"theoretically smart" - @arjun_ramani3

California, USA
Joined October 2012
Manan Shah retweeted
To push the open source frontier for RL + LLMs, we need scalable, modular environments with real-world complexity, beyond math benchmarks. Today, we’re releasing *benchmax*. An open-source framework to build, run, & scale useful RL envs for LLM fine-tuning, with integrations to verl & verifiers (more coming soon!).
4
24
2
85
16,374
Sharing some cool RL results @manans99 & I just witnessed in reasoning models for science! Our research question: 🤖 LLM performance limits agentic capabilities. 🎓 Humans learn new skills by practicing them. 🔬 Can LLMs learn to reason by solving graduate textbook problems?
1
2
1
7
1,649
Training & serving LLMs can often feel like "magic" -- that perceived simplicity is possible thanks to beautiful innovations in system design under the hood. Building intuition for LLM systems: part 1—low-latency, on-device inference using ggml ↓ mananshah99.github.io/blog/2…
3
1
13
649
Part 2—on the mechanics of @ollama's beautiful on-device language model "containerization", scheduling, and serving experience: mananshah99.github.io/blog/2…
1
2
270
Manan Shah retweeted
my favorite part of this project was showing that we have general ways of spending much more than the limit of ~$20 per problem to continue to improve scores. it’s incredible that we live in a world where you can trade compute for more intelligence
New verified ARC-AGI-Pub SoTA! @OpenAI o3 has scored a breakthrough 75.7% on the ARC-AGI Semi-Private Evaluation. And a high-compute o3 configuration (not eligible for ARC-AGI-Pub) scored 87.5% on the Semi-Private Eval. 1/4
10
10
6
185
27,915
Manan Shah retweeted
Curious about what I’ve been working on over the last two years? Join me next Thursday for the launch of Kumo.AI’s groundbreaking declarative ML platform that will transform the world of predictive AI as we know it. 🚀 More at info.kumo.ai/revolutionizing… #AI #revolution #GNNs #deeplearning
5
31
1
306
62,029
Still can't believe I had the opportunity to present our work on teprotumumab-related adverse events at the European Thyroid Association's conference in Milan last month! Grateful to my mentor @KosslerMD for this incredible opportunity!
1
3
14
2,731
Manan Shah retweeted
Kumo team at the @databricks summit! We're proud to partner with the Databricks team to bring ML to the enterprise Lakehouse. Excited to kick off the event with this stellar crew!
2
7
1,385
Manan Shah retweeted
We are very excited to announce our annual Graph Learning Workshop on September 28th. You can learn more about the event here: snap.stanford.edu/graphlearn… We’d love for you to join us! Register here to reserve your spot: eventbrite.com/e/stanford-gr…
2
33
208
Manan Shah retweeted
I write constant notes to myself which mostly feel useless in the moment but sometimes I look back through them and it’s *amazing* to have a record of exactly how you felt at one particular moment in your life. This was... 2018?
9
7
5
340
Manan Shah retweeted
Wow, the mental health effects due to the pandemic are extreme , across nearly every way of slicing the population (gender, age, race/ethnicity, income level, education, urban/rural). That right hand column: 👀 Please check in with your family/friends & suggest help if needed!
8
82
5
288
1/ Covid (@UCSF) Chronicles, Day 150 Today, 150 days since I began my Covid tweets, I’m going to do something odd: write the speech that Trump should give. I have no faith he’ll do so, but it’s worth recognizing how little it would take to change course & save lives. Here goes:
146
1,329
513
4,293
This is horrifying. It appears the system is biased against less affluent schools. It factored in grades of past graduates. This is how AI can be used to reinforce social and cultural disparities
This is actually unbelievable. In the UK, students couldn't take A-level exams due to the pandemic, so scores were automatically determined by an algorithm. As a result, most of the As this year - way more than usual - were given to students at private/independent schools. 😩
1
32
3
100
A very short history of some times we solved AI togelius.blogspot.com/2020/0…
13
53
28
228
I’ve been somewhat puzzled by the intensity of my own annoyance with AI hype. I think I’ve just discovered in this rant that, beneath irritation with badly-motivated, exaggerated claims, there’s deep sadness that the scientific opportunities are being lost. >
3
2
2
47
Hurts foreign students. Hurts America's universities. Hurts freedom. Hurts America. Hurts the World. "Foreign college students could be forced to leave the U.S.... according to guidance released by Immigration and Customs Enforcement (ICE)." axios.com/online-classes-col…
23
184
29
551
just spent 5 hours figuring out whether or not a proof was obvious
28
45
8
612
Manan Shah retweeted
Arrogance has fueled a lot of Silicon Valley’s response to coronavirus. It takes humility to consider that you might be wrong. It takes even more humility to understand how being wrong can be harmful 👇 stanforddaily.com/2020/04/11…
2
2
3
20
Manan Shah retweeted
Startup idea: a clothing store for college students called “It’s a Good Fit at This Time” where you can show a rejection letter from this week and get 20% off
1
1
34
Paper accepted to #cikm2019 on improving image classification using contextual information! Grateful for support from @GoogleAI and @sigir2019 to attend the conference in Beijing this November. Preprint: ai.google/research/pubs/pub4…
1
4