1. Introduction to ChatGPT:
The AI chatbot ChatGPT, which was released in November 2022 created a buzz and a revolution among digitally savvy enthusiasts. It became mega-popular in just 5 days of its release, reaching one million users. In 60 days, it reached one hundred million users. From a historical perspective, TikTok took 9 months to reach one hundred million users. Instagram took 2.5 years and Netflix took 3.5 years for its first million users.
GPT is short for “Generative Pre-trained Transformer”, which is a Natural Language Processing (NLP), Generative AI [1], [2] model.
With Generative AI, we can create content like audio, images, text, simulations, software-code, videos and more just by asking the AI model in a simple language communication. This opens boundless possibilities to come into reality.
ChatGPT is developed in an interactive chat interface model, which is pre-trained to understand your questions, and provides you with a human-like response, just like a subject matter expert (SME). It is based on the transformer architecture [1], [2], and it was trained on a large dataset, from which information is pulled during the construction of the AI responses to your questions.
This chatbot generates human-like written text responses meticulously designed for the context within which your question was asked. The responses are lightning fast, and the scope is to the extent the transformer was trained on, which is most of many things on Internet, till September of 2021 for GPT-3.5. GPT 4 can query the Internet and use calculators and other tools to complement its work.
Within the scope of the information ChatGPT was trained on, you may ask a simple interactive question and it will provide you with the exact answer based upon the context of your question. This way you can develop knowledge extracting conversation with ChatGPT. This has made it into a viral sensation and seen as embryonic version of AI in the hands of masses. It is bringing knowledge to the grasp of the common person like never before. One needs no unique skills or training to use it. If you know how to use WhatsApp, you are good to go. This is revolutionizing the information digital age, which is as big as, or even bigger than the eighties invention of Internet or the smart phone revolution of 2007 [1]. This AI technology is a disruptive one with far-reaching implications to society, and we will examine those in this paper.
2. AI & Chatbot History:
Artificial Intelligence (AI) is a field of computer science and engineering associated with creating machines with human-like intelligence. That means the new AI software and hardware combined must have visual perception, speech recognition, decision-making, and language translation capabilities, to say the least.
AI idea was first observed in the ancient Greece, where myths and stories featured robots with associated automatons. However, the modern field of AI took shape in the 1950s. In 1956, at Dartmouth College, New Hampshire, USA, researchers established a goal of creating a “thinking machine” for one of their conferences. That is the first brick in the AI foundation and in the coining of “AI” acronym as a word [1] that we use today.
In the sixties, considerable progress was made by AI researchers with the development of the first computer game-playing programs and the first expert systems, with a limitation of narrow-domain decision making. The next era from 70s to 90s is labeled as “AI winter”, as not much funding and research went into AI. The focus of this era was on the development of advanced and powerful computers. We all know what that gave us.
In the last two decades, AI is again in the front headlines, where specialized robots were brought to the market. Enhancements in the field of Machine Learning (ML) and Deep Learning (DL) play a vital role in this.
ML uses algorithms to learn from big data to perform tasks without explicit programming whereas DL on the other hand uses complex algorithmic models, based upon how human brain operates to process unstructured data (documents, real world images, text, etc.).
In mathematical terms, DL is a sub-set of ML, and it is more complicated and hence capable of handling more complex and granular tasks than ML. A distinguished feature of DL is the inherent capability to learn from its own mistakes and improve its performance over time.
Sophia [1] a social humanoid robot developed by the Hanson Robotics a Hong Kong-based company, which was designed to mimic human social behaviors and inspire feelings of love and compassion. She (Sophia) is equipped with sophisticated AI, ML and DL software for general reasoning, chatting, imitating human gestures, facial expressions, recognizing people and maintaining eye contact.
Of course, there are multiple other robotic younger and older siblings of Sophia [1] that exist today. BTW Sophia birthday is on Feb 14, 2016. The trend has taken off by the creation of animals like robots and in many other form factors.
3. ChatGPT timeline & People behind it:
GPT (Generative Pre-trained Transformer) is a family of AI advanced language models developed via deep learning (DL) capable of generating human-like text. GPT is also called generator only Open Domain Question Answering (ODQA) system [1]. OpenAI Transformer is a member of GPT model, which is trained on large textual datasets namely, research papers, Internet web sites, social media interactions, digital books, and all other forms of publicly accessible data on Internet and dark web. (Exact training material is not documented for the public.)
