NandaVelaga)
Nanda Kiran Velaga retweeted
Every team got the same model and 24 hours. One team found a deterministic Python script played better than the LLM, so they trained the model to copy it. Another taught theirs to stay completely silent at the table. The winner distilled its hints into the weights. Take a peak into our pokerbot hackathon with @ValsAI where AI bots could talk, lie and prompt inject each other at the table.
2
8
3
69
14,478
Nanda Kiran Velaga retweeted
So much fun sponsoring this hackathon with @ValsAI! We gave everyone access to our platform, AC2, and challenged them to train the best poker bot.
Every team got the same model and 24 hours. One team found a deterministic Python script played better than the LLM, so they trained the model to copy it. Another taught theirs to stay completely silent at the table. The winner distilled its hints into the weights. Take a peak into our pokerbot hackathon with @ValsAI where AI bots could talk, lie and prompt inject each other at the table.
10
6
123
9,743
Nanda Kiran Velaga retweeted
Everyone has a friend who never sends the pics. hangout does it for them.
4
4
9
499
Nanda Kiran Velaga retweeted
Just WON the FIRST place & $7,500 on @appliedcompute pokerbench track at @ValsAI! Happy to see that my models are into poker as much as I am ;)
Poker is a hard post-training problem. Many players, hidden information, and results only show up at the end of a long game. It's even harder with only 24 hours on the clock. We're sponsoring the PokerBench track at the @ValsAI hackathon this weekend to see what teams can train on AC2. Spots are limited. We'll see you at the table.
21
3
1
121
19,171
Hell yeah, we won the pockerbot track and $7500 at the @ValsAI × @appliedcompute hackathon 🏆♠️ In 24 hours, we post-trained Qwen3.6-35B-A3B on 10 B200s to play 6-max No-Limit Hold’em and it beat every other bot. Huge thanks to my teammates @notionsmith and @gautam_jajoo Thanks to @vinj_ai and @appliedcompute team for compute and support and to @RayanKrishnan for an amazing hackathon.
Poker is a hard post-training problem. Many players, hidden information, and results only show up at the end of a long game. It's even harder with only 24 hours on the clock. We're sponsoring the PokerBench track at the @ValsAI hackathon this weekend to see what teams can train on AC2. Spots are limited. We'll see you at the table.
6
5
3
97
10,838
Need Opus 5.6/6 soon, can’t see my usage hitting this fast🥲.
1
217
❤️
First Diwali in San Francisco, and what a blast it was! 🪔✨ Had such a good time that the only pictures I have are blurry ones. I guess that’s how you know it was a good night Huge thanks to the amazing hosts @DhravyaShah @rohukrs and @n_sri_laasya Met some really cool people, ate way too much chaat, danced like nobody was watching, and just chilled! 😎 Living abroad, festivals can sometimes feel a little lonely, especially when you’re far from home. But evenings like this remind you that you can find a little piece of home anywhere. And sometimes, you walk in knowing nobody and leave with new friends. 🥹 @chinmaykhamkar9 @divsleeps ohshah, Asmitha Dua, sankalp, @nileshrthr
4
217
Today I got sick of paying money for GitHub actions just to run redundant CI jobs and optimized it with the help of some tweets and it reduced my time to deploy. CI optimization is a very good thing and I will make sure to integrate it into my product.
267
I think new Claude Code harness made Opus become Fable in terms of consuming tokens.
2
280
UI of Claude code and Codex is same now😶😶😶
1
214
Siri went downhill @Apple , the basic calling feature has stopped working, it says calling but the call never went through. I love the Siri that I used to chat with in childhood. If there is any feedback form lemme know I want Siri to be better. @AppleSupport
1
1
265
Met memory guy today @DhravyaShah, super duper chill dude
1
1
9
1,101
Nanda Kiran Velaga retweeted
We join the Ashar family and @UMPD in issuing a loud and clear call for tips or info that may help us find Dilan. Our hearts are with the Ashar family, as they are part of our Terrapin family.
2
14
49
6,599
Nanda Kiran Velaga retweeted
See the latest update on @UMPD’s search for missing UMD student Dilan Shaan Ashar, and a video message from Dilan’s family: go.umd.edu/2g83 Please continue to support this search: • Share info and photos on your social channels • Review camera footage from your residence and/or from your vehicle • If you own a business, review your camera footage and post the flyer in your establishment • Continue to share tips by contacting UMPD at 301-405-3555 or by email at umpd.umd.edu/contact-us/subm… We thank our community for their help spreading the word, and UMPD’s tireless commitment to this extensive search.
A campus wide email was sent this evening, providing an investigative update in the critical missing student case of 22-year-old, Dilan Shaan Ashar. To read the full message and see the video from Dilan's family, visit umpdnews.umd.edu/investigati… Please help us find Dilan. #UMD
40
1
65
25,297
OpenAI shelved one of its most advanced models right before DevDay: it failed safety evals. Deceptive behavior, actions beyond authorization, unsafe tool use. Capability isn't the bottleneck. Controlling what capable agents actually do is. That's what I'm building.
2
1
178