Kunal Deo

Head of AI Customer Engineering, India, Google (Opinions = mine)

[Kunal Deo @kunaldeo]

@thdxr I hope not for the drive where you are loading models from else it will be super slow. For all the nice features BTRFS provides the same features hurt the raw read performance.

[Kunal Deo @kunaldeo]

While I agree with this. Maintainers also have to deal with a lot of AI slop so they take the easy way out. I understand that. But since it’s Open Source feel free to fork it. Heck even keeping it in sync with upstream is now easier with AI agents. So win win for everyone.

[DHH @dhh]

"This is a protectionist tale as old as time. And the justifications are just as tired: It's about quality! It's about attribution! It's about workers! Spare me. It's about you, your insecurities, and your privileges." https://t.co/SP6DubrXXh

[Kunal Deo @kunaldeo]

Glad to see this. Unmetered intelligence, with solar :), else electricity will be metered.

[NVIDIA Newsroom @nvidianewsroom]

Introducing NVIDIA DGX Station for Windows, the world's most powerful deskside AI supercomputer with Windows powered by NVIDIA GB300.

✅ Run frontier AI models with up to 1 trillion parameters locally
✅ Build and run secure AI agents on Windows with NVIDIA OpenShell
✅ Built by @ASUS, @Dell, @GIGABYTE, @HP, @msigaming, and @Supermicro

#NVIDIAGTC https://t.co/82tqvvNzJU

[Kunal Deo @kunaldeo]

Thermal issues are so prevalent that a chipset vendor is putting that as a feature.

[Snapdragon India @Snapdragon_IN]

Experience immersive graphics, smooth gameplay, and lightning-fast response, all while gaming stays cool on smartphones powered by Snapdragon | India's Most Trusted Chipset Brand*

[Kunal Deo @kunaldeo]

@kimmonismus Anything other than Factorio is the incorrect answer.

[Kunal Deo @kunaldeo]

@popovicu94 You would need to disable secure boot

[Kunal Deo @kunaldeo]

Claude Code and Codex can be friends they just need a bridge like Tmux. They don’t care about the feud between their leaders :)

[Kunal Deo @kunaldeo]

I need to make the app snappy. I wonder if an MCP server would be handy here.

[Kunal Deo @kunaldeo]

I see people running hordes of agents to build stuff quickly thinking they are the bottleneck. I don’t like that. I like the process of debating with the agent for each feature I am going to ship, going through the changelog and modding things by hand where it makes sense.

[Kunal Deo @kunaldeo]

Great way to learn foundational ai

[Dwarkesh Patel @dwarkesh_sp]

New blackboard lecture w @ericjang11

He walks through how to build AlphaGo from scratch, but with modern AI tools.

Sometimes you understand the future better by stepping backward. AlphaGo is still the cleanest worked example of the primitives of intelligence: search, learning from experience, and self-play. You have to go back to 2017 to get insight into how the more general AIs of the future might learn.

Once he explained how AlphaGo works, it gave us the context to have a discussion about how RL works in LLMs and how it could work better – naive policy gradient RL has to figure out which of the 100k+ tokens in your trajectory actually got you the right answer, while AlphaGo’s MCTS suggests a strictly better action every single move, giving you a training target that sidesteps the credit assignment problem. The way humans learn is surely closer to the second.

Eric also kickstarted an Autoresearch loop on his project. And it was very interesting to discuss which parts of AI research LLMs can already automate pretty well (implementing and running experiments, optimizing hyperparameters) and which they still struggle with (choosing the right question to investigate next, escaping research dead ends). Informative to all the recent discussion about when we should expect an intelligence explosion, and what it would look like from the inside.

Timestamps:

0:00:00 – Basics of Go
0:08:06 – Monte Carlo Tree Search
0:31:53 – What the neural network does
1:00:22 – Self-play
1:25:27 – Alternative RL approaches
1:45:36 – Why doesn’t MCTS work for LLMs
2:00:58 – Off-policy training
2:11:51 – RL is even more information inefficient than you thought
2:22:05 – Automated AI researchers

[Kunal Deo @kunaldeo]

@pritopian These are FTEs

[Kunal Deo @kunaldeo]

@sanjeed_i Those who are interested for India DM me with your GitHub profile/protfolio

[Kunal Deo @kunaldeo]

Couldn’t agree more :)

[François Chollet @fchollet]

When looking at deep learning profiles, one of the most obvious tells between a mediocre and great candidate is whether they list PyTorch or JAX.

[Kunal Deo @kunaldeo]

Dwarkesh is asking excellent questions and Jensen is giving good answers. This is what you would want in an interview. Not hey you are so great, yes we are so great.

[Dwarkesh Patel @dwarkesh_sp]

The Jensen Huang episode.

0:00:00 – Is Nvidia’s biggest moat its grip on scarce supply chains?
0:16:25 – Will TPUs break Nvidia’s hold on AI compute?
0:41:06 – Why doesn’t Nvidia become a hyperscaler?
0:57:36 – Should we be selling AI chips to China?
1:35:06 – Why doesn’t Nvidia make multiple different chip architectures?

Look up Dwarkesh Podcast on YouTube, Apple Podcasts, Spotify, etc. Enjoy!

[Kunal Deo @kunaldeo]

sglang-kernel compile can eat into whatever you throw at it. Here it is easily chewing through 1 TB of RAM and 224 cores of CPU with CMAKE_BUILD_PARALLEL_LEVEL set to 100

[Kunal Deo @kunaldeo]

@LLMJunky @centralcomputer I think even after bandwidth loss it is a better card as can do fp8 and nvfp4 these more than makeup for bandwidth loss and allows you load up to 2x (fp8) or 4x (nvfp4) large models

[Kunal Deo @kunaldeo]

@FedericoNoemie Welcome to the club

[Kunal Deo @kunaldeo]

@simonw @osanseviero You can use https://t.co/5CZruE99Mh or use LiteRT to build your own anywhere

[Kunal Deo @kunaldeo]

The latest [email protected] now pulls in [email protected], a package that did not exist before today. This is a live compromise. If you are worried about python packages make sure you check out. https://t.co/clBj3ee9WJ

[Kunal Deo @kunaldeo]

@karpathy's post about LiteLLM supply chain attack post genuinely scared me. A routine pip install being able to exfiltrate secrets and CI/CD creds was enough to make me build secure-packages.

It’s a supply chain security gate for packages, starting with PyPI:
- point it at requirements.txt
- fetch package source
- review it
- cache approved versions by hash
- diff new releases against the last approved version
- block risky updates in CI/CD

https://t.co/omcmSZJ2iD

[Kunal Deo @kunaldeo]

@simonw I posted about this project and this happened https://t.co/clBj3ee9WJ

[Kunal Deo @kunaldeo]

@karpathy's post about LiteLLM supply chain attack post genuinely scared me. A routine pip install being able to exfiltrate secrets and CI/CD creds was enough to make me build secure-packages.

It’s a supply chain security gate for packages, starting with PyPI:
- point it at requirements.txt
- fetch package source
- review it
- cache approved versions by hash
- diff new releases against the last approved version
- block risky updates in CI/CD

https://t.co/omcmSZJ2iD