A deep learning model by the name “Transformer” was first introduced in a research paper in 2017. ChatGPT Transformer architecture model is based upon a refined version of this 2017 paper, re-written in 2018 and used a decoder-only transformer network with 117 million parameters. ChatGPT uses the same architecture but with different parameters and training data. ChatGPT is fully capable of answering your questions, based upon any of its trained data via your interactive questions.
In 2015 Greg Brockman (Chairman & President of OpenAI), Sam H Altman (CEO of OpenAI), Elon Musk (left OpenAI in 2018), Reid Hoffman (Microsoft LinkedIn cofounder), and few others (Jessica Livingston, Peter Thiel, Amazon Web Services, Infosys [1] and YC Research [1], [2]) formed a nonprofit research organization named it as OpenAI. They raised $1 billion from these funders initially and established the mission statement of OpenAI as …
“Artificial Intelligence should benefit all of humanity” [1].
OpenAI goal was to democratize the access of AI to public and prevent its misuse.
2024 Update: Elon Musk is no longer a part of this consortium. In fact, Elon Musk filed a lawsuit [1] against OpenAI and its CEO, Sam Altman, alleging a breach of contract. Claiming that OpenAI violated its founding charter by prioritizing profits and commercial interests over the public good in the development of artificial intelligence (AI). Specifically, the lawsuit points to a multibillion-dollar partnership between OpenAI and Microsoft, which Musk believes diverged from the original mission of OpenAI.
OpenAI in 2020, announces GPT-3, a language model, trained on large Internet datasets. GPT-3 is aimed at natural language answering of questions, it can translate between languages and coherently generate improvised-text with the help of an inbuilt API. For understanding purpose, as somebody has said on the Internet…
… to consider the backend GPT-3 “language model” as the car engine, and ChatGPT “web interface” as the body of your car. Only when they are combined are they useful to the masses.
As per Stanford University, GPT-3 has 175 billion parameters and was trained on 570 gigabytes of text. For comparison, its predecessor, GPT-2, was over one hundred times smaller at 1.5 billion parameters.
In Nov 2022, OpenAI launched ChatGPT chatbot, based upon GPT-3.5. OpenAI forward revenue projections for 2023 are $200 million and in 2024 it is expected to do around $1 billion. It is also expected that AI market in Cybersecurity is set to surpass USD 94.3 billion by 2030 [1]. If this happens to be true, it will revolutionize AI and associated sister industries going forward.
Back in 2019, OpenAI joined forces with Microsoft making them a key strategic investor, where a one-billion-dollar funding boost was extended to OpenAI in exchange for an exclusive license to the OpenAI’s technology by Microsoft.
In Jan of 2023 OpenAI was in talks for funding that would value the company at $29 billion [1], double its 2021 value. On January 23, 2023, Microsoft announced a new multi-year, multi-billion-dollar (reported to be $10 billion) investment into OpenAI. In less than three weeks (Feb 7, 2023) Microsoft announced it is incorporating ChatGPT AI technology into its Bing search engine, Edge browser, and Microsoft 365 product line. As an unofficial word out there, Microsoft has acquired 49% stake in OpenAI, and all OpenAI product models training, testing, and production hosting are running on a global scaled Microsoft Azure cloud platform.
3.1 AI Image generation via prompt
In 2021, OpenAI introduces DALL-E and later DALL-E 2, which is a text-to-image generation tool, capable and proven to create and edit realistic real-to-life high-quality images and artwork from a description provided to DALL-E in natural language. DALL-E is available via an API interface and many web sites are utilizing it. [1], [free CrAIyon v3 2], [3], [4].
A competitor to DALL-E is Stability AI’s [1] text-to-image program Stable Diffusion [1], also available via DreamStudio interface [1]. Here are a few high-quality images created by it [1]. While DALL-E 2 has several filters at the API level to prevent problematic images from being generated, plus a watermark on images it generates. The open-source Stable Diffusion does not, meaning bad actors can utilize this technology for deep fake image creation and other malicious activities. OpenAI also released GLIDE [1], [2], [3] a scaled-down text-to-image model that rivals DALL-E performance.
