A quick guide to what Generative AI is and isn’t.

When I talk to people that have discovered and started using #ChatGPT and other tools, you would think they’ve found the elixir of eternal life. They say, ‘Nick why are you not excited, you don’t get it, this changes everything!’ You can imagine my answer.

There is a lot of general misunderstanding about #GenerativeAI and #AI #artificialintelligence in general, what it is, how it works and the pitfalls and issues, and so this article may help lay out the landscape, describe the underlying tech, what it is and is not, the structures, backends and the platforms.

There are now dozens of Generative AI systems, platforms, tools, products from US and many forget from China, huge companies not mentioned much. And yes I wrote this article on my own using my own creativity, grammar mistakes, oddities of my own style.

Why?

Because if you ask an AI to write an article IT IS NOT YOURS…

Although proving ownership in some ways will be almost impossible, the point is made. The generating part uses data and source material not identified or known — ‘caveat venditor’.

Last week I spoke at a Generative AI conference/meet up at a large law firm on London’s Bishopsgate, attended by a healthy bunch of AI enthusiasts, investors, geeks and AI users. All seeking further insight into what has become the phenomenon – GenerativeAI, which is principally a Large Language Model (LLM), as preceptors such as ChatGPT. The GPT bit stands for Generative Pre-Trained Transformer. The wider label — Generative AI. https://en.wikipedia.org/wiki/ChatGPT.

One has to remember Generative AI isn’t real AI, it is human controlled, led and biased. The humans behind it are the gatekeepers, holder of the official keys, confirm the provenance or not of the data, they control the output. ‘Output’, the provenance isn’t know, trusted, fact checked let alone permissioned. Yet this trend remains popular because it’s a new and exciting, a new toy, and does some cool stuff.

Entrepreneurs and Start Up Use

During the London event the primary discussions centred on Use Cases and quite rightly that ChatGPT and the 30 others are amazing tools, create a perfect storm for entrepreneurs to use to build their offerings, start their company and refine the business model. No longer do you need to hire staff, get service providers to work for you at cost and need to raise pre seed money before doing anything.

These tools reduce time to market, through accelerated ideation and to get operational, build the test business model, design the products and offerings ahead of time. All before you go ask friends and family and VCs for cash. These tools can create all the key assets, materials and if you must — the content you need, structuring and laying out, testing and further refining.

Enterprise Use

Also discussed, the impact of Generative AI in the corporate, large enterprise world, reflecting on the current adoption rates and uses. The room was in general agreement there are too many barriers for their use across the ‘enterprise’ to be green lit, reasons that also highlights many of the fundamental issues with this emerging tech.

The major blocker is of course compliance, given the Generative AI cannot prove provenance of sources, accuracy and thus copyright, or the reliability of the output. As of now the platforms (OpenAI) and products do not as yet publish their algo’s or source data. Indeed the wide deployment, integration of said Generative AI systems, will involve lots of testing and no doubt a rule book, legal opine and HR study as to the impact of productivity, loss of roles and presumably workers rights. Same old nonsense and why corporates are slow, can’t compete and will get left behind.

Highlights the real issues

In a sense though, these barriers to entry into the enterprise software world is the very reason many are concerned that Generative AI is moving us in the wrong direction, and most concerning is the revelation these tools also invent. There is an emerging tide of copyright and theft claims in the pipe, but then the legal systems is also not on the pitch. How many lawyers would understand the complexity, and yet the hardest part is collecting the evidence would take years to mine as proof at the back end. The loop hole, get out of jail is the assembly of outputs as next word based predictive models.

Of course we know and understand the big tech firms do not care about your, our data, as they have freely been harvesting us through behaviours AI for decades and selling it what is ours, denying access if we don’t agree — full Web2 head lock. ChatGPT is yet another tool for ‘human harvesting’ but this time you’re being raped with your eyes open and mouth closed. Ask yourself who gets the better deal?

Data Inaccuracies — it makes shit up!

What many mind blown users do not realise is when the Prompt or Question is asked, the LLM cannot strictly go outside its scoped parameters, (has data only up to 2021 apparently), but it will fill in the gaps generating its own data (answers) to deliver the output. This false truth, is then added to the data and further data to be mined and used again and again. So in a sense things get more inccurate and the truth gets further buried, depite users publishing unchecked data as facts and truths already.

Having said that, ChatBots using LLM are quite useful provided you do not use them purly for content generation, to do your work for you, to create stuff, instead is a great productivity tool for organising, structuring and for discovery search.

Generative AI — what are they?

