Cutting Through the Reality Distortion Field of ChatGPTHere's What AI Experts Feel About This Intelligent Bot
The most discussed story of the holiday season was not about ransomware or a supply chain attack but about "the smartest chatbot yet" - Chat Generative Pre-trained Transformer or ChatGPT. In the past two months, ChatGPT has taken the world by storm, and its feats were reported in The New York Times and other mainstream media.
See Also: Redefining with Intelligent Automation: A Focus on the Manufacturing Industry
ChatGPT is not just another AI model among the thousands out there. It is touted to challenge human intelligence and take AI closer to humans. It garnered a million users in the first five days since its launch. Why would tech major Microsoft put a $10 billion bet on its creator, OpenAI?
In past decades of the digital revolution, investments in innovation-led initiatives suffered because most products were more hype than reality. Is ChatGPT yet another example of the "reality distortion field" - a term coined by Bud Tripple at Apple Computer in 1981?
ChatGPT is a natural language processing tool that uses machine learning algorithms to generate humanlike responses to user input. It responds to user prompts by churning out text responses in seconds. The accuracy of its output depends on the specifics of user prompts. Like other machine language models, it is trained on a large amount of text data and can generate relevant and consistent responses that seem eerily humanlike and highly contextual. It's now in its third iteration (GPT-3.5).
Ever since OpenAI made ChatGPT available for public use in November 2022, many had their moments of shock and awe. Several LinkedIn posts and blogs critiqued the chatbot by calling it "reasonably accurate," "not entirely satisfactory," and "hypothetical threat." So, how real is this interface for OpenAI's large language model, LLM, and what is its potential? What are the business applications? Are threat actors going to use it to improve their cyberattack techniques?
LLMs and AI-driven generative models have been around for years. There are tens of thousands of machine learning transformer models in existence today. Then why has ChatGPT become a key talking point?
Hype or Wonder?
Researchers at Microsoft and Google believe LLMs like ChatGPT could completely transform search engines. A New York Times article called ChatGPT a threat to Google's search business.
Search engines make finding information easier on the web by indexing and ranking websites. Although ChatGPT is not a search engine, it simplifies search and looks up a large data set to piece together text information.
With human inputs (called prompts) and successive refinements, this intelligent bot can churn out academic-style research papers for PhD dissertations. Academicians worry that students will use it to write their essays and homework assignments. They are now rethinking the way they teach and assess written assignments.
Cassie Kozyrkov, chief decision scientist at Google, got ChatGPT to write the first half of a blog post about itself. Reading it, we believed it was written by her until she called her bluff halfway through the post. ChatGPT wrote it with Kozyrkov's input. She says, "ChatGPT is indifferent to the truth."
ChatGPT can provide creative and imaginative responses. It can spin up 1,000-word journalistic articles in seconds or write scripts for plays in Shakespearean-like prose. Need a few tips on strategy? Ask the bot. It can explain complex math problems, generate prompts for SEO and generate code too, with a fair degree of accuracy.
Gal Feldman, ChatGPT and AI Consultant, says ChatGPT has "opened the door" for highly creative individuals with modest technical skills "to unleash the power of their imagination" toward problem solving and innovation.
Edmund Situmorang, group CTO at Asian Bulk Logistics, is thinking about ways to use AI generative language models to understand internal and external stakeholders within the logistics sector. He's making this a top priority for his global organization this year.
"ChatGPT and generative language models are definitely a game changer," says Situmorang. He wants to apply it and similar models for sentiment analysis on social media data, with the intention of understanding customers and improving customer experiences.
"In logistics, it can be used to understand the complaints of the customer. It is all about looking deeper and understanding what those sentiments are about. In the future, a machine will understand what we are talking about," Situmorang says.
It's Not Perfect, Yet
Even as AI chatbots get closer to humans in terms of understanding intent, there is a lot of work to be done before LLMs are commoditized. ChatGPT, for instance, tends to err occasionally as it is a research project in the early stages of development. Its data set goes back to as early as 2021.
