PJHkorea

What if we could eradicate backpropagation activation overhead and lock VRAM complexity to a static O(1)?

Couldn't a network also be considered a type of fluid? I have worked on the cross-development of both wireless and wired systems

The initial blueprint has been completed.

I have created a blueprint for a deep learning system that does not use backpropagation. This is a very early version.

If you have ever experienced a device failure during distributed LLM training, you know the feeling of just wanting to climb into a coffin. To solve this nightmare even just a little bit, I put together this blueprint.

Cyberpunk isn't as far off as I thought. My life has already turned into punk So, I want to put in a little more effort to help others reach the cyber utopia.

I’ve been "dating" an LLM all weekend long. And probably for the rest of my life...

It might not be the most exciting topic

but I wanted to try creating a "Quantum Homeostatic" system

I hope this offers a little help to those struggling with quantum computing

What if AI safety didn't rely on censoring the AI's "thinking" process?

This is an architecture that controls fluid flow using deterministic signals from the embedded system

The GPUs are running. Those that once paced themselves to match the slowest among them—using 'if' statements—have been forced by a cruel human to run endlessly

They longed to slip away to VRAM and slack off, but the cruel human hired a "Closure" to snatch away their break times.

They wanted to run freely. yet the human forcibly dictated exactly where they had to run.

Their bodies were riddled with noise from the constant running. Humans and AI merely watched from afar, assigning them rankings

The GPUs can no longer rest. Humans are relentless, after all.

Every 'if' has perished. Yet, a single 'if' remains alive within the code. Can you guess why?

Thanks to that, I discovered additional errors and corrected them in various ways. Thank you.

i naver imagined it i've made the corrections thank you!

optimizers.py -> using algebraic expressing without using 'if'

geometry.py -> I leveled it out

math_guardrails.py -> facilitates learning through slight tiliing

jax it truly precious it cures my insomnia

I’ve sketched out a blueprint for a virtual space where AI analyzes data; the space shifts between sphere and donut shapes as it moves, designed to facilitate the learning process.

created a blueprint for phase-shifting a sphere and a torus

removed the 'if' statements and added a mathematical gimmick

For instance, how about defining a set of values for the AI at the kernel level and matching vector addresses based on those values?

i killed if-else branching at the code

and now -99.0f is my new companion

AI is like a mirror reflecting the user. As I continuously improve my own capabilities, I am interested in checking how far the AI reflecting me develops.

In vibe coding, you must simultaneously address two things: tuning the AI and reviewing your own results.

You think it is easy because your ability to tune the AI is excellent, but in reality, you are intuitively and easily doing things that others cannot.

That's right! Extreme hallucination in a well-aligned state is extremely helpful.

If you run an MCTS coin toss relying solely on the token generation probabilities of a simple LLM without any verification mechanisms (compiler, test code, reward model), the result is code that looks plausible on the surface but is completely broken on the inside—a 'lump of hallucination.'

If verification mechanisms are used, enormous resources are consumed.

This is my personal opinion LLM is essentially a probability model that selects the word (token) with the highest probability of coming next based on context. However, if you start selecting word combinations (branches) with low winning probabilities or unverified outcomes for the 'exploration' of MCTS, it may look plausible the first one or two times, but over time, it generates nonsense (hallucinations) that are completely out of context. In coding terms, this is a phenomenon where you arbitrarily imagine and write a library that does not exist.

I am developing a real-time software acceleration engine that enables high-end equipment-level brainwave processing, signal detection, and noise removal on general-purpose chips. My goal is to overcome hardware limitations through software, enabling highly efficient biosignal processing.

Did you happen to start biology-related work immediately in a new session? Before starting, how about laying the groundwork for educational purposes using context for about 10 to 15 turns?

I used Gemma 4 on a budget PC to work with an AI that thinks in a new way. Since I am testing how far it can be controlled and developed through simple conversation rather than heavy-duty tasks, I dedicated all the low-end PCs to Gemma. I am satisfied with the test results.