Another popular AI image creation AI tool via a prompt is Midjourney [1], which is available from Discord servers [1]. This is a subscription-based service with three plans. Plan #1 Basic $8/Month; Plan #2 Standard Plan $24/Month and Plan #3 Pro from $48/Month.
4. ChatGPT inner workings:
In a chatbot the system understands input and output text as strings of “tokens,” which can be words and punctuation marks as a part of words. Behind the scenes, AI is constantly generating a mathematical function called a “probability distribution” to decide the next token a.k.a. word to output for us to read, considering all previously output tokens. GPT has made trillions of connections between words, and it is called tokens. This is how it digests data from its trained sources.
In the case of ChatGPT, after the distribution is generated, OpenAI server does the job of sampling a token according to its distribution and inserts randomness so that same question asked to ChatGPT the next time, has a slightly different response.
Context is important in any chatbot interaction. In your discussion with ChatGPT, you can go for follow-up questions, ChatGPT remembers your previous questions in your end device cache memory, from which it builds the discussion context. This is super valuable to users because we don’t have to explain everything to it from scratch with every question.
Watermarking AI generated text is a controversial subject with good arguments on both sides of this discussion. In research published by the German security institute CISPA (Helmholtz Center for Information Security) published a paper [1] in March 2021, titled “Adversarial Watermarking Transformer: Towards Tracing Text Provenance with Data Hiding.” This paper theoretically talks about how AI generated text could be watermarked.
OpenAI is said to be working on this theory utilizing cryptography-based approach to statistically watermark the outputs of their AI text. This process would embed an “unnoticeable secret signal” indicating where the text came from.
Jack Hessel, a research scientist at the Allen Institute for AI, pointed out that it’d be difficult to imperceptibly fingerprint AI-generated text because each token is a discrete choice.
4.1 Neural Network Role:
Success of chatbots goes to “neural networks” logic of ML, which are mathematical models with tweaked internal parameters to change what they output. A neural network is what teaches a chatbot at a prominent level what it needs to speak out based upon its model training.
A subset of machine learning is deep learning (DL), where neural networks are expanded into sprawling networks with many sizeable layers that are trained using massive amounts of data.
These deep neural networks have fueled the current leap forward in the ability of computers to carry out tasks like speech recognition and computer vision. There are several types of neural networks with different strengths and weaknesses, for example:
Recurrent Neural Networks (RNNs) are a type of neural networks well suited to Natural Language Processing (NLP), which aids in understanding the meaning of input or spoken text (a.k.a. speech recognition) to the machine. The other one is Convolutional Neural Networks (CNNs), which is designed for visual image analysis.
Together these AI techniques are used in robotics and autonomous car driving.
5. ChatGPT Use Cases:
With the enhancements of AI software and power of today’s hardware, plus the advancements in the quantum computing field, chatbots are ready and fully capable of aiding humans in enhancing their efficiency at work and removing the knowledge acquisition barriers.
Literarily there are use cases in every professional field for this technology. For example, creative writing is just becoming easier, and one can get ideas and contents for their topics within minutes. Then generate different sections and so on.
AI, because of its speed and agility, could be used for cybersecurity threat detection and even in incident responses. These days we get so much threat data, it will not be possible for humans to analyze the correlated data in real-time, identifying patterns and anomalies that would otherwise get missed by traditional security systems. End-to-end automation could be accomplished if we allow AI to make decisions and I guess people may not be comfortable doing that as of now. The speed that AI brings to the table in potential threat detection scenarios is what makes it the winner. The same power of AI can also be utilized in incident response and breach prevention effectively. For sure there will be products on this particular AI use case very soon.
Program code-generation or error debugging is another area already utilizing AI. For example, one can write a Docker file for Nodejs application or a python program using any library, for example, let us take TensorFlow. You can change all the variables in this question and ask ChatGPT in simple statements and you will get a response in the form of real programming code that you can put back in the right platform and your code will be ready in minutes rather than days or weeks. You may enhance this response by asking it to add more features. Of course, you may have to do small tweaking here and there per the uniqueness of your requirements and environment. GitHub Copilot [1], which is also an AI-powered OpenAI code generator behind this logic. Conclusion is AI generated code is fairly accurate and revolutionizing modern day programming.
Current DevOps teams can take full advantage of AI powered tools to assist in their regression, functional and user acceptance testing. With siloed data from a large organization, this kind of testing will be consistent and comprehensive across different departments.
