Romin

[Romin @RominPlusFans]

@wokecisally booo

[The Vigilant Fox 🦊 @VigilantFox]

Neil deGrasse Tyson: “Whatever [the next outbreak] is, we ain’t ready for it. We still have anti-vaxxers running around.”

“I don’t trust scientists. I saw a YouTube video, so I’m not going to take it.” (mocking)

“I don’t want you to ever forget this story.”

“20,000 years ago, we’re in the cave. Do you know what the life expectancy was?”

Shannon Sharpe: “10 years? 15 years?”

Neil deGrasse Tyson: “30. Half of everyone born was dead before they were 30.”

Shannon Sharpe: “Wow!!!”

Neil deGrasse Tyson: “Fast forward to 1840… everyone born in the world was dead by the age of 35. We gained five years of life expectancy. And every one of them ate organic, breathed clean air… Science matters here.”

“We’ve doubled the life expectancy with antibiotics, vaccines, and sanitation. The three biggest forces operating on our longevity. So to come around and say I don’t need vaccines because I’m not getting sick, that’s like saying, why are you using dandruff shampoo? You don’t have dandruff.”

Shannon Sharpe: “Well, I don’t want to get it.”

Neil deGrasse Tyson: “That’s my point. If you’re successful, people think you don’t need it when that’s what’s creating the ongoing success in the first place.”

[Romin @RominPlusFans]

@lindaxie Have you been doing virtual experiments, or at home labs?

If neither, here is a free resource for some experiments: https://t.co/hiyKKoTibe

[Romin @RominPlusFans]

CEOs of companies should also be held accountable when they're aware their LLMs produced images of child pornography. Time is ticking Elon.

[Elon Musk @elonmusk]

This needs to change

[Science girl @sciencegirl]

The Olympic symbol can be drawn in one stroke.

[Romin @RominPlusFans]

@keysmashbandit A pretty fun idea. I did it with the normie gemini "deep research". I have never used it for brain stuff. I guess there is a deep connection in the training data about robots being interested in their consciousness? It sounds sci-fyish so maybe there is a lot of literature on it?

[Romin @RominPlusFans]

This is why education does not correlate with intelligence. It's also why we need to lower healthcare prices in the USA, so idiots like this go into different fields. Disgusting that someone like him can touch your body with money being the only thing on his mind.

[Romin @RominPlusFans]

Tl;dr: "I'm an insecure code monkey who happens to do physics, and I can't understand basic quantization ideals so I'm going to pretend I can understand your fields because I can't understand my own :("

[Mayukh @mayukh_panja]

There is an objective hierarchy among academic fields and therefore not all PhDs are equal. Before you get upset, hear me out.

I am a computational physicist, I solve partial differential equations (PDEs) on a computer to simulate real world phenomena. But if I were to try to understand string theory, it would take me a full year before I can even scratch the surface. Because I don't have the mathematical rigor for it. For a string theorist though, it wouldn't be too difficult to understand what I do.

Similarly, I could pick up a psychology paper and understand most of it in a first read. But obviously the reverse is not true. It would be impossible for a psychologist to understand differential equations without any prior exposure to physics or maths.

Humanities papers require almost no specialised training. Anybody with common sense and some English language comprehension skills can read, or even write a humanities paper.

In fact, read about the Sokal affair. A professor of physics (Alan Sokal) wanted to test the intellectual rigor of a cultural studies journal. He produced some garbage that sounded good and flattered the preconceived notions of the journal editors and voila the paper was accepted. It became a big scandal back in the day.
By and large, humanities papers are not very intellectually rigorous or demanding. Hierarchies exist almost in every realm of human endeavour. Not all sports are equally physically demanding. We readily accept that. Let's stop pretending everyone is equal. It is doing more harm than good.

Before you come at me with pitchforks, I am talking specifically about academic research. I have deep respect for authors, journalists, musicians, artists, anybody doing anything creative and original.

[Romin @RominPlusFans]

A small reminder that every field has crackpots, and language models will only make it worse.

[vixhaℓ @TheVixhal]

The Future of Language Modeling is a Wave, Not a Particle

The AI world has been obsessed with one fundamental architecture, the Transformer. It powers everything from ChatGPT to Grok.

It is, without a doubt, a brilliant piece of engineering.

But what if the fundamental way Transformers look at language is flawed?

To understand the future of Large Language Models, we can borrow an analogy from physics. It is the difference between a particle and a wave.

The standard Transformer architecture treats language as a collection of discrete particles.

Every word, or token, is an individual marble in a bag. To understand a sentence, the model uses a mechanism called self-attention.

This forces every single token to interact with every other token so that the model can figure out their relationships.

If you have a sentence with 100 words, that results in 10,000 interactions.