To help the next part of the article covers the core technologies and platforms behind Generative AI. The entire product, tools and thus the hype centres on three essential parts:

1. The approach to harnessing the power of machine learning, deep learning.
2. The hardware specific computation — to crunch the data, for training the models.
3. The Service model on which the app you are accessing is created.

Generative AI a very expensive game, is developed, run and owned by the big tech firms, the usual suspects because only they can afford to make the vast investment required in gathering the vast data sets, building the LLMs and investing in the vast processing hardware. So in many ways there are few creators of Generative AI and lots of imposters that sit in front as a facade. Offering a new UI/UX, a different MRR pricing model which follows a premium period — only to find out the same back ends are often visible, aka Open AI, Dall-e and GPT 1,2,3 are running the show.

Personally, I prefer to go direct to the source and use the primary engines, libraries, tools and environments.

1. Popular Frameworks:

All ChatBots, what Wiki refers to all Generative AI, use a range of common frameworks, tools sets, libraries and approaches principally 5.

Here are a few examples of different approaches/components that involve Machine Learning (ML) and Neural Networks (NN to creating Generative AI service models and platforms. Can include text to text, voice to text, text to image and image to image — the actual learning process.

1. GANs (Generative Adversarial Networks) comprise two neural nets, one a good cop and one a bad cop (adversary) working to provide a version of truth, good/bad, true/false. Between real and generated data ( a key point) we shall return to.

2. VAEs (Variational Autoencoders) are a type of generative model that learns from a compressed representation of data and can generate new sample data using a Deep Learning models.

3. Transformers (the ’T’ in GPT) used in ChatGPT they rely on self attention popular in natural language processing (NLP) are heavily reliant on GPU processing.

4. RNNs (Recurrent Neural Nets) used for sequence generation (what comes next) that give some predictive performance opportunities.

5. Probabilistic Models such as Gaussian, Markov and Bayesian networks focus on probability and statistics, the prediction of what (letter/word) comes next.

2. Common Processing Environments:

Many LLMs and environments are hardware specific:

Most likely they will be using a combination of these hardware and application specific frameworks, but also consider other options from the Asian tech giants.

A. TensorFlow by Google an open source deep learning framework used to support AI tasks, generative models like GANs, VAEs and transformers.

B. PyTorch by Facebook another widely used deep learning framework, uses a dynamic computation graph supporting research for generative models.

C. NVIDIA Cuda a parrallel computing model to unlock the power of GPUs for deep learning tasks, supported by the cuDNN (Deep Neural Network Library) for optimising processing.

D. Intel One, a deep learning neural network library that provides the building blocks for deep learning apps on Intel architectures and CPU training models.

E. AMD ROCm (Radeon Open Compute) an open source platform for GPU compute on AMD GPUs. Tools and libraries.

F. OpenAI Triton a back end system designed for deploying deep learning systems including an approach for generative systems.

3. Most used Service Models:

Open AI GPT (Generative Pre-trained Transformer) tools and model based on formal GPT series 1,2,3. Opening large data sets through NLP tasks and text generation.

NVIDIA DL GPU’s (GPU processing units includes the Tesla Quadron GPUs and Tensor core GPUs widely used by many for training and running large generative models that need parallel processing.

IBM Watson — a suite of AI tools and services supporting NPL. Offers AI powered apps to support the creation of generative AI models, on IBM hardware

Microsoft Azure Cognitive Services a collection of AI services for developers to build AI based apps. Includes — LUIS Language Understanding Text Analytics.

Amazon SageMaker a cloud based model using AWS to simplify building, training and deploying machine learning models for generative AI.

SalesForce Einstein is an AI powered platform integrated into Salesforce CRM applications for generating personalised content, recommendations and Chat BOT deployment.

Adobe Sensei an Ai machine learning framework used in various Adobe products like Photoshop, Illustrator and Premier Pro. Delivers AI driven content generation.

Unity ML Agents a toolkit allows developers to integrate machine learning models including generative AI based on Unity simulations, games and platforms.

Tanks on the Lawn

My issues with Generative AI stem from its use and user behaviours. Oh before we start I want to make it clear — the tools like ChatGPT are not intelligent or good at problem solving. There is no actually intelligence going on, no defined cognitive processes, no experience recall, no emotional context, no ethics, no morals and the source material (and the algo’s) are of course completely biased and remain private. In the sense of human direction, involvement and objective setting.