There are multiple AI medical use cases [1], including one that can predict diseases before onset, based upon the genes, historical inside cell conditions, state of enzymatic reactions in correlation with other patient’s vital medical data.
2024 Update: I have a dedicated article on this titled “AI for Healthcare Professionals” and could be accessed from here.
ADA [1] is an AI-powered app that helps people understand, manage, and take care of their unique medical symptoms in minutes. ADA Application has access to a medical conditions dictionary, disorders and associated medical symptoms from which a diagnosis including personalized assessment report with probable causes and next steps is pulled, utilizing AI techniques. ADA application is available in both Google and Apple stores.
Babylon Health [1] provides 24/7 access to doctors, personalized care plans, seamless referrals to specialists, digital health tools, chronic disease management and more for its members.
Other important use cases are avoiding medical errors, effective disease diagnoses, assistance in prescription medication, chronic disease management, DNA Sequencing [1], [2] and so on. Elicit [1] AI Research Assistant is another example in this category.
There are use cases for highly skilled areas like architects, lawyers, doctors, engineers, scientists, and common users, to use this technology. You can search past court cases PACER (Public Access to Court Electronic Records) service [1] and few other Apps in Legal space making inroads with AI are Casetext [1], Legal Robot [1], DoNotPay [1], Latch — next generation contract assistant [1], Lawgeex — AI Contracts [1], NeLaw [1], LegAI [1], PatentPal [1].
With ChatGPT 4.0 you can search medical research papers in PubMed [1] repository for specific medical research topic between specific dates and you will get the exact findings. You may even ask ChatGPT to summarize one of these papers in two paragraphs or in the form of the top ten points and you have the answer in seconds. Elicit [1] is an AI Research Assistant that is free and is also doing the same.
ChatGPT AI power could very well be utilized for your documentation and manual creation purposes. Of course, students can get help with their assignments and drafting their papers. Writing Ph.D. thesis is now many folds easier and quicker. There are so many use cases, a book could be written on these.
6. Large Language Models (LLMs) Landscape:
Let us pay some attention to different LLMs [1], [2], [3] currently available, who is behind those and how their power is measured. LLM power measurement is done by the number of parameters a.k.a. tokens those models were trained on. Another way to measure the power of a Chatbot based upon a LLM is to measure its IQ. According to IHIQS (International High IQ Society) [1], [2] high IQ are those who have an IQ of 125 or above. An average human IQ is between 85 and 110. Below 70 is low and over 160 is genius. AGI is the term that refers to “Artificial General Intelligence”, which in AI parlance means AI is performing at or above the intelligence of a genius human capacity. The question is with all these AI around us now, have we reached AGI yet? Interesting question, I will leave that for you to decide.
⁂ OpenAI ChatGPT-3.5 around 117 million parameters, which is approx. 45 TB of text data. This model is designed to handle 175 billion parameters [1].
⁂ OpenAI ChatGPT-4 got released on March 14, 2023. it is an Embodied Multimodal Language Model (EMLM) [1] just like PaLM-E of Meta. EMLM is capable of processing text, numbers as well as images and videos. It is trained on 175-280 billion parameters [1], [2 sign-up for a private beta].
Note, the free version of ChatGPT is still on GPT 3.5 but the paid Plus version of ChatGPT has moved to GPT 4.
⁂ OpenAI ChatGPT-5 — Release Date kept a secret — Believed to be currently getting trained on NVIDIA GPUs. IQ expected to be in the genius range.
⁂ Microsoft New & improved Bing Search uses Prometheus [1], [2] LLM model, which is based on an earlier version of ChatGPT 4. Microsoft internal code name is Sydney [1] and it is capable of dynamically querying Internet to get current data. You can ask it about current date, or events and even ask it to multiply (sin(0.12345) x π) and it will use a scientific calculator to resolve complex equations and provide you with the answer. Bing IQ is evaluated to be 114.
⁂ Microsoft AI 365 Copilot [1], [2], [3] = Microsoft 365 Apps + Microsoft Graphs + Microsoft’s new LLM
Microsoft 365 Copilot combines the power of LLMs exclusive with your (company’s / private) data within your very own Microsoft tools (calendar, emails, chats, documents, meetings, etc.) thereby making Microsoft 365 Apps powerful productivity tools for us. Plus, there is nothing new to learn, as we are aware of Microsoft Office tools!
