The Elephant: NLP AI Primer

Intro (P1)

Intro

The Wright Brothers made a glider and crashed it in a pond.  Enough people came to watch that there was a crowd in the picture a photographer took.  That was enough to be a tipping point.  An old dream stopped being a dream.

ChatGPT was that for a thing we do not yet have as clear a name for as "flight".

There were gliders before the Wright Brothers, plans, sketches, and the idea of humans flying was older then writing.

Whatever the "flight" is for NLP AI (Natural Language Processing Artificial Intelligence, which is what powers ChatGPT), it's not a new thing; most of us have touched it and used it as Google Translate if nowhere else.  It's a visitor from a family we have had come calling before;  calculators, PCs, the internet, and search engines are cousins that started as strangers and now live in our homes and pockets.  

Much like the newspaper readers of 1903, at some level we are starting to grasp that this ends in industries and tremendous change.  "Flight" meant the world got a lot smaller, more wonderful, fortunes were made, and it also meant bombers and bombs.  

NLP AI changing our world is not “a future”.

What should you do?  

1. Get Educated

2. ???

3. Profit

A Starting Point

What Happened: 1. NLP AI models became publicly available 2. that allow us to make and imagine much more powerful tools and 3. got enough press that the world started to notice.

What we are looking at here is not "ChatGPT" any more than one person named Chris is the US.  It's not a new type of search engine, a chat-bot, or a way to automate customer service, though it will be used in those ways.  It's not really NLP AI (Natural Language Processing Artificial Intelligence), as models using other types of ML (Machine Learning) can do the similiar things.  It's not AGI (Artificial General Intelligence) AND it doesn't need to be to change the world.

Step 1: Accept that It's Not "Optional"

This will change the world in the same way that "Agriculture", "Industrialization", and "Globalization" did.  Even if you avoid it or don't like it or just have a life where it does not touch you much, it will affect you and everyone you care about.  Being angry will not help.  Those that choose to use it will "win" any competition with those that don't.

You can either go with denial and fear or acceptance and education.  The second option is better.

Step 2: Naming It

The word we need should summarize what these new tools do in the same way "flight" means "tools that fly".

Though I do not think it will stick, I will offer a step towards that;  "QCRAT (“quickrat”).  It describes the function in the same sense a hand drill is an electric motor that turns a shaft with a thing on the end.  In a sense, this is the “flight” of “NLP AI”.  Wordsmiths, please do your thing.

What this means in depth is a rabbit hole better gone into a few steps later; follow the flow or skip to “Step 5” below, your call.  

Step 3: Describing It

It's a "drill" that works on "digital products" we currently use "educated humans" to create. 

It’s a tool for doing “tasks” - Human/AI Collaborative Task Completion.  Using the “demo version” (ChatGPT) it has been successfully used for tasks like optimizing and writing code, creating hundreds of Genre novels for sale on Amazon, and a lot more; YouTube is full of “world’s firsts” ranging from silly to amazing.

If you can take a task and describe it as a “story problem” in a way that the model “understands”, it can do that task.  That's “writing the query”, and more on that later. 

It’s not a thing that replaces people; it’s a thing that makes tasks more efficient on three main axis - faster, more precise, and requiring less subject matter expertise, and yes, in the end we know that ends in fewer people doing the same jobs.  

Describing new things is a lot like "triangulation" - you offer known points that describe the thing in the middle.  Here are a couple more attempts at that:

Step 4: Chart the “Fronts” - Vectors

These are broad parts of this technology, as things you can buy or sell and develop.  Each "Vector" will become an industrial "Sector" with products you can have delivered by Amazon.

Interface - To allow Users to use these tools

Models - To power these tools

Data - To fuel these tools

Ethics/Social Adaptation/Education/Values - To deal with the effects of these tools

Users - To provide better operators for these tools

The fourth vector is something we sometimes don't see as industry, a bit harder to name and describe, and still, leaving it out is a mistake.  Fortunes will be made and lost here as well, and Amazon will deliver it.

Step 5: What is QCRAT?

Query

Completion

R(Reaction, Revision, Repeat)

Analysis

The T is for tasks, and in a way it could be “silent” as all of this is about Tasks.  It’s not, because we might slip and fall and think this is about making a chat-bot say naughty things or a better way for us to shop online.

Using an Interface:

1. User makes a Query

2. Sends it to a Model

3. The Model does Analysis to it

4. And sends a Completion back

5. which is displayed by the Interface to the User who Reacts

Then:

1. IF it's good enough, good.

2. IF its not - its Revised and a new Query is sent back to a Model (Repeat until 1.)

Through all this, the A (Analysis) is happening, the most obvious being what the Model does to the Query in step 3.  As this technology is developed and the tools get better, the A will happen in a lot more places.  To keep that short:

1. Pre-Processing - the Query itself will be Analyzed before it is sent

2. Post-Processing - the Completion will be Analyzed after it returns from the Model

3. Meta-Processing - the Interface will be Analyzing the User, the Model's owner will be Analyzing the User, the Queries, and the Model, and more in that vein.

All of that becomes Data.  Different Models will be used to Analyze different parts of "the Elephant".  It gets messy.

