PeacefulCoder's profile picture. Google Engineer | Former Apple and Microsoft Engineer | Education Volunteer | Subpar Tweeter | HQ Trivia Winner

Kiranbir Sodhia

@PeacefulCoder

Google Engineer | Former Apple and Microsoft Engineer | Education Volunteer | Subpar Tweeter | HQ Trivia Winner

Kiranbir Sodhia reposted

what did i just read bro

amritwt's tweet image. what did i just read bro

Kiranbir Sodhia reposted

I think this is literally as close as you can get to winning a #WorldSeries without actually doing so. Unreal. That was an incredible 7-game series. Congrats to the Dodgers.

DMC_Ryan's tweet image. I think this is literally as close as you can get to winning a #WorldSeries without actually doing so. Unreal. That was an incredible 7-game series. Congrats to the Dodgers.

Kiranbir Sodhia reposted

Imagine losing first authorship because you got hit by a blue shell on the last lap 💀

luismbat's tweet image. Imagine losing first authorship because you got hit by a blue shell on the last lap 💀

LLMs are injective and invertible. In our new paper, we show that different prompts always map to different embeddings, and this property can be used to recover input tokens from individual embeddings in latent space. (1/6)

GladiaLab's tweet image. LLMs are injective and invertible.

In our new paper, we show that different prompts always map to different embeddings, and this property can be used to recover input tokens from individual embeddings in latent space.

(1/6)


Kiranbir Sodhia reposted
wintermoat's tweet image.

The trend for $GOOGL continues. Non-Search revenue is closing in on being 50% of revenue generated.

Couch_Investor's tweet image. The trend for $GOOGL continues. 

Non-Search revenue is closing in on being 50% of revenue generated.


Kiranbir Sodhia reposted

hmmm

whatRanjuSaid's tweet image. hmmm

Kiranbir Sodhia reposted

Harvard just dropped a book on ~ML systems engineering~ and it's 100% free in PDF (it will be published in print by MIT Press). systems engineering is the hot skill companies die to see in candidates, but few books explain it in the context of ML. What it covers: → ML system…

Hesamation's tweet image. Harvard just dropped a book on ~ML systems engineering~ and it's 100% free in PDF (it will be published in print by MIT Press).

systems engineering is the hot skill companies die to see in candidates, but few books explain it in the context of ML.

What it covers:
→ ML system…

Kiranbir Sodhia reposted

Excited to expand our partnership with Adobe, more deeply integrating Google's latest models in their apps

At Adobe MAX keynote I love that Google DeepMind is our great partner with @elicollins speaking on the main stage. Love Nano Banana and Veo 3.1 I use every day. Two of my favorite AI models for creativity at the moment.

icreatelife's tweet image. At Adobe MAX keynote I love that Google DeepMind is our great partner with @elicollins speaking on the main stage. 

Love Nano Banana and Veo 3.1 I use every day. Two of my favorite AI models for creativity at the moment.
icreatelife's tweet image. At Adobe MAX keynote I love that Google DeepMind is our great partner with @elicollins speaking on the main stage. 

Love Nano Banana and Veo 3.1 I use every day. Two of my favorite AI models for creativity at the moment.


Kiranbir Sodhia reposted

TPUs go brrr!

Today, we announced that we plan to expand our use of Google TPUs, securing approximately one million TPUs and more than a gigawatt of capacity in 2026.



Kiranbir Sodhia reposted

> You do Leetcode > I do LeetGPU We are not the same bro.

asmah2107's tweet image. > You do Leetcode
> I do LeetGPU

We are not the same bro.

Kiranbir Sodhia reposted

Announcing the completely reimagined vLLM TPU! In collaboration with @Google, we've launched a new high-performance TPU backend unifying @PyTorch and JAX under a single lowering path for amazing performance and flexibility. 🚀 What's New? - JAX + Pytorch: Run PyTorch models on…

vllm_project's tweet image. Announcing the completely reimagined vLLM TPU! In collaboration with @Google, we've launched a new high-performance TPU backend unifying @PyTorch and JAX under a single lowering path for amazing performance and flexibility.

🚀 What's New?
- JAX + Pytorch: Run PyTorch models on…

👀

Google officially starts selling TPUs to external customers and competes directly with Nvidia now



Kiranbir Sodhia reposted

Excited to release new repo: nanochat! (it's among the most unhinged I've written). Unlike my earlier similar repo nanoGPT which only covered pretraining, nanochat is a minimal, from scratch, full-stack training/inference pipeline of a simple ChatGPT clone in a single,…

karpathy's tweet image. Excited to release new repo: nanochat!
(it's among the most unhinged I've written).

Unlike my earlier similar repo nanoGPT which only covered pretraining, nanochat is a minimal, from scratch, full-stack training/inference pipeline of a simple ChatGPT clone in a single,…

Kiranbir Sodhia reposted

2024 evals can it count letters 🥺 can it do college stuff 🤓 are its solutions diverse 👉👈 2025 evals has it worked for 30 hours yet 🦾 has it increased gdp 📈 has it discovered novel math 🧮


Kiranbir Sodhia reposted

Quite the contrary: We're using the language that was designed as a glue language for gluing pieces together that are written in the language(s) that were designed for peak performance. Everything working exactly as designed.

its an ironic twist of fate that the most performance intensive workloads on the planet running on eye wateringly expensive hardware are run via one of the slowest programming languages with a precarious parallelism story



Kiranbir Sodhia reposted

1K GitHub stars in 2 days! 🔥 github.com/google/tunix

penstrokes75's tweet image. 1K GitHub stars in 2 days! 🔥

github.com/google/tunix

Today, we’re launching Tunix, a JAX-native LLM post-training library! With Tunix, you can do the following: - Finetuning (SFT, PEFT) - RL (GRPO, PPO, etc.) - eg.: teach your LLM to reason - Preference tuning (DPO) - Distillation, etc. For more details, refer to this blogpost:…



Kiranbir Sodhia reposted

RL research is becoming like pretraining/modeling. This is a huge vibe shift. Most research published on RL isn't using enough compute to make many of these decisions matter as much. This is slowly shifting.

practical, modern GRPO tweaks as described in Meta's Code World Models paper

iScienceLuvr's tweet image. practical, modern GRPO tweaks as described in Meta's Code World Models paper


Kiranbir Sodhia reposted

Agreed. GRPO is technically wrong.

This post is unavailable.

Kiranbir Sodhia reposted

It's delightful how easy it is to deploy working prompt injection attacks via LinkedIn

i can't believe this shit actually works

cameronmattis's tweet image. i can't believe this shit actually works
cameronmattis's tweet image. i can't believe this shit actually works


Kiranbir Sodhia reposted
thekitze's tweet image.

Kiranbir Sodhia reposted

- you are - a random CS grad with 0 clue how LLMs work - get tired of people gatekeeping with big words and tiny GPUs - decide to go full monk mode - 2 years later i can explain attention mechanisms at parties and ruin them - here’s the forbidden knowledge map - top to bottom,…


Loading...

Something went wrong.


Something went wrong.