Samir Char

- Data Scientist @Microsoft
- MS in Data Science @Columbia University
- Runner, weightlifter
- I love reading books

[Samir Char @Samir_char]

@cdomingoenrich @randall_balestr In the embedding geometry, scaling the text encoder improves uniformity but worsens image-text alignment. Modality-specific weight decay recovers alignment while preserving uniformity.
📄 https://t.co/ubQ4ycaTdp
💻 https://t.co/p4hOS7XSa0
🤗 https://t.co/sEj2iQdF5W

[Samir Char @Samir_char]

Bigger CLIP isn't always better. 🧵

The vision↔text split is usually an unexamined design choice. We trained 30 CLIP models, scaling each encoder independently, and evaluated downstream performance on 40 datasets.

with @cdomingoenrich and @randall_balestr

[Samir Char @Samir_char]

@cdomingoenrich @randall_balestr We trace the degradation to modality-specific overfitting: past a point, growing the text encoder lowers training loss but hurts zero-shot. The fix: modality-specific weight decay. More decay on the text than the vision recovers every degraded config; the reverse hurts.

[Kevin K. Yang 楊凱筌 @KevinKaichuang]

When training protein language models, people usually discard metagenomic proteins that don't cluster with at least 1 other protein (singletons).

With @riavinod_ @Samir_char @avapamini @lorin_crawford , we show that this is probably not the right strategy.

[Samir Char @Samir_char]

@NandoDF So is the proposal RL right from the start? Or which learning paradigm?

[Samir Char @Samir_char]

@mckaywrigley Even if the next models "solve" software, does anyone know how users and companies will support all the massive inference cost required for a new and sustainable paradigm? Would profits be enough to dedicate full data centers to LLMs instead of other use cases (e.g., cloud, ads)?

[Samir Char @Samir_char]

@tunguz What skills do you think will be key in the next generation and how to build them? Analytical and scientific reasoning?

[Samir Char @Samir_char]

@KevinKaichuang @alexijielu @avapamini @lorin_crawford Having worked with the team, I can't recommend this opportunity highly enough!

[Kevin K. Yang 楊凱筌 @KevinKaichuang]

You have until Dec 1 to apply to the bioml PhD research internship!

This is where you apply to work with me, @alexijielu @avapamini @lorin_crawford or Kristen Severson!

(new) link and some instructions below

[Randall Balestriero @randall_balestr]

At Chicago Booth to speak about our recent JEPA research! We need more researchers from all fields to join our efforts and to keep pushing for better JEPAs, LeJEPA is only the V0.

[Ava Amini @avapamini]

single-cell models tend to learn from the many - and miss the rare

we introduce an Adaptive Resampling approach to help models learn from underrepresented cells, improving generalization & discovery

https://t.co/CIpAW1RwWA
https://t.co/xfCc13dMe5

great work by @NavidiZeinab!

[Samir Char @Samir_char]

@KevinKaichuang Also waiting for the name...

[Kevin K. Yang 楊凱筌 @KevinKaichuang]

Are you a PhD student interested in machine learning and biology or health? Come do an internship with me, @avapamini, @alexijielu, @lorin_crawford, or Kristen Severson at @msrne!

Applications are due Dec 1: make sure you include a research statement!

https://t.co/HfyZshDtof

[Gary Marcus @GaryMarcus]

The last few months have been devastating for LLM dreams:

• Apple reasoning paper and the ASU mirage paper and many others confirmed that LLMs still can’t solve distribution shift.

• GPT-5 came late and fell short.

• Turing Award winner Rich Sutton thanked me for my critiques of LLMs and agreed.

• Karpathy just said agents aren’t anywhere close, and that AGI is a decade away.

• And Hassabis just blew up some wildly overhyped claims from OpenAI about math.

Game over, man.

LLMs have their place, but anyone expecting the current paradigm to be close to AGI is delusional.

[Pedro Domingos @pmddomingos]

Andrej Karpathy explains why reinforcement learning sucks and is not the path to AGI:

[Samir Char @Samir_char]

@karpathy @karpathy It would be amazing to learn how you manage the learning vs shpping tradeoff. How you balance coding from scratch and without looking at other code vs just borrowing implementions from others. It will help others create great repos like yours!

[Ava Amini @avapamini]

Applications for @MSFTResearch undergrad research internships for rising juniors and seniors are due Monday Oct 6!

Apply to work with us in BioML 👉 https://t.co/QYjsdyEpRQ

w/ @KevinKaichuang, @alexijielu, @lorin_crawford, Kristen Severson, @ntenenz, @SarahAlamdari, and more!

[Samir Char @Samir_char]

@iScienceLuvr @KevinKaichuang

[Samir Char @Samir_char]

@randall_balestr This is exciting!