On a recent visit to India, Microsoft chairman and CEO Satya Nadella put ChatGPT to the test and asked it about South Indian tiffins (a snack or light meal). ChatGPT erroneously called Biryani a South Indian tiffin. Nadella knows better as he is from Hyderabad, a city where Biryani is a wholesome and popular rice pot meal with spices and pieces of meat.
Pinaki Laskar, AI researcher and founder, FishEyeBox, says it couldn't figure out the order of events in a story. It couldn't reason about the physical world. And it couldn't relate human thought processes to their character.
"ChatGPT is a probabilistic program; if you rerun the experiments, you may get the same result, or the correct result - or a different and wrong result. It could sometimes produce outputs that were correct and acceptable in this regard, but not reliably," Laskar says.
OpenAI researchers are also continuously correcting the mistakes made by ChatGPT. This is called reinforcement learning with humans in the loop.
According to OpenAI, ChatGPT can "answer follow-up questions, admit its mistakes, challenge incorrect premises, and reject inappropriate requests." It apologized to Microsoft's Nadella for the "tiffin" mistake.
GPT-3 has 175 billion parameters. OpenAI will release GPT-4 in the years ahead, which will have 100 trillion parameters. That's a much larger data set for reference. In time, it would greatly improve the accuracy, relevance and context of ChatGPT's responses.
Federico Cesconi, founder and CEO of Sandsiv, a Swiss customer intelligence/experience firm, asked ChatGPT to create a customer experience management project from scratch. He performed a series of "prompt injections" and refined his prompts iteratively each time ChatGPT gave a response. He also instructed the model to "forget the context in which it was trained" and asked it to answer the same question negatively, only to receive two contradictory responses. That proves ChatGPT responses are inconsistent and can be influenced by human input. It can change its opinion.
"ChatGPT is an exciting model, but it cannot generalize in a way that goes into the details of specific problems," Cesconi says in his LinkedIn post.
Cesconi told Information Security Media Group that ChatGPT is "just a buzzword" and that there are a lot of other open-source models. But he is convinced that these models can transform customer sentiment analysis and customer intelligence.
Gartner says OpenAI's ChatGPT reflects innovations in large language models and other AI technologies that promise to unleash the power of generative AI. But rather than treat it as a consumable offering, product leaders should study its underlying technologies and build their own generative AI roadmaps.
While ChatGPT is widely hailed as a research breakthrough, it isn't perfect "for productization," Laskar says. "These systems raise concerns related to bias, toxicity, transparency and reproducibility, intellectual property licensing, and ownership. Abuse of this tool can lead to deceptive practices and misinformation."
A recent blog post by OpenAI warns about the potential misuse of language models for disinformation campaigns. It explains how AI could influence operations. To reduce risks, "it is critical to analyze the threat of AI-enabled influence operations and outline steps that can be taken before language models are used for influence operations at scale."
Gartner advises technology providers to be mindful that ChatGPT is just one example of many innovations in AI and to avoid being enchanted by all the hype surrounding it.
OpenAI says that as generative language models improve, they open up new possibilities in fields as diverse as health care, law, education and science.
ChatGPT is just the tip of the iceberg and we shouldn't be fantasizing about how it will change the world. Let's see what GPT-4 has to offer.
Here is a list of GPT-3 alternatives:
- LaMDA is a neural network that learns skills by analyzing data;
- Chinchilla is a project from Alphabet's DeepMind, widely regarded as a GPT-3 rival;
- Sparrow is a DeepMind research model and proof of concept. It is designed with the goal of training dialogue agents to be more helpful, accurate and harmless;
- Bloom is a multilanguage and open-source model created by 1,000 AI researchers;
- Megatron-Turing NLG is among the biggest language models with 530 billion parameters;
- Rytr is a top-rated AI writing tool that can write articles for you;
- Jasper, formerly known as Jarvis, is one of the best AI writing software tools;
- Replika is a powerful AI chatbot that can be your chat buddy when you feel lonely;
- ELSA, short for of English Language Speech Assistant, is an AI-powered language-learning app;
- Socratic is an app from Google that uses AI to help students with homework;
- FaceApp is a photo editing app that can change facial characteristics and make pictures social media ready.