Shadowy Figure Has AI Secrets

Autopay or Autopsy ?

AI systems are just sophisticated AutoCorrect app/programs.

AI is not Magic

Despite what AI Spokes-men-women-Hucksters would have you believe AI Data is usually contained in two files, not dis-similar to how a traditional database works: Data & Indexes.

(Although it is surprisingly difficult to research due to their landslide of deliberately obscure new terms. And maybe because some Two Trillion invested dollars depend on keeping it mysterious.)

(Please – do not confuse what this article describes – with systems that provides you with answers. Those working apps/programs that you interface with are called “Models.” This article is Not about Models at all. I’ll describe those in a subsequent article.)

AI Data is just a few Indexes. Large but not gigantic Indexes; as in Terabytes; Multi-dimensional Indexes.

They index relative value strength of words or tokens associations to each other.

(Opinion: There are some Two Trillion dollars invested in these systems. For some reason, some AI experts get really upset with my simplifying their data structures into familiar, yet “accurate” (as other experts agree), terms like “index.” I wonder if the money and removing obscurity have any correlation?)

An AI index can have hundreds to thousands of dimensions (Values per indexed word. e.g. a single index points to your zip code, a multi-dimensional index points to your zipcode and city, state, etc).

Those many, thousands or millions, of data values for each Word/Token are named Vectors or Tensors. (Pointers with strength values)

Large Language Models (LLMs) do not merely use and store Tokens (words or parts of words). Contrary to popular belief they do store English words and terms in a Vocabulary file: e.g. vocab.json, or tokenizer.json

A Vocabulary file does “vocabulary mapping” and includes a tokenized index. That index (“lookup table”) has a unique identifier for the smallest parts of each vocabulary noun / term / idea.

Example: “Football“ might tokenize into “Foot”and “ball”.

Connections File – The (“Neural Network”) :
This is the two Trillion dollar heart of AI. LLMs store Connections or weighting values in Index files named “model.safetensors,” (newer filetype) or “pytorch_model.bin” (legacy filetype) files.

Example: The term “Football“ has millions of connections such as “Sport,” “Game,” “Team,” “Win,”  “pigskin,” etc. stored in the connection file. This is the file created by training on data.

The many connections use numbers indicating strengths of specific associations. So for “Foot” its connection to “Ball” or “Sport” might have a greater weight (or proximity) than to “pigskin.”

The difference between a typical data index and Connection file indexes (“Neural Network”) is specificity vs relativeness. A typical data index, like a Spelling Checker or AutoComplete, have an alphabetized list of exact matches. (Newer Spelling Checkers typically do include a small set of common misspellings and synonyms, meaning that they are more than an exact match lookup.)

A “Neural Network” index is not that different, except in the numbers of data for each key. They have many index keys, possibly millions, that have values for how close or how related two words (tokens) are.

Putting AI to Work
What an AI Inference Engine does is reconnect “Football” and “Sport” for your questions.

*** Remember AI is just a sophisticated AutoCorrect app/program that REALLY Does  NOT understand anything.

It only Predicts what are the highest probability answers.

And we know how well AutoCorrect apps work don’t we? (Example: AutoCorrect turning “friendship” into “Fried Shrimp” and autocorrecting “caulk” into “cock“, and transmogrifying “I did speak with Eric” into “I did sleep with Eric,” or converting “Portugese” into “port of geese” and “porkncheese” or “diet sodas” into “dirt sofas” or “autopay” into “autopsy“)

I refer to AutoCorrect programs as “Smell-Chickers” (that’s Spell-Checkers – misspelled) due to the sometimes amusing errors they make.

# # #

Note 1: Eye Raising — If you want to see Search engines Choke (AI or not) ask how big (in Petabytes) the Claude Connections files are — (e.g. model.safetensors, or pytorch_model.bin.) Their Masters seem to want that info hidden.

Note 2: Connection Data File Size: While the Connection file “Neural Network” is formed from analyzing billions upon billions of examples, its size is remarkably “small.” I expected it would take petabytes. But no. I overestimated by a factor of ~500.

None of the leading American AI firms publish their file sizes. However we can calculate size roughly. With a trillion connection values, the data file may need only two terrabytes of disk space!

DeepSeek’s data (DeepSeek-V4-Pro) only needs ~1.6Terabytes.

Note 3: Regarding this article’s Accuracy. When I asked an AI system to identify any incorrect claim in this article, it responded —
“highly accurate conceptual simplification of how modern neural networks, particularly Large Language Models (LLMs), represent and process information.“

And that was my goal – to provide you with a highly accurate simplification of this enigmatic subject.

… More to come . . .

1. Fuzzy vs Determinative (Probabilty vs Certainty)
   

Quantum processing gives fuzzy, varying, statistical, non-replicable, erratic, or analog answers.

