An 80s American TV series called “Small Wonder” had everyone glued to their sofas during prime time. It was about a robot modeled as a 10-year-old girl who was adopted as a daughter by her maker and his family. The amusing conversation between “Vicki” — a morphed translation of Voice Input Child Identicant (V.I.C.I.) — and the family members as she tried to pick up human behavior through her super-powered learning system made for exciting TV watching.
Fast forward to 2022, and we have a chatbot that can translate and respond conversationally in almost any language on earth and is increasingly being seen as a breakthrough in artificial intelligence history.
US AI research lab OpenAI’s ChatGPT is fine-tuned from GPT-3.5, a part of the GPT-3 language model trained to produce text. GPT stands for Generative Pretrained Transformer. Such “transformer” models are sequence-to-sequence deep learning programs that can produce a sequence of text given an input sequence. ChatGPT was optimized for dialogue by using Reinforcement Learning with Human Feedback (RLHF), a method that uses human demonstrations to guide the model toward achieving human-like behavior.
As such, it is capable of producing articles, essays and even poetry in response to user inputs or prompts. ChatGPT uses generative artificial intelligence (AI) algorithms to create content, including audio, code, images, text, simulations and videos, from large data sets fed into it from the internet. AI-generated art models like DALL-E (a combination of surrealist artist Salvador Dalí and Pixar robot WALL-E) can create extraordinarily beautiful images on demand. Generative AI falls under the broad category of machine learning.
Further, the ChatGPT system is designed to maximize the similarity between outputs and the dataset it has been trained on. However, at this stage of development, ChatGPT makers maintain that “such outputs may be inaccurate, untruthful and otherwise misleading at times.”
When it comes to the truthfulness of the chatbot’s responses, its makers claim, “ChatGPT is not connected to the internet, and it can occasionally produce incorrect answers. It has limited knowledge of world events after 2021 and may also occasionally produce harmful instructions or biased content.”
Can It Be a Disruptive Technology?
In the world of AI, many large language models perform tasks ranging from the simplest to the most complicated. Large language models can be game-changer for productivity as they can access and process real-time information, tackle complex problems and get more done in less time. AI can be used to create prediction/planning models and also curate massive data sets. For instance, the GPT-3 model was used by a UK theater group to write a play. The system generated a story based on the inputs of the writers. The story was further edited before the final version of the narrative was ready for the play. In another instance, the US news agency The Guardian used the GPT-3 model to write eight different articles, which were later compiled into one.
In a recent development, Chinese researchers from the Institute of Oceanology, the Chinese Academy of Sciences and Nanjing University of Information Science & Technology built an AI inference and prediction system for the Indonesian Throughflow (ITF), one of the largest movements of water on the planet, that can make valid ocean current predictions seven months in advance.
The researchers used a convolutional neural network (CNN or ConvNet), a network architecture that can learn directly from data through deep learning methodology. CNNs are used for detecting patterns in images to identify objects, classes and categories. They are also effective for classifying audio, time series and signal data.
The researchers used sea surface heights between the Indian and Pacific Ocean basins to design their AI model and trained the model with oceanic data sets. According to the researchers, the new system reported in the journal Frontiers in Marine Science could provide a new tool for studying ocean circulation and climate change in the Indo-Pacific Ocean and simplify real-time oceanographic observation. Similar data crunching and analysis using generative AI will likely follow suit in other industrial sectors, potentially changing market dynamics.
How Is Generative AI Different From Search Engines?
At a basic level, the range of memory is perhaps the biggest differentiator between the two. Chatbots can retain data for an extensive length of time, whereas search engines typically have short-term memory. Chatbot outputs can be comprehensive and specific as the bot analyzes the query inputs based on the entire historical data that it has been trained on. Search engine software scans its index of web pages to find relevant responses to the user's query. The results are ranked by relevancy and displayed to the user in the form of links to web pages, images, videos, infographics, articles, research papers and other types of files based on previous searches. Furthermore, chatbots depend on AI neural networks, NLP, audio, video and media processing to interact with users. Essentially, chatbots are function-specific, while search engines have a wide range of indexed information.
Technically speaking, search engines generate revenue through performance-based advertising. So, search engines like Google get paid by advertisers if their links are clicked. However, chatbot results have no links but are fairly comprehensive, so instead of checking the prompted links one by one, searchers can get the details in one post. It would be interesting to observe if the consumer trend might shift toward a chatbot response rather than search engine results, which could give search engine businesses a run for their money. However, the latest GPT3 model has been trained on data up to the end of 2021, which will potentially limit its capacity to identify emerging trends and, as a result, generate a bland output.
Ban on AI-Generated Content
The superior capacity of generative AI to create content in real-time has professionals in some sectors rethinking their career prospects. Teachers and academics are wary about the quality of written assignments and have already started banning the use of AI-generated applications on their campuses and universities. They feel that the bots can be used to plagiarize essays, which could be hard to detect for invigilators who are pressed for time. Other concerns related to AI-modeled texts are that they could even influence public opinion based on what data has been fed. In some instances, racially biased information has been generated, potentially in violation of the ethical code of conduct of the specific location. Another argument against AI-generated content is its ability to curtail human brain power for critical thinking. Proponents stress that such a technology can render humans lazy rather than productive.
Investments Pouring In
Despite a mixed response related to generative AI from various quarters, massive investment in the technology from both venture capitalists and tech giants such as Google, Microsoft and Amazon is painting a different story. According to Pitchbook, venture capitalists have increased investment in Generative AI by 425% since 2020, to $2.1 billion. Furthermore, rather than spending a hefty amount of capital to clean and compute big data, the introduction of APIs and open-source tools is allowing entrepreneurs to develop existing foundation models using generative AI and fine-tune specific verticals like gaming, graphic design, marketing materials, media, entertainment, etc.
It is not hard to see that generative AI has the potential to streamline a lot of business activity through automating patent writing, generating drafts of marketing materials or computer codes and generally optimizing the virtual world experience through new use cases. And the sophistication will likely increase as more and more inputs are reviewed and fine-tuned over time.
Conversely, the United States, the European Union and some additional countries are grappling with how to regulate the use of technologies such as biometric data, facial recognition and artificial intelligence to prevent racial bias or errors in data input. Ultimately, in all cases, the evaluation of data remains a key factor for the meaningful and purposeful use of these technologies. As of now, there is no alternative for humans to qualify for this critical responsibility.
