Nitin Gupta

[Nikhil Hooda @_nikhilhooda]

i’ve published to neurIPS but i still can’t understand research papers

that’s why we made arXivisual to make every paper feel like a @3blue1brown video

🏆 winner @ TartanHacks ‘26

link: https://t.co/4noZCuhxTy
repo: https://t.co/Zq4ngYCVYB

built with: @AjithBondili @_rajshah6 @armaangupt0

[Yann LeCun @ylecun]

I think the disagreement is largely one of vocabulary.
I object to the use of "general" to designate "human level" because humans are extremely specialized.

You may disagree that the human mind is specialized, but it really is. It's not just a question of theoretical power but also a question of practical efficiency.
Clearly, a properly trained human brain with an infinite supply of pens and paper is Turing complete.
But for the vast majority of computational problems, it's horribly inefficient, which makes it highly suboptimal under bounded resources (like playing a chess game).

Let me give an analogy: in theory, a 2-layer neural net can approximate any function as close as you want. In practice, almost every interesting function requires a impractically large number of units in the hidden layer. So we use multi-layer networks (that's actually the raison d'être for deep learning).

Here is another argument: the optic nerve has 1 million nerve fiber. Let's make the simplifying assumption that the signals are binary. A vision task is therefore a boolean function from 1E6 bits to 1 bit.
Among all the possible such functions, what proportion are implementable by the brain?
The answer is: an infinitesimal proportion.
The number of boolean functions of 1 million bits is 2^(2^1E6), which an unimaginably large number, about 2^(1E301030) or 10^(3x1E301029).
Now, assuming that the human brain has 1E11 neurons, and perhaps 1E14 synapses, each represented on, say, 32 bits. The total number of bits to specify the entire connectome is at most 3.2E15. This means the total number of boolean function representable (computable) by the entire human brain is at most 2^(3.2E15).
This is a teeny-tiny number compared to 2^(1E301030).
Not only are we not general, we are *ridiculously* specialized.
The space of possible function is vast.
We don't realize it because most of those functions are unfathomably complicated to and us and look completely random.
I love this quote from Albert Einstein: "the most incomprehensible thing about the world is that the world is comprehensible"
It's pretty incredible that among all the random ways the world could be organized, we can actually find a way to understand a tiny part of it.
The part we don't understand, we call entropy.
Most of the information content of the universe is entropy: things we simply cannot understand with our feeble minds and choose to ignore.

[ali @waterloo_intern]

We beat Nvidia’s cuBLAS kernels on B200s in 170 LOC.

Using zero CUDA. Just pure Mojo.

Here’s exactly how we went from 1% to 106% of Nvidia benchmark perf from scratch (with code) 👇🧵

[ali @waterloo_intern]

final results 🚀
we just blew past cuBLAS by up to 6% on Gemma 3-27B
the future of AI compute?
it's hardware-agnostic...
it's mojo 🔥

[Akshay Agrawal @akshaykagrawal]

Coming soon: bring any agent to your data in @marimo_io notebooks.

marimo is the first notebook to implement @zeddotdev's Agent Client Protocol — making it possible to use @AnthropicAI's Claude Code, @GoogleAI's Gemini, @OpenAIs Codex, or any other ACP-compatible agent.

[sunny madra @sundeep]

Groq Dammam powering Humain Chat

[Kostas Anagnostou @KostasAAA]

An Introduction to Neural Shading, Siggraph 2025 course presentation slides https://t.co/GBramFaoMI and materials https://t.co/rQGy0JDpwc

[Groq Inc @GroqInc]

OpenAI’s open models are live and already running on Groq. Try gpt-oss-20B and gpt-oss-120B today.

Groq delivers 128K context and built-in tools such as code execution and browser search. For the first time, developers and enterprises can deploy open models backed by OpenAI instantly, anywhere, at scale.

Start building now. Links in comments.

[Klaas @forgebitz]

people with setups like this always manage to ship absolutly nothing of value

it's kinda impressive

[Deedy @deedydas]

Google DeepMind just dropped this new LLM model architecture called Mixture-of-Recursions.

It gets 2x inference speed, reduced training FLOPs and ~50% reduced KV cache memory. Really interesting read.

Has potential to be a Transformers killer.

[Sebastian Aaltonen @SebAaltonen]

DX12 is 10 years old now. It's not modern anymore. Most of the complexity of DX12 and Vulkan APIs is simply not needed anymore. We don't need to support 13 year old Intel iGPUs and 12 year old mobile chips anymore. Metal 3.0 binding model is much more clean. Just GPU pointers.