Step 6: Queries

Queries are not new.  You could get PhD's in analyzing them in the 60's.  They are a part of any database and every search made using the internet.  What is different is the Models that are now available.  Queries are no longer for searching for numbers in a database or for information on the internet anymore.  Instead, the right query to a Model with the right Data available makes a new thing - Queries are no longer for “Finding” - they are for “Making”.

This is a difference similar to roads before and after cars; we have some vocabulary, which helps, and we have guys who know how to make real good roads for carriages and horses, and that helps too, but it's a new thing at the same time.

Part 1: Anatomy of a Query

A Query is, in essence, a "story problem".  If you can describe a desired outcome as a story problem, this (a QCRAT Platform) can "drill" it.  This is the most complicated part of using and designing this new technology; "Framing the Question''.

Queries have three main parts:

Task/Template - What result the user wants.  

This is sometimes not explicitly stated - the outcome can be inferred.

Examples: {A} a business plan, {B} the solution to a math problem, or {C} a joke.

Parameters - How the task is modified and described, as well the direct parameters used by the Model.  

Sometimes there aren't any, or they happen "behind the scenes".  A NLP AI can infer them or they can be implicitly stated.

Examples: {A} a 1 page result, {B} using this Model and this amount of randomness, or {C} in the style and voice of this dead comedian.

Data - The Data to be used in the "story problem".  

The Model may have access to more Data and the Model "has some Data in it" (see Model State and "Context").  

Examples:  {A} your draft and notes collected for a business plan, {B} the numbers and symbols that describe the math problem (2 + 2), or {C} your favorite jokes.

Part 2: How NLP AI Models change Queries

Keeping it to a few examples:

1. They can (somewhat) understand common language

Example: You can ask in common language for a Salesforce Query that will pull a list of all your customers who have birthdays next month.

2. They can tell you what data is missing from your "story problem"

Example: They can not only assemble an estimate for a job but tell you what data you are missing to make a complete estimate and give you a worksheet with directions to get that information, and explain the concepts in common language.

3.  If they have enough examples of similar things they can “mimic” (establish a Template and match using your Data and Parameters).

Example: ChatGPT has “seen” enough “thank you emails” that this is a “template” you do not need to define.  

It’s “seen” enough of most anything written (yes, that's also math and code and Chinese), and if a “task” or “template” is not in it or new, you can feed in enough examples that it can understand that too.

This means many tasks that are now time consuming, expensive, and require specialized knowledge will need less of all three.  Of course, knowing how to phrase a query, using these tools, and producing “training data'' (used to make and customize Models) become new “specialized knowledge”.

Fun fact - if you want to make “training data”, it goes a lot faster with a tool for that - see the next step.  Any QCRAT interface will passively create it just by being used.  ChatGPT sure does, even as a demo.

Step 7: Envision the First "General Purpose" Tool

A good way to understand what something means is to anticipate the first "general purpose" tool that will come from a new technology.  A "plane" described like this - that it has an engine and wings, it leaves a runway, a person "pilots" it, that it can lift cargo, and that it needs to land - is a good start.  From there, you have enough to start anticipating what it will mean.

As a big ugly word someone will make much shorter for us before long, let's imagine a: “General Purpose QCRAT Interface".  For short, lets call it a “Model T”.  If you want to imagine it, imagine a work truck - a car, a backhoe, or a semi are better for specific tasks, but you can do most things with a F-150.

Imagining this “truck” mostly has to do with describing the "interface" or app you will use it through - everything that is not the engine and fuel.  Your experience of a truck is mostly the chair, the wheel, and the dashboard.  If you see it from outside, it looks like a box with windows and wheels.  

The details of the publicly available Models and what Data you can use will change rapidly.  This is like engines and grades of fuel available for a truck; the basic functions and design don't change for the “Interface”. 

Being able to use different “Engines” and “Fuel is the first thing that we will define.

1.  "Model T will be able to mount any engine (Model) and use any fuel (Data) with minimal hassle"

It will make a broad variety of tasks more efficient.

2. "Model T will work on any task you can define as a story problem."

It will have an interface that lets you do most functions without fuss and see what is going on on your "dashboard".

3. "Model T with have enough of an interface a non-expert can use almost all its features"

4. "Model T will display what is going on in a way a non-expert can use"

Accessing a QCRAT platform through a web-client or a phone or a desktop app or a smart watch doesn't really matter either.  Import/Export and connecting to other apps is also basic.

5.  "Model T will connect to anything reasonable with minimal effort from a non-expert"

It will collect your Data and let you see and use it later.

6. "Model T will save your Data, to include series of queries, completions, and their parameters"

7. "Model T will be able to search your Data and display it to you visually"

It will have basic security features.

8. "Model T will have locks, keys, and modes where your kids can use it without inappropriate access or financial damage"

It will handle common issues, like Queries or Completions being too large to a single QCP (Query Completion Pair) and use your accounts for connections, automatically.

9. "Model T will automate common functions"

You will own it and be able to customize it, though customizing it might take expertise.  If you want to use your own Data and your own Model and for all that to be secure and isolatable, it will do that.