Yes many Generative AI systems are supported by Large Language Models, models that are vast and yes impressive, trained on billions and billions of source material, books, articles, papers, databases mostly written, prepared and owned by others — where permissions have it seems may not been sought or granted. And yes some of these sources contain hateful content, human madness, lies, distortion of truths, re-wrting history and likely elements of the ‘Woke’ nonsense sweeping the world.

You can argue the developers of Generative AI systems are ‘Knowingly Concerned’ as designers fully understand proving sources and accuracy is vitally important, not because of how LLM work, but because its the right thing to do for a tool adopted by hundreds of millions already. They have an obligation, yet the founder Sam Altman then says he’s concerned, worried and a little scared? Is he saying it’s an Oppenheimer moment?

The claims of THEFT, COPYRIGHT infringement and HARMFUL content are on the rise, complaints the source materials are not fact checked or filtered sufficiently or at all. The source data is consumed in vast quantities into predictive LLM — pixel/letter/word by word — what comes next, the predictive bit. Turning Inputs into Outputs as answers?

Overwhelmingly the issue with Generative AIs is the User, the public are initially wowed by this tool but naively think by asking questions, asking for the tool to ‘generate’ content, a story, code, a song, a picture, a report, a presentation thinking this OUTPUT is universally theirs?

Yes, people really are that naive to think this often detailed and valuable information that is being ‘mined’ is for free. Some believe they will be better at their jobs, more creative and get a promotion — until the first legal claim comes in. Or the employer hires someone on 10% of your salary or another AI to asked the questions and come up with the Prompts from the now extensive libraries. Free stuff always comes with a cost.

That is why I don’t buy the broader arguments for wider job losses. Many complain that writing jobs, journalist jobs and other creative pursuits will be replaced by these tools. But offer no basis of fact or understanding of what is really going on structurally. Have you actually read what a Generative AI produces? The output is largely dull and uninteresting at best, and fake at worst with no emotion, totally bland with little sub-text, lacks purpose and conviction.

But some will say it’s because you’re not asking the right questions or using the right prompts. True but also not true. AI’s producing interesting and unique content isn’t there? Wrong, I hear people say. It’s worth remembering that everything the AI uses is based on human driven content, original thinking and sources that are already out there, as the starting point. But now it gets dark. The Generative AI sees and knows all and rams chunks (words) together to generate apparently great, unique, creative stuff. But this isn’t true either.

In any event when an AI writes text, a story and an article they are without an understanding of human context, emotion, experiences, uses data someone on someone else’s bias and opinions that buried in the source material. Albeit the developer has rountines that wash the data and maybe even change the bias, all true meaning and human context is completely lost. Have you actually read this stuff?

The resultant text output is, unexciting, untrustworthy, unreliable. As for images well that’s a different level of argument but also similar in experience. Pictures created from other images, pictures, reference material, photos, maps, visuals — a mash up of things that were human created, in the beginning. I make this point is as we use generative AI the source data is also changing, and thus much is already AI generated.

Are humans going to give up being creative, selling their soul to an AI?

Generative AI tools will not replace creativity unless humans allow it. If we start to rely on these tools rather than ones own creativity is considered a crime against what it is to be human. We are not ready to become sheep although that is the wish of WHO, WEF and many governments. To dilute our humanity, freedoms and take our assets. Having tried to take away our freedoms through lock-downs, they intend to take our money (CBDC), and have already taken away our health with unnecessary vaccines. Do not let them take our creativity, our intellectual freedoms. I know this was a bit ranty but I feel better now.

Generative AI is a tool for us to use, and is NOT a replacement abilities at any level. After all there is no real thinking (thought) going on, no real intelligence, no real cognition although big tech will claim otherwise. I argue differently having been in tech for almost 50 years, learning computer science in the 1970’s. Computers are dumb, will remain dumb, cannot feel, think or see the future. They are programmed by software which is human generated for the most part. Although AI now writes code. Is it sentient, not yet. A ChatBot did a good job at faking it (the basic Turing test).

I am affraid I fall into the Roger Penrose camp believing the current Von Neumann based compute will never be able to mirror humans or capture consciousness, but Quantum Machines will…

A reminder — Computer hardware:

Cannot understand

Cannot think for themselves

Have no emotions

Cannot see the future

Cannot see reason

They don’t know

Cannot be autonomous

Cannot see the impact of its outputs

Cannot understand emotion, context

Have no view of ethics

Don’t know right from wrong

Cannot recognise bad people

Cannot choose

Are not creative

Don’t care about biases

Don’t care about people, humanity

Don’t care if their outputs harm people, destroy lives

Don’t be Fake

So here is the thing. If you ask ChatGPT to write an article for you, a book, a paper or script please don’t claim as yours, at least own up to using a tool to write it. Do not ignore your role in what could be deemed an assisted theft of data.

