Tarun @atemyipod

eternal founding MLE | prev upenn, msft research, quant | opinions here are my own San Francisco, CA Joined March 2016

Tweets

6K
Followers

793
Following

2K
Likes

13K

Sachin Goyal @goyalsachin007

11 hours ago

1/Excited to share the first in a series of my research updates on LLM pretraining🚀. Our new work shows *distilled pretraining*—increasingly used to train deployable models—has trade-offs: ✅ Boosts test-time scaling ⚠️ Weakens in-context learning ✨ Needs tailored data curation

3 32 142 10K 109

Download Image

Divyat Mahajan @divyat09

4 hours ago

Really interesting findings!

Sachin Goyal @goyalsachin007

11 hours ago

Really interesting findings!

3 32 142 10K 109

Download Image

0 1 1 269 0

roon @tszzl

a day ago

latent space is best explored in person

68 34 687 50K 53

Tarun @atemyipod

19 hours ago

never deleting this app btw 😭 😭 😭

0 0 1 108 0

Download Image

roon @tszzl

a day ago

i can’t believe they recreated harvard university from the famous movie the social network

103 75 5K 165K 118

Download Image

sankalp @dejavucoder

a day ago

sydney sweeney reveals in an interview that her current favourite model is gpt-5. "claude code's been an enjoyable experience. i have loved sonnet/opus 4 for coding but lately anthropic is fucking up and sonnet 4 is pushing too much slop. gpt-5 in cursor/codex been a saviour"

43 47 1K 95K 130

Download Image

Decart @DecartAI

4 days ago

Introducing Oasis 2.0: Our most advanced AI model Transform game worlds & styles in real time. 1080p, 30fps Here’s how to play it live (and also a Minecraft Mod)! 🧵

103 196 2K 250K 1K

Download Video

Long Le @LongLeRobot

2 weeks ago

Thanks AK for sharing our work! 😊

AK @_akhaliq

2 weeks ago

Thanks AK for sharing our work! 😊

5 30 177 21K 85

Download Video

0 1 14 1K 3

Tarun @atemyipod

2 weeks ago

Dot product is essentially a 1-head Transformer without softmax: y = (xW_Q)(wW_K)^\top where the router is hard-coded and the only “expert” is the weighted similarity. And then you only need to train the projection weights and the attention just follows along.

Shuangfei Zhai @zhaisf

2 weeks ago

14 21 344 38K 255

0 0 1 142 0

Shuangfei Zhai @zhaisf

2 weeks ago

ReLU is essentially an element wise MoE with shared router and expert weights: y= I(w^Tx >= 0^Tx) w^Tx + I(w^Tx < 0^Tx) 0^Tx, where the two experts are parameterized by the w and 0 (zero) vectors. And then you only need to train the expert weights and router just follows along.

14 21 344 38K 255

Lakshya A Agrawal @LakshyAAAgrawal

2 weeks ago

youtube.com/watch?v=H4o7h6… Thanks @CShorten30 for creating this amazing tutorial on using GEPA to optimize a Listwise Reranker!

2 12 79 9K 65

Kaustubh Sridhar @_k_sridhar

2 weeks ago

FYI, if you’re trying out RICL-pi0 and need tech support, ping me here or on email (or raise a GitHub issue) and I will help fix any issues you run into.

Kaustubh Sridhar @_k_sridhar

3 weeks ago

FYI, if you’re trying out RICL-pi0 and need tech support, ping me here or on email (or raise a GitHub issue) and I will help fix any issues you run into.

6 29 225 46K 172

Download Video

0 1 5 526 2

Kaustubh Sridhar @_k_sridhar

3 weeks ago

Robot AI brains, aka Vision-Language-Action models, cannot adapt to new tasks as easily as LLMs like Gemini, ChatGPT, or Grok. LLMs can adapt quickly with their in-context learning (ICL) capabilities. But can we inject ICL abilities into a pre-trained VLA like pi0? Yes!…

6 29 225 46K 172

Download Video

Mohak Mangal @mohakmangal

3 weeks ago

We came second in the GPT-5 Hackathon. We are building ML for the electric grid. The grid will define the next 100 years. We want to redefine the grid. We are looking for a founding MLE and a founding engineer. Talk soon :)

159 221 7K 660K 1K

Download Image

Rohit @rohitrango

3 weeks ago

🧠 Excited to share our DIO work published in Medical Image Analysis! 🚀 Traditional deep learning methods for medical image registration sacrifice optimization robustness for speed and very flexible inference-time customization. DIO bridges this gap with implicit optimization.

1 3 6 332 0

Download Video

Dhruv π @dhruv31415

3 weeks ago

For a second I thought someone had dropped the most cursed mixed precision strategy

tobi lutke @tobi

4 weeks ago

For a second I thought someone had dropped the most cursed mixed precision strategy

23 8 175 74K 14

7 16 409 32K 19

Rohit @rohitrango

3 weeks ago

are there any massive 3d datasets that we wanna try this on?

Rohit @rohitrango

4 weeks ago

are there any massive 3d datasets that we wanna try this on?

0 2 7 511 2

1 1 2 352 0

Lakshya A Agrawal @LakshyAAAgrawal

3 weeks ago

Very excited to share that GEPA is now live on @DSPyOSS as dspy.GEPA! This is an early code release. We’re looking forward to community feedback, especially about any practical challenges in switching optimizers.

9 39 323 43K 146

Download Image

Rohit @rohitrango

4 weeks ago

3D convnets are powerful but very memory intensive - a lot of representations like high fidelity occupancy maps, 3d image volumes, climate data, cosmological data cannot be processed at scale because of memory requirements. (1/2)

1 2 8 595 3

Rohit @rohitrango

4 weeks ago

now that dinov3 is released, cannot wait for the same braindead drill in medical imaging: - someone benchmarks it on mri and compares it to nnunet - someone does a CT comparison - someone claims SOTA registration for the 3312th time (but still unsolved)