Digital computing by contrast gives explicitly digital, replicable exact answers.

Example: Wind Speed Estimation vs Password Cracking

We simply can’t crack a lost password with analog, non-replicable or fuzzy answers.

But for problems that don’t need exact answers like estimating wind speeds, quantum computing may(?) be worthwhile. Or for many-dimensional problems involving millions of variables like weather modeling they might someday be helpful.

2. Quantum Error Correction is Unsolvable

There is no such thing as Determinative Quantum Memory

Just as quantum processing results are fuzzy, the errors it produces are also fuzzy. And can’t be moved to or processed with digital memory.

3. Quantum Computers Cannot Exist without Determinative Quantum Memory

. . . which does not exist (See #2 above).

So why are you throwing good money at Science-Fiction?

# # #

Notes:

“Quantum computers: what are they good for? For now, absolutely nothing.” Brooks, Michael (24 May 2023). Nature. 617 (7962): S1–S3.

How Far Can Quantum Computing Go without Quantum Memory?
George Sidman, 2024 Medium.com

We keep hearing stories about Quantum Computing advances, like D-Wave’s Quantum Annealing. Stories designed to keep us on the edge of our seats – expecting a Breakthrough landmark, any minute now. . .

Sorry to throw Cold water on those “promises.”

But Quantum Computing is Not going to happen in your office, in our lifetimes. Why?

1. Non Repeatable; non-Replicable Answers ! Quantum computing has an outrageously poor Qubit Error rate of about 1 error in every 1,000 operations(2) (vs classical computer error rate of ~ 1 error in a sextillion operations).

Is this serious? Well, Imagine using your calculator giving you a wrong answer every thousand times you use it. Oops. Or imagine your car’s brakes won’t work every thousanth time you step on them. BIG OOPS. Except its not predictably the thousandth time, it could happen twice in a row!

This is not merely Fatal, its Incurable BECAUSE Quantum computer’s Probabilistic operations are inherently non-Deterministic – diplomatically called “Noisy.” * *

Deterministic means exact and reliable (standard computing) – vs approximate and variable.

And Quantum “Error correction” schemes are —

1 a) Classical – not Quantum (I’m trying not to laugh), and

1 b) Based on the Pooling of Ignorance principle — “In quantum error correction, multiple faulty qubits are combined to make reliable qubits, called logical qubits…”

1 c) Next to impossible to “… diagnose errors without ever directly observing them . . .”

Despite all that – could they be used for estimating? Questions where exact, correct (non-wrong) answers are not needed. (Obviously this excludes decryption which demands deterministic exactness.)

2a No Quantum RAM exists to store Qubits – either momentarily or long-term (Fatal)

(According to the no clone theorem, no quantum state can be reproduced completely meaning we cannot make quantum RAM)

2b And one Can’t use non-quantum hybrid RAM because logically we can’t read or transfer quantum data to conventional RAM – thanks again to the no clone theorem.

3. Quantum processors need to run at near Absolute Zero temperature (at minus 459 degrees F)  (Kind of eliminates “outside the lab” use. Fatal.) & . . .

4. Hoover Sucks: Quantum processors need to run in an “absolute Vacuum” (No “outside the lab” use.  If not Fatal, it is in Intensive Care on life-support.)

5 “Eye-watering cost” of a quantum processor (bad, but not necessarily fatal or eventually incurable)*

# # #

Notes

(1) Many of the facts and rationale described here are derived from 2024-2025 papers by George Sidman of Trustwrx.com

(2) A few basic arithmetic calculations in Excel might require 1000 cpu cycles.

* So far some $40 Billion has been spent for Quantum computing and anther 25 billion committed.

** Noise is not just quantum probabilities. It is also caused by any exterior phenomena like temperatures above Absolute Zero, data cable / conductor RF noise, and random cosmic rays.

Some further reading —

How Far Can Quantum Computing Go without Quantum Memory?
Why the Challenges of Quantum Physics May Deny Quantum Computing a Commercial Future

Why Quantum Computers Will Never be a Threat to Encryption
The Qubit Noise, Scaling & Memory Barriers Remain Insurmountable

More Technical Reading —

Why Quantum Computers Cannot Work – UC Davis

Fixing the Internet’s Fundamental Flaw with Trustworthy Packets
Giving IP Packets a Trusted Identity to Dramatically Improve Internet Security

This content is password-protected. To view it, please enter the password below.

Password:

Downgrade 50%

Yes, you can downgrade a Windows 7 system from 64 bit to 32 bit with only a little bit of inconvenience and cost in time while (contrary to most web advice) you  get to keep all your documents and folders AND your 32 bit programs and their settings !

Recently I obtained a laptop with Windows 7 (Ultimate) 64bit with only 2 gigabytes of ram memory.