10. "Model T will be something you can own, modify, and isolate"

Of course, you will be able to use a “free” one that collects your data or prods you with banner ads.  You may care that your results are censurable by giants, or maybe not.  I wrote this in Google Docs.

Real talk, something like this will be available for use and purchase soon.  It's not now and your guess is as good as mine, and that is important to know.  That something like a “Model T” exists somewhere isn’t a bad bet: it's hard to imagine that the institutions that make the Models do not have tools like this - if only to make the “training data” needed to build the Models and test them.  Many working pieces of a “Model T” exist Open Source on GitHub right now.

Step 8: What does a "Model - T" Mean?

For one, lots of things you may want to do will be easier and prices for some things will get cheaper.  Bad news, that applies to everyone.  

In practice, it is safe to bet that a lot of "white collar" jobs will be done much more efficiently.  That means lay-offs and that means less high-end consumers buying stuff, and they will still have student loans and mortgages.

It means all those “Vectors” are opportunities, and that's good.   Who thinks those vectors will employ as many people as they displace?

Remember the "Ethics/Social Adaptation/Education/Values" Vector?  It means we will need that.

Step 9: What We Have Now or "The State of Play"

As of now, this what you can access easily: 

Part 1:  “Individual” Access

Right now, most access to NLP AI is through OpenAI.  My suggestion is you take my word for nothing and go find out yourself by using these things directly.

Right now, you get "mostly free" "partial trucks".  They are "demo models" at best.  Here are available options as they stand today:

Low Barrier Access:  ChatGPT - "one wheeled truck"

• “Oracle” style interface

• Remembers some of your earlier “conversation” (1.5 screens or so on average)

• Connects to one Model (ChatGPT 4.0 as of yesterday) which is customized to not do anything that will cause a bad public reaction (which limits it somewhat)

Medium Barrier Access:  OpenAI Playground - "two wheeled truck"

• Several Interfaces

• One of them remembers some of your previous “conversation”

• Connects to a list (9 right now) of Models

• Allows easy access to some Model Parameters

High Barrier Access:  OpenAI API - "three wheeled truck"

• No Interface unless you make it

• Allows you to store files which can be referred to in a Query

• That can be a log of your “conversation” as long as you want or data of your choice

• Connects to many Models, and if your interface can handle it a single Query can be sent to all of them at once

• This access is enough to use “QCRAT”.

Using them, you can see a lot of what a truck will be and while they are not easy to use you can do "work" with them.  They are a lot like a version of a corded drill you can only use in the showroom that has a 6 inch cord; you can still make furniture with it if you have the patience and it is already sometimes faster then doing it at home with the tools you have, but mostly it's there to convince you to buy one later when better versions are available.

Not surprisingly, all data you use and the data generated by your use of these interfaces are claimed by OpenAI - you do own the “products” and can use/sell them commercially.  Still, if you care about privacy, like any business that doesn’t want its IP or financials to be seen, it's a barrier.

Part 2:  “Enterprise” Level Access

So, maybe you think your business could use this, but need that security?

Enterprise Level Access: “four wheeled truck”

• Your Interface

• Storage is on you, and the Data is yours

• Connects to a Model you control

If you want a "four wheeled truck" the good news is that the Models (even making one yourself using things you can download Open Source) and the Data are not serious issues.  The issue is the "Interface" part and training your Model (of which the hard part is generating Training Data).

As a rough estimate, you can get on the net right now and download or buy all the parts and even hire people (demands high) to “assemble” it for you.  Again, the biggest pain in that is the Interface (custom software development) and generating the “Training Data” (both because you want a tool for that and people with the specialized knowledge are not plentiful).  

If you have some Tech people and want a minimal working model, an interface for a specific purpose, and data to make it useful:

10-500k? A few months?  

If you want to buy everything and make something to serve estimating for a large General Contractor or similar business:

Millions and a few years?

Or, you can wait.  The best bet for most people and businesses is to wait and adopt when it's easier and cheaper.  Of course, anticipating this, planning for this, and watching this unfold is probably key.

My Reaction:

I want a “truck”.  I want you to have a “truck” too.  I don’t trust you all to have “trucks” but I am more afraid of a world where only the few have “trucks” - just like with actual trucks - then I am of everyone having trucks.  If we both have trucks I can compete, even in an uphill climb, and I am fine with that.

Good information on this topic is hard to find and that concerns me.  To be more direct, when I got done learning about this myself (I went hands on first) and went to see what other people had to say, I was a bit stunned.

To keep it simple:

I am more concerned with tilting the table towards everyone having access to trucks and information than the layoffs and upheavals this will bring (and that's already started), because the harms are inevitable, and access to good trucks and information is not.  Even if it happens sooner, if we see it coming and can make choices, this ends up better.  

I am working on a truck (small software startup).  I wrote this.  My intent is to get into depth on different parts of this, one at a time, in my spare time.  What I focus on and how much time I spend on this is somewhat up to you - this has to help you, and me, or it's not worth my time.  

If you want a truck, think this is worth knowing more about, or would like to help:

Bookmark this, Share this, Subscribe, or if you're really serious, contact me.