⁂ Meta’s LLaMA [1] reported be available in 7, 13, 33 and 65 billion parameters versions.
⁂ DeepMind’s Chinchilla 70 billion to 1.4 trillion parameters [1].
⁂ Google’s BERT (Bidirectional Encoder Representations from Transformers) is just like ChatGPT 3.5, but with much less power. It was trained on 345 million parameters [1], [2].
⁂ Google’s PaLM 540 to 800 billion parameters [1], [2], [3].
⁂ Google’s GLaM (Generalist Language Model) [1], [2].
⁂ Google’s Gopher trained on 280 billion parameters [1], [2], [3].
⁂ Clearview.AI is a LLM for Facial Recognition, believed to have 20 billion images indexed on Internet.
⁂ Stability.AI is another opensource LLM model with various products, most popular being DreamStudio.
⁂ WolframAlpha — Computational Intelligence — is a powerful LLM capable in the field of Mathematics and Physics related AI magic. More on this in the conclusion section.
⁂ Megatron-Turing NLG trained on 530 billion parameters considered to be most powerful generative monolithic transformer language model [1].
⁂ Humanloop [1] is a service build on GPT-3 to add value.
⁂ AI21Labs [1] offers a range of customizable Large Language Model APIs.
⁂ MLPaaS: Outerbounds [1] provides a production-grade ML infrastructure platform for your ML projects data and pipelines, capable of integrating with your systems and policies.
⁂ Adept [1] is a unique AI model that can interact with your computer components.
⁂ Taplio [1] is an AI program designed to manage your LinkedIn profile and your personal brand. Whereas TapIO [1] is the first Indian made AI and IoT solution to streamline business operations.
⁂ ML Model synthetic data provider:
※⁑ Synthesis AI [1].
※⁑ TONIC for fake AI Data [1].
※⁑ Datagen AI [1] provides synthetic Image Datasets for Computer Vision.
※⁑ AI Foundation [1] is extremely unique and provides synthetic ML data based upon your selected personal values and goals.
⁂ Opensource ML Models:
※⁑ LangChain is an opensource LLM, with which you can build your own powerful AI applications [1].
※⁑ Rasa is a conversational AI infrastructure available under both opensource [1] and commercial [1] (Rasa Pro) licensing.
※⁑ Hugging Face [1], [2] is a community that does ML from build, to train, to deploy.
Here is a list of top global players / companies eyeing for the AI Cybersecurity market share [1].
Apple’ Ferret [1]— Updated Feb 2, 2022
The future versions of Apple’s Siri, Messages and Apple Music will be using generative AI, according to Bloomberg’s Mark Gurman.
In Oct 2023 Apple released a Ferret, an Open Source MLLM (Multimodal Large Language Model), which is a Guided Image Editing LLM. Apple’s Ferret is a 7B MLLM Generative AI model (a.k.a. MGIE AI), comparable to or even better than GPT-4, as talked about in technical circles. Time will tell the truth.
MGIE AI a.k.a. Ferret can edit images based upon the user’s natural language instructions. (Steps: #1 Just upload ur image here. #2 In simple English write what change u want to make. E.g., “Add a USA flag colored hat”. Voilà, it will happen in 20 seconds.
This is FREE, Open Source and ready for use now! The other two popular AI models are Midjourney [1], [2] and DALL-E [1]. Both need considerable technical skills to use.
Google’s Gemini [1] is a trained AI model that is publicly available. It can communicate and generate human-like text in response to a wide range of context alive (real-time) questions. Below are four advantages it brings:
Multimodality: Can process and learn from data types like text, audio, and video, potentially leading to richer understanding and responses.
Real-time information access: Because it integrates with Google’s knowledge graph and Google internet searches, providing potentially more up-to-date responses.
Transparency: It offers information on its sources and reasoning behind responses, allowing for better trust and understanding between the users and the AI.
Image generation: It can generate creative text formats and images.
— End of Update!
6.1 Different OpenAI software manifestations.
⁑ OpenAI Codex is another AI model that parses natural language and generates computer code in response. That is the logic and brains behind GitHub Copilot, which is a programming autocompletion tool for your IDEs [1]. It is optimized for professional and armature programmers and access to it is available from GitHub.