You already know the source and reference material isn’t known or validated, imagine if your material, invention, ideas were mined as part of this — which it probably is. You know the source material is owned, copyrighted and thus created by someone else. Be sensible.

Of course ChatGPT will say the AI is essentially a predictive ‘word game’ guessing what comes next. They will also claim the AI does its thing without malice, intention to defraud, without undermining someone’s work but only they know the source data it is trained upon and what the algo’s are told to generate.

The Law

Yes our laws are inadequate when it comes to crimes where software is involved. We’re not talking IP. Today AI (software) cannot be accused of anything? Indeed let us call them ‘acts of the software’ are not covered by law. Who then is responsible? The owner of the software (licence holder)? The programmer(s) and team? The founders? The person that source the data? The platform that extends the UI? The promises made on the website? So who is accountable? Good luck with that one.

Proving intent and what is going on in the complex world of AI. OpenAI is a research lab that is meant to share development techniques, tools, approaches and code. Unfortunately for you ChatGPT code and algorithms aren’t open sourced.

The Rest of the pack

Generative AI’s are all pretty similar and many use the same backends, components and algorithms.

For most ChatGPT and other AI tool users https://themeisle.com/blog/chatgpt-alternative/#gref

1. Google BARD
2. Microsoft Bing Chat
3. Chatsonic
4. Github Copilot
5. Wix ADI
6. Jasper.ai
7. WriterZen

Worrying Developments

There are amongst these tools new AIs that scrape content from websites, platforms, social media to harvest the data for use in the Generative AI tool landscape. I may be old fashioned but this is most definitely theft in plain sight is it not? As permission has not been asked. There are tools that can scrape (sc-rape) LinkedIn for example and collect personal and business data used in outreach campaigns — ignoring the User Settings of Privacy and User Data.

Hopefully you get the point. However software code can extend its parameters and redefine limits, ranges and adjust its efforts based on recent experiences, and what is learned, and the extent of the data sources offered. Essentially right versus wrong outputs, or degrees of almost getting it right and wrong— as the creator has set the objective and also has the expectation of the result looked for.

Remember the Big Tech firms are looming in background.

The systems, architectures and frameworks are mostly owned by big tech — Amazon, IBM, Microsoft, SalesForce, Intel and Adobe and in China — Tencent, Huawei, Baidu, Alibaba also have LLMs and ChatBots being deployed with user bases that dwarf everyones.https://techwireasia.com/2023/07/generative-ai-is-progressing-well-in-china/ Arguably the west is already behind and ChatGPT because of its mention of the CCP leader will be banned.

For this article I have referenced Generative AI options in the US and Europe. Although interestingly to note in China, to use a ChatBot you need to enter and register with your National ID card, food for thought. And it is well known China sensors technology firms. It is obvious certain Generative AI systems will be blocked or censored outside their regions.

The Real AI

The really scary stuff goes on in the field of general AI systems sometimes called Generality or Artifical Genegeral Intelligence (AGI) that are not, as many people might think, a collective of AI’s, a grouping and co-exiting to plot to bring humanity down.

Real scary AIs work diffrently, where there is no reference data, nothing to steal, copy or mine. These systems are much more powerful and they are for me genuinely scary. The question always raised is the dystopian threat. will Ai end humnity or will it indeed help save it?

With a prevailing wind behind us and getting rid of the idiots (mostly old men) running most countries, humanity stands a good chance if making it. These AGI yes can be used for bad things, but these systems will also help shape humanity, prevent humanity from destroying our planet and themselves (hopefully) and they are not a distraction as Generative AI are now.

General AI principles are different and work on the basis of envisaging scenarios, seeing into the future, searching for outcomes, covering ground and possible pathways that vastly outstrips human ingenuity, cognitive reach and imagination.

These systems are similar to game and targeting systems as they operate within boundaries and rules, (checked, validated and policed rules) but can play out a gazillion scenarios instantaneously, and by doing so will extend human capacity for understanding, growth and survival. The question — should we? The Hippocratic oath? In many ways these systems can police humanity, but in the wrong hands find our vulnerability, although I will argue we all know our biggest vulnerability as a race — one of greed, power and control.

I write about these General AI systems in more detail next time.

About the Author.

Nick Ayton is a Futurist works in Deep and Frontier Tech..