Paul Edward Harris 4th

More interested in what's next than what's done. . Building things, breaking assumptions. Unfiltered

[Paul Edward Harris 4th @PaulEHarris4th]

Everything I'm building, in one place:
🌀 Portfolio: https://t.co/TfesaYbcup
🧠 GitHub: https://t.co/6ADdFN3Laj
💼 LinkedIn: https://t.co/J1WtHvz9TG
🔬 ORCID: https://t.co/0BwOwcN7kV
📘 Facebook: https://t.co/rk0ypGUxQR
🐦 X: https://t.co/kZwRCQVcCa
Unfiltered. Building things, breaking assumptions.

[Paul Edward Harris 4th @PaulEHarris4th]

@otokyo__ Nien

[Paul Edward Harris 4th @PaulEHarris4th]

@solopribuilds Well do what all the Greats do. Same taken/stolen idea... just marketing is much better. Sorry but the truth. I know, I hate it too.

[Paul Edward Harris 4th @PaulEHarris4th]

@matthias_mrc Lust

[Paul Edward Harris 4th @PaulEHarris4th]

@The_Alex How are you doing this? How did one give a model l access to build such videos or games or simulations. Please tell

[Paul Edward Harris 4th @PaulEHarris4th]

@MatthewBerman How and where is this being made? Videos as well. please someone explain, what you give your model access to what platform and how? I am just confused

[Paul Edward Harris 4th @PaulEHarris4th]

My paper compares a Transformer, Mamba, an LSTM and a new architecture I built on ordinary differential equations, to find out whether architecture is the differentiator for long-horizon coherence in chaotic systems when the forecast gets fixed correction, a gated correction I introduced, or no correction at all.

"This is the groundwork for my next paper, where I'll test the same models on irregularly timed data, which is what my architecture was originally built for."

[Paul Edward Harris 4th @PaulEHarris4th]

Where my first AI research paper stands: Done: rebuilt my test twice, and found real-data correction lifts forecasts past the old floor. Next on A100/H100 GPUs: a floor-free measure, then my model vs Transformer, Mamba and LSTM, same corrections for all.

[Paul Edward Harris 4th @PaulEHarris4th]

@ValerioCapraro Like everything in life and in our reality there needs to be BALANCE. What that looks like time will tell.

[Paul Edward Harris 4th @PaulEHarris4th]

@niqudrift Should be over of yours... We'll get you there😉

[Paul Edward Harris 4th @PaulEHarris4th]

@aslambeg84 Well as a noobie I am leaning python and trying to figure out what other programming languages I should learn and there's a lot of hype about Rust. As of it's something I better learn. I also believe it and might learn it anyway. Thus, RUST

[Paul Edward Harris 4th @PaulEHarris4th]

What survives the floor fix: Feeding real data back into the forecast lifted 4 of 5 runs past the floor. No correction: 1 of 5. My learned gate: 2 of 5. Next GPU run adds a standard, floor-free measure to show exactly when each forecast fails.

[Paul Edward Harris 4th @PaulEHarris4th]

Found a second flaw in my own AI test. It could only call a forecast "failed" after filling 20 memory slots, so nothing could ever score below 380 steps. 9 of my 20 results sat exactly at 380. Caught it before spending more GPU time.

[Paul Edward Harris 4th @PaulEHarris4th]

@niqudrift Thank you.. That means a lot. I don't have anyone to talk to about this stuff and, I am kinda in a weird spot. I am not a coder but technically a vibe coder. I would say, more like someone who really knows how to get the most out of AI and it's tools if anything

[Paul Edward Harris 4th @PaulEHarris4th]

Lesson from my first research paper: check your measuring stick before you trust your results. My old test failed a forecast after one unlucky streak. The new one looks at the last 20 windows. Spoiler: I found a second flaw too.

[Paul Edward Harris 4th @PaulEHarris4th]

@Miss_G_1 Self love

[Paul Edward Harris 4th @PaulEHarris4th]

@the_codewala Your can cook your own hamburger home (Linux). Healthier, cheapest but requires some effort. Then there's McDonald's (Windows) it's everywhere, effort less not healthy and reaches for your pocket.

[Paul Edward Harris 4th @PaulEHarris4th]

@solopribuilds Using more AI

[Paul Edward Harris 4th @PaulEHarris4th]

@KomalWCode Cost, daily usage cap, and context widow limitations.