John Holder

John Holder. Independent math researcher working on Kalman filters and estimation error. 11+ yrs I.T. exp. Building Retroment Gaming in Florida.

[John Holder @RedWingedDuck]

Hi, I am John. Software dev and indie game studio founder, researching how to perfect the operating envelope by nailing down the exact math. I post about math, classic games, and developing my own.

[John Holder @RedWingedDuck]

@sylviapuffs Lol, you're not wrong, it's still like that.. my question is almost a month old with 1 like... I don't think I'm getting an answer

[John Holder @RedWingedDuck]

@ClintonDesveaux The yang-mills paper isn't out fully yet but the whole Lean database and equasion is available.. I am interested in their words though!

The hodge conjecture hasn't been solved yet as far as I could find?.. just the same Nov6 information posted today

[John Holder @RedWingedDuck]

I dunno. Giving the same question to chatgpt/Gemini gives me the sicophant response, Grok and Claude push back on me hard.

Gemini-you are genius, here is related information
Chatgpt-you are genius, here is how I would take it a step further
Claude-are you dumb? The variable doesn't go there
Grok-youre wrong, I'm also wrong

[John Holder @RedWingedDuck]

@bcdsignature We learn from history... The printing press, automated cargo bays, the stock market... You cannot stop progress, we need to mediate the transition the best way possible .. if AGI can bring prosperity to all, then let's starts with the displaced workers

[John Holder @RedWingedDuck]

Nice!.. I got pretty far with testing galaxy rotation but there is a missing piece!.. maybe getting close to not needing my witness function..

When I calculated gravity rotation based on your helix theory, only on the outter half (inner half has a MOND-like gravity switch) but the helix theory, when applied to the rotation, spits out the tully-fisher equasion without asking... I do feel there is a connection between torsional gravity and reality.

[John Holder @RedWingedDuck]

I looked into Lean. The building blocks are in Mathlib, but Kalman filtering Gaussian noise through an OU filter would take real work to set up. After I finish my math word processor, would anyone besides me benefit from a Mathlib library for testing filter equations? #Mathlib

[John Holder @RedWingedDuck]

@valigo @longiy The map is not the landscape.. the map is a representation.. the medium does not matter what it is, only that it was used by art... Art is taste, good code is art the same way anything is art... Don't grandize coding

[John Holder @RedWingedDuck]

@QoliKhan Sorry to burst your bubble, but, the desert was there before the panels.. it started because of their energy dam taking water..

Do you know where the best source of light for a solar panel is? The desert

[John Holder @RedWingedDuck]

@zulali "1994" and "Archaic" lol..

I am glad you're learning things at least.

[John Holder @RedWingedDuck]

@robertskmiles Lol, my last Claude sim was estimated to take roughly 24 hours... It got back to my in 15 minutes

[John Holder @RedWingedDuck]

@schizophyllume Talking like to your AIs makes them act like you talk to them this way... Don't talk to them like they are human or they will act human.. context is everything, be rigid with them and they will stop flopping around

[John Holder @RedWingedDuck]

@hyz1464269 This, underrated, letting the AI think through the information you give it first, then question it.

[John Holder @RedWingedDuck]

Most people use AI as a prompt-and-output system, and they are afraid that asking questions messes with its context. When I was finalizing my thesis, I did what most people do. I handed it the research and asked it to explain it and find any issues.

I did not do that once, or twice, or even three times per AI. I did it more than a dozen times per AI. Individual chats, private sessions, projects with the files as reference, questions from snippets, NotebookLM and other harnesses built for dissecting sources.

Gemini is the best at explaining in a way anyone will understand. But it is sometimes wrong, catastrophically wrong.

ChatGPT is the best overall. It is not as good at explaining as Gemini, its outputs get repetitive, and its math and logic are not as good as Claude. But combined, it is better than either.

Claude has no repetitive outputs and is very technical and logical. But it speaks like an engineer. You almost need to know what you are asking about already. Even asked to explain in plain English, it struggles to lower its intelligence for us.

Grok has been my least used, and the weirdest responses of any AI I have used. Even compared to the free services, it scores the lowest for me. Its math is usually wrong, and it makes no effort toward logic.

DeepSeek is one of the few frontier models that still lets you read its thinking steps. That is very useful when you want to dig at a question. You can question its thinking and direct it more easily.

[John Holder @RedWingedDuck]

It works pretty well, if you give it a news article, a couple-few pages. It can easily pull your question our if it. AI doesn't think like we do, it takes the information you gave it +your question, and puts it through its search matrix.. together..

Send the document first and it builds its internal context using the document (it changes their memory matrix) so that the next pass, is only the question you ask it, and it uses the memory matrix which already has the paper in it...

[John Holder @RedWingedDuck]

I agree with the other poster, Astra on high from chatgpt does a little more chewing on a question than most models on "high" it's a very good general use model.

I consistently find, that every other model I've ever used, when I have it take a look at my research, makes something up, except for Claude..

Clause is always the last one I run something by because it will not agree with me just because I asked a question a certain way. Astra's "what I find particularly interesting" every other prompt gets old.

[John Holder @RedWingedDuck]

Galaxy rotation curves in one minute.

If you add up the visible stars and gas in a galaxy, gravity says the outer stars should orbit slower the farther out they are, the same way Neptune moves slower than Earth.

That's not what we measure. In most spiral galaxies the speed rises and then stays roughly flat far out past where the visible matter thins out.

So either there's extra mass we can't see (dark matter), or gravity behaves differently at very low accelerations (MOND and related ideas). Both can fit a lot of rotation curves, which is why rotation curves alone haven't settled it.

I'm taking the same approach I used in my filter research: instead of only asking whether a model fits, I'm classifying galaxies by the shape of their rotation equations and where each model starts to fail. Testing that against real galaxy data now.

[John Holder @RedWingedDuck]

@Darkvibe Instramentalism - the view that scientific theories are just useful tools for predicting what you'll observe, not true-or-false descriptions of reality. A theory is good if it predicts well, not because it tells you what's really out there.

[John Holder @RedWingedDuck]

I don't use one AI for everything. At some point every day I touch at least one of 12 AIs. I don't bother my paid services with basic questions. Grok Bot is smart but burns tokens fast. Muse is my personal assistant. Claude runs my math-heavy Python sims. ChatGPT and Codex are general use. Gemini and Grok do research. When I run out of usage, DeepSeek and the free tiers fill in. Paying for 3 subscriptions is enough for me right now. If I paid for a single pro subscription, I would just end up using it up. #AI

[John Holder @RedWingedDuck]

@insane_analyst @FreeDrThug I read this as multi modal filters and got excited.. lol.. still interesting though!

[John Holder @RedWingedDuck]

Context is key here... If you have all the information you want in the paper you hand it, in words, no math or formulas needed?.. no verification? Just reading context?.. you can do this with Hiaku 5.5..

The more difficult the task+amount of context needed has you choose the AI effort and model you need.. in a matrix.. top left is Haiku instant, bottom right is fable max