This is the infamous O(N²) quadratic bottleneck. It works, but it is brutally inefficient.

As we demand longer context windows, the particle approach requires exponentially more compute.

What if we stop treating language as discrete particles and start treating it as a continuous wave?

Human thought does not happen in isolated bursts. It flows. A sentence begins, rises in tension, and resolves. It has rhythm and frequency.

Architectures such as S4, Hyena, and FNet are abandoning the pairwise handshake mechanism. Instead, they use signal processing tools such as the Fast Fourier Transform.

Here is the shift:

The Input:
Instead of seeing 1,000 individual marbles, the model sees one continuous 1,000-unit long signal.

The Processing:
Instead of pairwise comparisons, the model moves the signal into the frequency domain and applies a filter to the entire wave at once.

The Result:
The math changes from quadratic O(N²) to near-linear O(N log N).

Google’s FNet showed that it can reach 92 percent of BERT’s performance while being 80 percent faster. There is clearly something promising here.

If waves are so much better, why are they not used everywhere?

Two major problems have prevented FFT-based models from overtaking Transformers.

1. The Echo Problem: Waves capture global patterns well, but they struggle with precise details. Language needs both.

2. Exploding Noise: In signal processing, waves can resonate endlessly if not controlled.

What I Did to Fix This

I have been working on a stabilized spectral architecture that uses physics-inspired constraints.

The key breakthrough was adding a control mechanism that acts like gravity for the wave function.

The solution has three main components.

1. A learnable decay mechanism that forces the wave to settle over time

2. A dual-stream design where one path processes the signal and the other path decides what information to retain

3. Complex-valued kernels that preserve phase information so that the model maintains causality

This solves the echo problem. By forcing the wave to settle, the model gains short-term memory for precise details while keeping long-range context through the spectral pathway.

It also stabilizes the math. The decay mechanism acts like a control rod in a reactor and prevents gradients from exploding.

This allows deep spectral networks to train without the instability that affected earlier attempts.

I will publish the full code on GitHub in two days. I am still running the final scaling experiments.

It is time to stop counting particles and start riding the wave.

[Romin @RominPlusFans]

BREAKING NEWS: A man who has relies on libraries and blackboxes built by others fools himself into thinking he has the intelligence to make logical systems himself. More at 11!

[Mustafa. @oprydai]

if i had to revise mathematics for the modern age, i'd rebuild it for engineers, not for exam rooms.

start from operations, not abstractions.
start from systems, not symbols.
start from behavior, not proofs.

the new sequence:
> numerical intuition
numbers as compressed descriptions of reality.
scaling.
orders of magnitude.
approximation over exactness.

> functions as machines
inputs → transformations → outputs.
no “f(x)” worship.
treat functions as modular components in a pipeline.

> calculus as dynamics
derivatives as sensitivity.
integrals as accumulation.
everything rooted in motion, feedback, energy, and change.

> linear algebra as geometry
matrices as transformations of space.
vectors as directions of influence.
eigenstuff as system behavior across time.

> probability as uncertainty engineering
quantifying doubt.
updating belief.
distributions as models of the unknown.

> optimization as intelligence
search. gradients. trade-offs. constraints.
the core engine behind ML, control, planning.

> differential equations as world models
the language of physics, robotics, fluids, everything.
treat them as simulation engines, not exam torture.

> discrete math as architecture
graphs, combinatorics, automata.
the backbone of algorithms, networks, compilers.

> numerical methods as reality filters
how computers actually compute math.
floating point. solvers. errors. stability.

> mathematics as compression
the entire field reframed as pattern extraction and reduction of complexity.

this version of mathematics isn’t for classrooms.
it’s for builders.
for people designing robots, systems, simulations, models, infrastructure.

mathematics as a toolkit.
not a religion.

[Romin @RominPlusFans]

@getjonwithit

[Romin @RominPlusFans]

@mochi_byte0 Soulcial media would never do numbers

[Romin @RominPlusFans]

@IamSreeman Write the book then

[Nirmalya Kajuri @Kaju_Nut]

@Gn_O_S

[Romin @RominPlusFans]

@mochi_byte0 Your kiwi post is a counterexample to the assertion that you're here to help 👀

[Romin @RominPlusFans]

@izzitrin Do we really need it? A question above my pay grade... congratulations!

[Michael "BWC" Clayton (2007) @Michael_CIayton]

Overrated cities:

-Charleston
-New Orleans
-Nashville
-Dallas
-Charlotte

Underrated cities:

-New York
-Chicago
-LA
-Miami
-Boston

[SXS Project @SXSProject]

We are excited to release a major update to our catalog of binary black hole simulations, available at https://t.co/LoIQoEUkN0! Such simulations are key to LIGO/Virgo/KAGRA being able to extract science from their gravitational wave detections.

1/14