Davinder C. Singh (@davindercsingh) header
Davinder C. Singh (@davindercsingh) profile picture

LP investor (PE/VC/RE). Bio AI, Drones, Logistics, Infrastructure, and Cybersec. Graduated Imperial College London Maths. Tech Founder. Per Aspera Ad Astra.

Palo Alto, CA
Joined April 2014
Photo by Davinder C. Singh (@davindercsingh)Photo by Davinder C. Singh (@davindercsingh)Photo by Davinder C. Singh (@davindercsingh)Photo by Davinder C. Singh (@davindercsingh)Photo by Davinder C. Singh (@davindercsingh)Photo by Davinder C. Singh (@davindercsingh)Photo by Davinder C. Singh (@davindercsingh)
Entrepreneurs are those beings in any given age who see that human progress is simply not fast enough...
2
4
3
37
Davinder C. Singh retweeted
Bye bye looped reasoning… and hello Energy-Based Reasoning⚡! Introducing Energy-Based Reasoners TLDR: - EBRs use 345x less inference compute and 2.6x fewer training FLOPs than the best baselines - EBRs are 7x-43x more sample-efficient than the best looped baselines at inference - EBRs achieve 99.9% on Sudoku-Extreme and near 100% accuracy on multi-solution benchmarks 🧵Thread
88
258
33
2,766
161,715
Is this what I'm doing?
17
R8 (not possible via any published method except on a supercomputer - enclosed late September 2026). M1 (not possible via any known method on any known hardware - soon to be enclosed). Both on a laptop.
3
89
The Analogy: Untangling a Massive Knot ​Imagine you are trying to map a colossal, tangled ball of yarn that represents a chaotic fluid flow. ​The Outer Strands (The Bounded Tail): On the outside of the ball, the strings straighten out and run parallel. You don't need to trace every millimeter of those outer strings because you can use a simple mathematical rule (a "fence" or bound) to predict where they are going. ​The Dense Core: Right in the center is a hyper-tangled, messy knot where every single thread crosses over, under, and interacts with every other thread simultaneously. Because it is so intensely interconnected, you can't use a shortcut or a general rule. You have to untangle and map that exact center piece-by-piece, connection-by-connection. ​That messy, highly interactive center is the dense core.
26
That 2.35EB matrix at full size is half billion columns and rows for one orbit... it really is a monster.
15
Davinder C. Singh retweeted
On the same day Jensen Huang lost Nvidia, a researcher published something that explains his whole story better than anything actually written about him. Will DePue found that almost every AI model carries a hidden fingerprint of the chip it was born on. Before training begins, a model is nothing but billions of starting numbers. How those numbers get laid out depends on the exact hardware. Then the model trains for months, on trillions of words, and every one of those numbers gets rewritten. The pattern is still there. He took models released by China’s top labs and read, in the weights themselves, the Nvidia Blackwell chips they were started on. Chips those labs are not supposed to have. Months of training could not wash out where they began. Nothing you build ever fully forgets where it came from. Jensen Huang washed dishes at a Denny’s as a teenager. In 1993 he started Nvidia in a booth at one. He had nothing, and he bet everything on a market that did not exist yet. He bet on CUDA when Wall Street called it a waste. He bet on AI a decade before the rest of the industry understood what a neural network was. He bought Mellanox for $7 billion because he saw, before anyone else, that the future was not one fast chip but thousands of them working as one machine. He did not manage Nvidia. He imprinted on it. Roughly 60 people reported directly to him. He knew the company down to the pay of its most junior employee. There was no successor because there was no version of the company that existed without his hand on every layer of it. Every major AI model you have ever used was trained on hardware that exists because of decisions he made when nobody agreed with him. He was its starting values. Then came the affairs, the lies about them, and a board meeting that ran late into the night. According to Bloomberg, the man famous for caring about everything started saying “whatever” and “I don’t care.” He called himself a piece of sh-t. That is not a man who does not care. That is a man watching thirty years of his life collapse and knowing he is the reason. Since then, more than 1,600 comments from inside the company have asked the board to bring him back. Months of scandal could not wash out the pattern there either. The fingerprint does not care whether you are proud of it. It keeps the brilliance and it keeps the failure. It keeps CUDA and it keeps the lies. It keeps the $7 billion bet that built the AI era and the moment he convinced himself nobody would ever look closely. That is what makes it real. A myth would only keep the good parts. A fingerprint keeps everything. And the people who worked beside him for thirty years are not asking for the myth. They are asking for the man, because they know the difference between someone who is perfect and someone who built something no one else could have built. We like to believe that enough success can overwrite who we were. That enough time can overwrite what we did. It cannot. It only buries it deeper in the weights. Jensen Huang’s fingerprint is on every chip, every architecture, every model that carries the AI era forward. It will be there long after the scandal fades and long after the man himself is gone. Everything you are is still sitting in the first decisions you made. Somebody, eventually, learns to read it.
20
67
4
227
17,757
Davinder C. Singh retweeted
This is how Jensen lost it.
667
2,346
534
24,555
3,349,826
Data centre cooling (low Reynolds / strain solved) is a massive upo opportunity we are exploiting right now.
1
1
108
I know a PhD spent 3 years on a data centre cooling control software. Our UPO work already outperform them. Demonstrating a new maths technique applicable across industries.
1
35
Dropped from O(Re^{9/4}) down to O(Re^{15/8}) - below quadratic.
2
85
This means a 178-fold surge per magnitude increase in Re to 75-fold (10^15/8 = 74.99) surge in days on the critical path. 1 week I am sure we'll get this below 10-fold (linear stepping).
19
Building a software framework that decides the stability and enclosure of chaotic Navier–Stokes flows proves that the most intractable physical systems yield not to heavier hardware, but to sharper foundational mathematics. That's our exact investment thesis and the more proof points the better the valuation.
55