Pranjal Srivastava

Would be fun to see Kapil Dev's iconic 175 runs innings from the 1983 World Cup come to life! #Google #Cricket

@JeffDean @petercrouch @JeffDean Didn't know you followed Football. Is it a recent phenomenon or been a while?

@petercrouch Not if you crouch?

If anyone gets hit by lightning it’s me isn’t it

Check it out here: developers.googleblog.com/google-tensor-… #GoogleIO

Happy to see my name in the Google I/O 2026 announcement! 🎉 The Google Tensor ML SDK has officially graduated to Beta — bringing a unified LiteRT workflow and 100+ models to the Pixel 10 TPU.

Agree with this view. We stand at a cusp of discovering several (beyond code) use cases of AI. Such a motto necessitates foundational research. I have a nice anecdote from Google, along the lines of what George Dantzig did in his Berkeley stat class. Mental blocks are unnecessary

@gabriberton Yeah, that was also surprising to see. It also leads to an added benefit of reduced training computation, as the heavy Encoder only has to process 25% of the input image. Overall, it is a win-win.

@gabriberton I was just reading this paper the other day. Fascinating how MAE solves the information density gap between pixels and words. In vision, pixels are redundant, hence 75% mask ratio is applied to force the model to learn semantic objects and not just interpolate features.

Thread on a staple of modern computer vision: Masked AutoEncoder (MAE) This 2021 FAIR paper proposes a new self-supervised technique to pretrain ViTs. It is one of the first ViT-specific SSL technique, which showed the world the flexibility of the transformer [1/6] 🧵

@gabriberton To give more context, the only loss function they use is the normalized MSE loss over the Masked patches.

@sethkarten Got it. Thanks for the reply. Hope you guys get sponsorship soon.

@pranjalks I think there is sufficient interest, but pending sponsorship to make it a reality. Also the potential competition will be slightly different based on current challenges in the field.

Everyone send me your pokemon ML papers. Im trying to complete the paperdex: github.com/sethkarten/awe…

@polymagelabs Looks cool. Added to my reading list. Also, great to see a deeptech startup in the domain of ML Compilers from India!

Finally, here's the paper on PolyBlocks describing how fully code-generating compilers for AI chips can be built! This is the culmination of multiple years of R&D and engineering. There is now enough reusable infrastructure in our toolkit to quickly build high-performing PyTorch/JAX compilers for new chips, no matter how weird or unique their capabilities are, and without relying on any "kernel" libraries or manual model optimization or porting. The paper isn't exhaustive, but it provides details on the key parts, the design choices, and why they are powerful. arxiv.org/abs/2603.06731

@VukRosic99 I guess you can look into Efficiency and Mechanistic Interpretability: …ter1-transformer-interp.streamlit.app

What should our LLM research goal be outside of working for OpenAI, DeepMind, DeepSeek etc? Should it be "Train GPT-5 level LLM for $100"? Does it make sense to research LLMs if we are not working at an LLM company, since we are not gonna make better LLMs than them? Should LLM research focus on helping big companies make better LLMs? Should we bail LLMs completely, work on JEPA or something else? How do we discover path forward, just doing research and seeing where it leads us?

@schopra909 This was a wonderful read. Thanks for sharing!

Turns out better VAE reconstruction quality can make your diffusion model worse. We learned this the hard way over 4 months. Writeup + open-source Image-Video VAE model. linum.ai/field-notes/va…

@o_v_shake This is a genuinely good resource! Nice domain name as well. :D

If ai summit is crowded (im not there), log on to workatafrontierlab.com so that the next time, you can get a key note speaker entry.

This is incredible. Each pixel in the camera captures only one colour (either red, green or blue), and the other 2 colours information is filled from the neighbouring pixels.

Tried creating a mock test for the GATE exam (used for engineering post-graduate admissions in India) and it worked flawlessly. This will go a long way in helping democratize exam preparation in India. #GATE #Google #Gemini

@real_deep_ml How about questions on normalization schemes like Dynamic tanh (arxiv.org/pdf/2503.10622) and Dynamic erf (arxiv.org/pdf/2512.10938)?

What paper should we make a question on next?

@willccbb Are the remote jobs only for US based folks or whether folks residing in the APAC region will also be considered?

@RisingSayak I do the exact same thing. Simple and helpful command!

I like to capture raw logs. One of the commands I overuse when running stuff from the console is: ``` <COMMAND_TO_RUN> 2>&1 | tee log.txt ``` Captures everything I need and stores it in `log.txt`. It's my default.

For those curious, this is the link to the first part of the blog series: huggingface.co/blog/Photoroom…