C S Krishna
@CSkrishna
Artificial Intelligence Researcher & Practitioner; Author: UnReal Elections
You might like
Correct take.
Hot take: DAgger (Ross 2011) should be the first paper you read to get into RL, instead of Sutton's book. Maybe also read scheduled sampling (Bengio 2015). And before RL, study supervised learning thoroughly.
We run full multiple tool call agent trajectories.
In case you didn't know my recent work Single-stream Policy Optimizaton (SPO), a group-free low variance policy gradient algorithm. Check this blog out: zhongwenxu.notion.site/Single-stream-… and paper: arxiv.org/abs/2509.13232
To summarize my thoughts across tweets: Policy gradient algorithms form the foundation for optimizing reasoning and agentic LLMs, with the REINFORCE algorithm at their root. This is because maximizing expected return is achieved via the log-likelihood trick. Modern algorithms…
if modi withdrew indian troops from ukraine this war would be over.
White House trade adviser Peter Navarro sought to raise pressure on India to halt purchases of Russian energy after the US imposed crippling new tariffs on New Delhi, casting the conflict in Ukraine as “Modi’s war.” He speaks with @jmathieureports bloom.bg/45FF9tW
Modi finally nominates Trump for the Nobel peace prize for getting India and China closer to each other. "Thanks to Trump bhai, it's Hindh-Cheeni bhai bhai once again," Modi wrote in his nomination letter 😅
India is essential in helping the U.S. move its supply chains from China. While Pres Trump works to bring manufacturing home, India stands alone in its potential to manufacture at China-like scale for products that can't be quickly or efficiently produced here. My latest w/…
i haven’t read enough Puranic literature to be certain, but the fact that western indology ~discards the historical usefulness of this entire corpus feels wrong yes Purāṇas intersperse myth, but even Thucydides includes e.g. oracular prophecies in his History
The "NASA finds Sanskrit is the best language for programming" is a nonsense spread long enough diminishing the real greatness of Sanskrit. But I have a strong intiution that Sanskrit, with its Paninian grammar and rigor in Astadhyayi sutras, will be a great language for AI…
I just came back from Independence Day celebrations in our Kalaivani Kalvi Maiyam school, filled with cultural programs, patriotic songs, sweets and (mercifully for students!) a short speech where I told our students that preserving our hard-won independence means working hard…
idk why people say that the VC theory doesn't explain deep learning it looks pretty clear to me that frontier labs managed to figure out generalization under pressure from their VCs
As AI models get commoditized, the value will be added in that final layer of orchestration. Not just routing to just one "best" model, but coordinating multiple models to combine strengths and create Chain of Debate.
New paper: Reflective Prompt Evolution Can Outperform GRPO. It's becoming clear that learning via natural-language reflection (aka prompt optimization) will long be a central learning paradigm for building AI systems. Great work by @LakshyAAAgrawal and team on GEPA and SIMBA.
How does prompt optimization compare to RL algos like GRPO? GRPO needs 1000s of rollouts, but humans can learn from a few trials—by reflecting on what worked & what didn't. Meet GEPA: a reflective prompt optimizer that can outperform GRPO by up to 20% with 35x fewer rollouts!🧵
There is little journalistic merit in tossing questions at Pakistani politicians about groups they neither control nor dare to name. These men and women are not decision-makers; they are fronts. Every interview with them seems scripted, evasive, and utterly pointless. If honesty…
Mahanadi Roaring ! Hirakud Dam on Mahanadi River in Odisha discharging flood through all the 34 sluice gates. Level - 616.6 ft/630 ft Inflow - 5,79,680 Cusecs Outflow - 5,60,614 Cusecs
Just realised, many of the Aman ki Asha with Pakistan types are now working for Al Jazeera.
What's India's potential? Well, Singapore previewed what China could look like. And Dubai may preview what India could look like. Because Indians are 38% of Dubai's population. While South Asians are >50%. And by working with Emiratis... They built a city that looks like this:
Saif Ali Khan's best sporting success so far may have been winning 100 meters sprint in an Amul Macho ad. youtube.com/watch?v=wbMuDR…
youtube.com
YouTube
Amul Macho(Saif Ali Khan) Advertisement
Did you know a pretrained graph transformer can tackle virtually any predictive ML problem? That means accurate predictions for churn, fraud, and recommendations without the endless cycle of feature engineering, model development, and training. This isn't just a concept. Our…
Chapter 7 of Concurrency Control and Recovery in Database Systems: Distributed Recovery. (from a cafe in Bologna)
The Chenab Rail Bridge isn’t just an Indian achievement it’s a testament to what humanity can build with vision and determination. Wishing for its destruction speaks volumes about a mindset stuck in animosity rather than advancement. And the irony is, when India takes a firm…
United States Trends
- 1. Jeremiah Smith 1,470 posts
- 2. Sunderland 93.6K posts
- 3. St. John 5,507 posts
- 4. Texas Tech 9,681 posts
- 5. #GoDawgs 3,403 posts
- 6. #iufb 1,224 posts
- 7. Mendoza 7,964 posts
- 8. Saka 32K posts
- 9. Obamacare 176K posts
- 10. Merino 10.3K posts
- 11. Mississippi State 3,887 posts
- 12. Philon N/A
- 13. Nate Frazier N/A
- 14. #SUNARS 7,831 posts
- 15. Ryan Graves N/A
- 16. Lebby N/A
- 17. Aden Holloway N/A
- 18. Letang N/A
- 19. Verstappen 24.8K posts
- 20. Shapen N/A
You might like
-
T.Hayase
@ThayaFluss -
Julia Kiseleva
@julia_kiseleva -
Irwan Bello
@IrwanBello -
Ashwin S Kumar
@ashwinskumar -
Kale-ab Tessera
@KaliTessera -
Yuri Campbell
@yuricampbll -
Brendan O'Donoghue
@bodonoghue85 -
Anirudh Thatipelli
@AThatipelli -
Kaká
@firqaaaa -
Krisha Mehta
@KrishaMehta2 -
Adam Kelleher
@akelleh -
James O' Neill
@JamesONeil21 -
🐈qeel🍉
@catqeel -
Tong Liu
@TongLiu15731072 -
Cozmitz
@CosmitzOMEGA
Something went wrong.
Something went wrong.