⁑ OpenAI CLIP is a neural network able to classify images by given categories. It was trained on four hundred million Internet images with long descriptions. The prominent use case for CLIP is its prediction capability to identify the given “image class” from its past learning.
⁑ OpenAI Five neural networks collectively referred to as OpenAI Five in an AI venture of a computer video game Dota 2 against professional players. It has already played 180 years of human timeline experience, against itself for learning and training purpose. Dota’s annual tournament pool prize is few million dollars every year. For the current event date and prize money, refer here [1], [2], [3].
⁑ OpenAI MuseNet is a deep neural network developed through unsupervised learning using millions of MIDI files. It can generate unique four-minute musical compositions with ten different instruments, based on a specific genre or a composer’s style.
⁑ OpenAI Wisper is a ML model trained to transcribe speech to text and translate contents from different languages to English.
If you want to integrate your AI project with OpenAI powerful models you may look here.
7. Future of AI Chatbots:
Top upcoming AI conferences for 2023 around the world from Forbes pages [1]. AI is a general-purpose technology which uses big datasets to identify patterns and predict events based upon ML/DL algorithms. The 20th century bottleneck was the cost of collecting and storing big data. We have resolved this bottleneck. The 21st century challenge is how to timely utilize this big data that we already have and perpetually adding to it. How can we utilize this big data to provide meaningful results within a reasonable period? Quantum computing is the answer going forward. This is to the word, what ChatGPT is doing minus the continuous newly generated data-usage part.
AutoML a.k.a. automated machine learning which is the process of automating iterative ML model development.
This is where we will move, utilizing quantum computing soon to scale ChatGPT or similar upcoming chat bots to keep up to date with their training data. Once AI chatbots go real-time, imagine the new use cases that will be added to the already overwhelming use case list. ChatGPT 3.5 model learning date stops at September 2021, and it is blind after that date. ChatGPT 4 (a.k.a. Plus) is a different case, as it can query the Internet and use external tools to complement its work.
Quantum computing can execute complex functions quickly, by harnessing the collective properties of superposition, interference, and entanglement.
That definition is a little farfetched for the ordinary mind to grasp. Let us rephrase the quantum computing definition in a way we can digest, and that will be…
Quantum computing brings faster processing power to our data and makes our environment and networks secure.
With the quantum computing context in the backdrop, let us analyze today’s data creation use case. Tons of data created by each one of us, on our applications, and systems each day. We are having a perpetual never-ending and growing data inventory. That is exactly what an AI/ML needs to function. Quantum computing can establish a bridge between our perpetual data creation process and machine’s ability to grab that data in real time for learning and utilization purposes. When this bridge comes up, Generative AI/ML/DL model training can happen on real time data, which could make AI more effective and efficient in predictions and decision making.
8. ChatGPT best practices:
ChatGPT is a long-form question-answering AI application. Like any other application, it has some unsaid rules that one must follow for optimum utilization of this platform. The first and most important rule is to breakdown your complex problem into simple small segment/groups and then reveal it to the ChatGPT in several steps building upon the previous one. It is good to ask ChatGPT to follow you and specifically ask it to add to your previous point.
8.1. ChatGPT value add:
ChatGPT has knowledge from all fields of functional areas that it was trained in. A reading of this paper gives you a high-level idea of what those are and now imagines a subject matter expert (SME) in all those areas is at your fingertips. Just ask a question and within seconds you will have an answer. It could be your assistant in your research projects, and it can generate high-quality text output based upon the context and simple iterative questions you must be asking it for a response. It will save you tons of time.
This AI chatbot has the potential to change the way we interact with machines. The proof is in the stats of how quickly it has its organic growth. A positive outcome is that Google, Apple Sri, and Amazon Alexa will be forced to expand their scope to incorporate ChatGPT like behavior soon. This will absolutely be a win for the user community.
ChatGPT can do Google like search, but with refinement to a finite level. The rules of engagement are different here and different from normal search engines, which will scrape info from the web and throw some links for you. Here you will get an exact tailored match answer to your question. It all depends upon how you interact with it, which is directly proportional to the framing of your questions and the context you set.
As businesses utilize quantum computing and AI, it is most exciting to consider how propelling these technologies can serve humanity. AI can aid the development of new cures for diseases, detangling cybersecurity threat traffic, and could identify and protect sensitive data. A new trend is emerging, where ChatGPT is behind the influx of new AI-written books on Amazon [1] and has become an easy-way to make money.
