Shridhar (@JupyterAI) header
Shridhar (@JupyterAI) profile picture

Researcher | PhD @eth_en| Past: @MSFTResearch @AIatMeta | He/him. Views are my own.

New York, NY
Joined August 2017
Photo by Shridhar (@JupyterAI)Photo by Shridhar (@JupyterAI)Photo by Shridhar (@JupyterAI)Photo by Shridhar (@JupyterAI)Photo by Shridhar (@JupyterAI)Photo by Shridhar (@JupyterAI)Photo by Shridhar (@JupyterAI)Photo by Shridhar (@JupyterAI)Photo by Shridhar (@JupyterAI)
Someone mentioned to me that @claudeai can think deeply and make plots to explain a concept. while testing an idea on bias variance tradeoff, I saw that it made a spelling mistake (flails vs fails). It was very surprising to see @AnthropicAI models making spelling mistake 🧵
1
9
744
So I checked if someone else used flail vs fail and found a blog where the author used the same word : flail Not sure if this was also intentional.
1
1
65
Flail exists as a word and it makes sense to be used in the context tbh. Now maybe human bias was in play and as AI becomes more and more intelligent, as a humans we might have to learn to adapt to AI better.
2
49
A student should confidently exploit what it knows and then explore its limits with a teacher. “Learning is often a mix of confidence and curiosity” Check out how we can balance the two for knowledge distillation at @aclmeeting in Vienna! #ACL2025
I am happy to share that our work SIKeD has been accepted to ACL 2025 @aclmeeting findings! More details below
4
428
Gave a simple image to @OpenAI O3 to look into and it zoomed in and out to make sure it’s “q” and not “9”. Is zooming in and out with corresponding text part of some alignment strategy? Or some form of augmentation that works ? Any papers in this direction ?
1
299
Thought #OpenAI's deep research would add URLs to BibTeX entries easily. Seemed like a perfect use case given I provide all the sources to look into. But NO, it decided to choose a couple of the entries and ignored all the others.
1
2
209
Game the system!
Llama 4 quietly dropped from 1417 to 1273 ELO, on par with DeepSeek v2.5
2
236
Claude 3.7 Max Thinking in Cursor is hands down the best for cloning anything 💻✨
Claude 3.7 Max Thinking is still the best model for Cursor. — Better at creating apps, refactoring and adding features — Better at figuring out connections between classes — Main benefit of Gemini 2.5 is the 1M context
1
244
A mismatch between teacher-generated rationales and the student's specific learning requirements is a huge problem in Knowledge Distillation. Solution: Teacher understanding the knowledge gap in student learning and providing tailored rationales that specifically address these weaknesses.
Is standard one-step distillation not enough for your smaller agents? 🤖 We present a new iterative distillation approach that enhances teacher🧑‍🏫-student🧑‍🎓 interaction resulting in improved student performance on math reasoning problems. 🧵
1
1
7
1,038
The approach works. Consistent improvements over standard distillation. More here: arxiv.org/abs/2504.02521
2
101
It’s crazy how the definition of small models is changing so fast. Now it’s 17B MoE with over 100B parameters. Not sure if this will help the open source community to train their own models which was the main reason why llama was so popular.
Introducing our first set of Llama 4 models! We’ve been hard at work doing a complete re-design of the Llama series. I’m so excited to share it with the world today and mark another major milestone for the Llama herd as we release the *first* open source models in the Llama 4 collection 🦙. Here are some highlights: 📌 The Llama series have been re-designed to use state of the art mixture-of-experts (MoE) architecture and natively trained with multimodality. We’re dropping Llama 4 Scout & Llama 4 Maverick, and previewing Llama 4 Behemoth. 📌 Llama 4 Scout is highest performing small model with 17B activated parameters with 16 experts. It’s crazy fast, natively multimodal, and very smart. It achieves an industry leading 10M+ token context window and can also run on a single GPU! 📌 Llama 4 Maverick is the best multimodal model in its class, beating GPT-4o and Gemini 2.0 Flash across a broad range of widely reported benchmarks, while achieving comparable results to the new DeepSeek v3 on reasoning and coding – at less than half the active parameters. It offers a best-in-class performance to cost ratio with an experimental chat version scoring ELO of 1417 on LMArena. It can also run on a single host! 📌 Previewing Llama 4 Behemoth, our most powerful model yet and among the world’s smartest LLMs. Llama 4 Behemoth outperforms GPT4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro on several STEM benchmarks. Llama 4 Behemoth is still training, and we’re excited to share more details about it even while it’s still in flight. A big thanks to all of our launch partners (full list in blog) for helping us bring Llama 4 to developers everywhere including @huggingface, @togethercompute, @SnowflakeDB, @ollama, @databricks and many others👏 This is just the start, we have more models coming and the team is really cooking – look out for Llama 4 Reasoning 😉 A few weeks ago, we celebrated Llama being downloaded over 1 billion times. Llama 4 demonstrates our long-term commitment to open source AI, the entire open source AI community, and our unwavering belief that open systems will produce the best small, mid-size and soon frontier models. Llama would be nothing without the global open source AI community & we are so ready to begin this next chapter with you. 🦙 Read more about the release here: llama.com/, and try it in our products today.
3
225
Shridhar retweeted
😅
29
132
20
8,317
317,051
What’s the equivalent of Vibe coding for AI agent? Anything specific people are testing?
2
97
ICML 2025's rebuttal process be like🤣: 👨‍💻 Authors: spend a whole week writing a careful rebuttal ✅ Reviewer: clicks "acknowledge" without reading 🚫 Author: not allowed to reply anymore So what does acknowledge mean here? "You speak. I pretend to listen. Conversation over."🙃
7
24
5
287
27,850
Sometimes I wonder what’s worse: A human lying (knowingly) or an LLM hallucinating (unknowingly). Some context: I booked an @airindia flight lately and wanted to upgrade my seats. I called and a human on the other side asked for some money to upgrade my seats.
2
1
128
I mentioned it’s the same seats. No upgrade. Apparently seats prices are dynamically set and you can pay to get the worse seats. This is what an AI system would have done too. Follow rules without any common sense applied to it.
1
58
Maybe the person had some interests on selling me those seats. He did it knowingly. An AI system would have done the same but unknowingly, following rules. Or maybe sometimes hallucinates something that doesn’t exist. I wonder what’s worse.
55