[mark @marktanft]

@Cursor works well if you know how to do architectural planning, but @roocode comes with an “Architect Agent” preconfigured to help design system components.

Sometimes the best AI tools aren’t the ones that move fast, they’re the ones that make you slow down and think.

[Nitin Gupta @nitingupta910]

Ollama now supports new multimodal models with its new engine.
https://t.co/PkPUSWwGmQ

[Vulkan @VulkanAPI]

This new sample demonstrates how to render a simple triangle using Vulkan 1.3 core features. It modernizes the traditional "Hello Triangle" Vulkan sample by incorporating Dynamic Rendering, Synchronization2, Extended Dynamic State, and Vertex Buffers

https://t.co/PSQHIdYVXM

[Bryce, the CUDA Colonel @blelbach]

It's the year of CUDA Python.

[Nitin Gupta @nitingupta910]

Create USB with multiple Linux distro installers without Ventoy

https://t.co/LHAveyiAiK

[Bryce, the CUDA Colonel @blelbach]

We've announced cuTile, a tile programming model for CUDA!

It's an array-based paradigm where the compiler automates mem movement, pipelining & tensor core utilization, making GPU programming easier & more portable.

I'm proud of my stellar team for all their hard work on this!

[NVIDIA Newsroom @nvidianewsroom]

📣 Announcing #NVIDIABlackwell DGX Personal Computers. #GTC25

AI developers, researchers, data scientists and students can prototype, fine-tune and inference large models on desktops.

Learn more ➡️ https://t.co/spraN9RBpP

[Sidharth @Cloudwatch199]

Thank Brown ( Indian ) people for:

The Concept of Zero – the foundation of modern computing and digital systems

The Decimal Place-Value System – the cornerstone of global finance and advanced mathematics

Negative Number Arithmetic Rules – essential for algorithm design and higher-level computations

The Chakravala Method – a precursor to solving complex polynomial equations and cryptographic algorithms

Trigonometric Functions (Sine and Cosine) – critical to GPS, navigation, and signal processing

Infinite Series for π – paving the way for calculus and engineering simulations

Partition Function in Number Theory – fundamental to cryptography and combinatorial data analysis

The Raman Effect – revolutionizing molecular spectroscopy and materials science

Bose–Einstein Statistics – underpinning quantum computing, superconductivity, and superfluidity

The Chandrasekhar Limit – a pillar in astrophysics for understanding stellar evolution

The Saha Ionization Equation – vital for plasma physics and fusion research

The Raychaudhuri Equation – integral to black hole singularities and cosmological models

Black Hole Quasinormal Modes – key to gravitational wave research and astrophysical observations

Bhabha Scattering – advancing quantum electrodynamics and particle physics

Wireless Radio & Semiconductor Diode (J. C. Bose) – paving the way for radio broadcasting and modern electronics

Fiber Optics (Narinder S. Kapany) – at the heart of high-speed internet and data communication

The Mahalanobis Distance – used in machine learning, data mining, and pattern recognition

Dual-Phase Steel – enabling lightweight, high-strength materials for automotive and aerospace

Genetic Code & First Artificial Gene Synthesis (Har Gobind Khorana) – pioneering genetic engineering and biotech

Ribosome Structure Mapping (Venkatraman Ramakrishnan) – crucial for drug design and structural biology

ATP Mechanism & Drug Discoveries (Yellapragada Subbarow) – key to modern pharmaceuticals and chemotherapy

The Ramachandran Plot – a cornerstone in protein modeling and computational biology

Oral Rehydration Therapy (ORT) – a lifesaving intervention adopted worldwide

The Green Revolution (M. S. Swaminathan) – ensuring food security through high-yield crop strains

The Pentium Microprocessor (Vinod Dham) – critical to personal computing and modern CPU designs

Universal Serial Bus (USB) (Ajay V. Bhatt) – standardizing plug-and-play connectivity across devices

Web-based Email (Hotmail) (Sabeer Bhatia) – ushering in a new era of global communication

Digital Video & HDTV (Arun Netravali) – forming the bedrock of modern broadcasting and video streaming

MIMO Wireless Technology (Arogyaswami Paulraj) – powering 4G/5G networks and high-speed Wi-Fi

The Jaipur Foot Prosthetic – restoring mobility through a cost-effective and flexible design

Lunar Water Discovery (Chandrayaan-1) – marking a breakthrough for Moon exploration

The Mars Orbiter Mission (Mangalyaan) – setting a record for cost-effective interplanetary travel

The Aryabhata Satellite – igniting India’s journey into space research and technology

So the next time you use a number, open an email, stream a video, or marvel at a space mission, thank brown ( Indian) people.