8.2. ChatGPT limitations:
ChatGPT has its own share of limitations. At the time of authoring this paper, it is not capable of taking matrix or tables as input data. It may not provide the right data despite its learning sometimes. Reason is what you put in is what you get. Take any topic on the Internet and that topic data is filled with arguments from all sides. Meaning, ChatGPT was trained on non-factual data. At times on contents, it may be wrong or biased. It is not its fault. It is our fault that we have filled Internet with all kinds of nonfactual crazy ideas. So, it is best to use human ingenuity after you generate anything from any chatbot to wrap it in your experience. This is also a reason it cannot 100% replace any specific human job that involves ingenuity. Yes, it can replace monotonous and mechanical jobs, but where human touch and thinking is involved, it is still a long way to go.
We are still miles away from bringing AI machines to near human replacement, as tomorrow we can have a new experience, meet new people, and get innovative ideas. An AI machine cannot create its own ideas unless we write and put those experiences online. On the same token, it is not a clever idea to get advice from an AI machine. The same logic goes here, if you are a high-tech professional, you will not seek advice from those who are from a completely different era and have no knowledge about your field.
8.3. ChatGPT negative use cases:
In 2016 Microsoft released Tay AI, which was promptly shut down for spewing foul language. Technology is not associated with any specific use case, although at the time of its creation, one or two use cases were in the mind of its inventors. As quantum computing and AI evolves, it is important to understand that validation of data is equally important just like data analysis. Data weaponizing, corrupting analytics, and derailing the experiential learning could be few negative use cases where AI could be utilized for cyberterrorism. On the other hand, ChatGPT could very well be used by malicious actors [1], [2], [3] to write high-quality phishing email / campaigns [1] and in the creation of malware, or the school assignment cheating.
Problem Amplification is a major concern of AI. For example, let us take a non-AI environment. Some people will make mistakes and produce or believe or use wrong data. In this case, the damage from using bad data is restricted to a very narrow segment. Now let us get AI into the equation and it has been trained on wrong or misleading data. The repercussions are huge when AI is involved. A massive number of users using the AI results, will build their decisions / data / plans on something that is not logically correct. Plus, the multiplier effect will be… future AI may be trained on this wrong or misleading or faulty data. You can now visualize the amplification / compounding of the misleading data seep-in, of the future AI training and how perpetually bad this problem will become with each iteration.
Data leakage is another potential threat as it is noticed employees are feeding sensitive business data to ChatGPT [1]. All these and more are possible, but that is the other side of AI that we must learn to live with and build awareness and secure processes around. However, this could easily be solved by the utilization of AI LLMs like Microsoft 365 Copilot or comparable products that could utilize one’s own / personal data.
9. Developer Corner for ChatGPT / Whisper API:
This section is specially designed with referenced links for developers who want to integrate ChatGPT and Whisper API calls into their own applications. For OpenAI ChatGPT / Whisper API documentation refer here [1], [2 in here search for “Whisper API for cost”]. Worldwide developers can join the GPT-4 API waitlist here [1].
To interact with GPT APIs through HTTP post endpoint requests from any language, via OpenAI official Python bindings and their official Node.js library cloud be downloaded from here [1] and API commands refer here [1]. For examples on API referenced applications refer here [1].
There are multiple OpenAI models optimized for different use cases. Ada is the fastest model, while Davinci is the most powerful. More on these here [1 in here search for “InstructGPT”]. The pricing model for GPT-4 API integration referenced here [1]. Pricing is based upon 1K tokens [1 in here search for “Managing tokens”]. The max number of tokens ChatGPT can handle in your prompt is 4096 and your billing is proportional to the number of tokens used in your prompts.
This cost is approximately $0.002 / IK Tokens and changes based upon even punctuation used.
There is a library called “TikToken”, which can keep track of the use of tokens by your application. Installed via PyPi: pip install TikToken — The tokenizer API document in the TikToken/core.py [1], [2 for which embedded tokenizer to use]. There is also a JavaScript library [1] that can do the same.
It is worth noting that the above GPT API pricing model is different from ChatGPT 4 Plus general AI service access [1] model.