After a bit of use, the laptop turned annoying to use because it was constantly running out of memory with only a browser running a few Tabs. It didn’t actually crash (permanently lock up or BSOD) that I can recall, but it would regularly slow to a crawl and give warnings to close programs because all memory was used up.

Since the laptop RAM was full, we couldn’t add any more memory, I decided to downgrade the 64bit system to 32bit to free some RAM memory up.

Oddly, I could find nothing on the web to tell me how much less memory the 32bit version would use. The best I could find was some articles with vague indications 32bit would use less ram, but nothing quantitative (other than MS which tells you to have 1Gb RAM for a 32 bit install and 2Gb for a 64 bit install), so I was relying in small measure on hope.

One of the difficulties in measuring memory use is the new and probably better way Windows 7 does memory management which now takes into account virtual memory, though how is not yet clear enough for me.

Nevertheless , after having (mostly) successfully done the downgrade I can now report that —

It can be done, AND you can keep all your files, documents and folder structure.

Even better, contrary to essentially all pundits’ warnings, you can even keep your 32bit programs and their settings without reinstalling (the programs) — IF you have desktop links to the programs and you have backed up the desktop links.

Your 64bit programs simply won’t work. You will need to get 32bit versions of your 64 bit programs – if they exist.

A couple of Caveats / Cautions:

• • Cost? You’ll need to obtain a new Registration/Activation Code. Microsquish uses a different Activation-code for 32bit and 64bit. MS may try to charge you for a new code, but if you have a legal 64bit copy they shouldn’t charge you.

• • You will lose your Windows System settings – such as mouse, keyboard etc.  Apparently you cannot (I didn’t test it) use the Easy Transfer Function to save System settings. There are some third party commercial (not-free) programs that claim to save these for you, but I didn’t test them either.

So here we go –

• • Of course, Backup everything — on a disk with a file system — where you can retrieve the select files you need.

• • Get a 32 bit Win-7 Install disk ready (DVD from an ISO or 4+GB USB drive). If you don’t want MS to charge you for a new OS, it should be the same type (Replace Win-7-Pro with Win-7-Pro 32bit – NOT Win7-Ultimate. You might consider calling MS in advance to see what their ever-changing policy is as of today. There are some programs that claim to save your activations state, but I tried two which didn’t work for me.

• • Run the install — as concisely and clearly described by Raymond from Malaysia

(Note: The “Custom” install does Not inherently format your drive, does not delete any folders or sub-directories,  does not delete any document files, or change the partitions – unless you want it to. It does rename your current/previous “windows” folder to “Windows.old.”)

NOW your 32 bit Windows 7 is running – in a “freshly installed” fashion. No additional programs are running. To get your previously installed 32 bit (64 bit programs won’t run any more), continue —

• • “Restore” the desktop icons(/pif) files for the programs you used on the 64 bit installation. (This will quite often let you run your 32 bit programs in the “Program Files (x86)” folder without a new install. Some programs that depend on setting stored in the Registry may balk at trying to help you.

• • Shazam or “Bob’s Your Uncle” your 32bit Win-7 is up and running and so are your programs – with all of your data !

Memory Reduction: Since I only took approximate ram usage measurements before the switch, and because of Windows new memory management I can’t tell you exactly how much less memory is used. Using Task Manager before and after, I estimate  the 32bit version before loading any programs, uses somewhere between 75-200 megabytes less than the 64bit version.

In any case its enough. I can say that in a week of use I haven’t run into either the slowdown problem or the low memory warning – even though I’m running two browsers at the same time, each with several open tabs!

So for my purposes – it worked ! — so far.

Now all I have left is to convince MS registration to give me an activate code for my OEM License that came with the laptop. Any hints ?

One nice review persuaded me to try CyberGhost.

Downloaded version 5 and installed it.

Just as it finished installing I got this outrageous message —

“You’ll need at least Microsoft Internet Explorer 8 to run Cyberghost 5 on your system !”

Then it quits and without my permission opens my browser to download Microsoft Internet Explorer 8. YOW !

For the uninitiated,

the best and fastest way known to attract viruses, robots and spyware

is to keep Microsoft Internet Explorer on your Windows PC.

Here’s where even the German Government officially advised against using it: German Government: Don’t use Internet Explorer

I was not amused and uninstalled Cyberghost immediately.

It gets worse.

After uninstalling CyberGhost “completely” and rebooting, I found a Cyberghost Service still running on my computer !

Now if I was to suspect NSA was “collaborating” with or bullying a VPN to see what we are doing, that’s exactly how I would expect a bad VPN to act.

You can rest easy, I will NEVER again use or even test Cyberghost.

For those who wonder how I managed to run a Windows system without Internet Explorer, here is a “How to” get rid of Internet Explorer article.

If you value your computer data, let me urge you to seriously consider wholly getting rid of MicroSoft’s Internet Explorer.