9.1 ChatML
On March 1st, 2023, OpenAI introduced ChatML [1], which is a document format just like JSON, consisting of a sequence of messages, where each message contains a header (which today consists of who said it, but in the future will contain other metadata) and contents (which today is a text payload, but in the future will contain other datatypes). This concept is like SAML, an open standard data exchange protocol for authentication (AuthN) and authorization (AuthZ) between parties (ID and Service Providers), which is an XML-based markup language for security assertions used for facilitating single sign-on (SSO). However, the use over here is data-exchange.
9.2 AI Laws and Regulations:
Although not much action is seen in North America on AI regulation, we can see some movement on the other side of the pond with the European Union’s (EU) AI Act (AIA) [1], [2], [3], [4].
US Copyright Office Rules for Generative AI was released on March 16th, 2023, and could be accessed from here [1], [2 same from Federal Register]. The “Register of Copyrights and Director of the U.S. Copyright Office” sign this doc. This could be looked as the first step in the direction of regulating the new space of Generative AI. This policy initiates steps to examine the copyright law and policy issues raised by AI technology. I bet this doc will be a part of coming copyright law cases.
In short, this policy states that copyright protection on Generative AI depends on whether the AI’s contributions are the result of mechanical reproduction, such as in response to text prompts, or if they reflect the author’s own mental conception. I know it is tricky to draw that line. For more, read the directive from the above links.
10. Summary & Conclusion:
Whether we like it or not, ChatGPT application is here and has reached millions of users. It has generated a hysteria among tech companies and user willingness to use it for the benefits it brings to the table. Programmers are using it for no code, low-code, and code-based fast quality programming.
Although ChatGPT is not the first publicly available AI application. That accolade must go to WolframAlpha [1], which is an AI knowledge computing engine. This model was trained on vast amounts of data related to world geography, economics, and name databases for facts. You can do unique mathematical computations with WolframAlpha, which will not be possible on ChatGPT 3.5, which is the basic version of it is freely available (however, GPT 4 is a different case). For many reasons it did not get the hype that ChatGPT got. You may use it and figure it out by yourself.
The free version of ChatGPT is running on GPT 3.5, which is the most popular large language model accessible by the public. It has a paid subscription model for USD $20 per month referred to as ChatGPT Plus [1] utilizing GPT 4. The next logical release of ChatGPT will be GPT-5 and the next release is expected to bring massive advancements. As we have seen the enhancements GPT-2 to 3 to 3.5 and 4.0. With its impressive performance and versatility, ChatGPT current version has the potential to revolutionize the way humans interact with machines and the way machines understand and process human language.
Today, AI has become an overarching umbrella architecture for any and every field, from cybersecurity (my area) to medicine, from finance to legal and all verticals of business and knowledge management faculties. You just name a subject and there is an application of AI in that area.
Responsible usage of AI chatbots is crucial to ensure that their impact is positive, lies on yours and my shoulder. The opportunity to lead the innovation curve is also ours. We need to work to enhance our laws and regulations to balance out this technology. Ask the people who have used it what it will be like not to use it in their daily work. It can speed up the work multifold and help you accomplish quality work in a fraction of the original time you otherwise would have taken.
ChatGPT is having a massive impact because of its ease-of-use and what it can do for users. Going forward, ChatGPT and similar AI technologies will be incorporated in the browser technology, and this has already started to happen. It will become a factor in a country’s economic variables in the future. Thereby becoming a foundation to economic growth. A huge population is already using it, and rest have tons of positive use cases to utilize it. Just like Internet and Smart Phones, this chatbot and other AI derived technologies soon be touching every ordinary individual with two goals. One to access knowledge from your fingertips and second to enhance your productivity multifold. When this materializes, the working pattern of the masses will change in a new direction. What that direction would be, we all will have to wait to see it.
About The Author
Asad Syed is a graduate of Mathematics, Applied Mathematics and Statistics. His experience spans in Security Architecture, Security Operation Management, Digital Investigations & Forensics, Crisis & Threat Simulation, GRC Management, Threat Hunting, Cybersecurity Emerging Trends & Threat Mitigation, Database Security, Identity & Access Management, and Identity Federation. His interests are in the application of newer technologies, to enhance the output performance of technologies with which he is working. He is a writer, teacher, and cybersecurity evangelist, who has worked for multiple fortune five hundred companies and currently providing cybersecurity consulting to the upstream operations of the oil and gas industry. Reach him via GRC at ASyed dot net. ■
