Separating the hype from the reality of ChatGPT

In 2015 a range of Silicon Valley thought leaders (including Sam Altman, Greg Brockman, Elon Musk, Reid Hoffman, and Peter Thiel) created the non-profit AI research laboratory OpenAI Inc. Then, in 2019, it created a “capped-profit” company called OpenAI LP. The mission statement of both companies is “to ensure that artificial general intelligence (AGI) benefits all of humanity, primarily by attempting to build safe AGI and share the benefits with the world”. Since 2020 OpenAI has publicly announced a series of AGI innovations, starting with the release of GPT-3 (which stands for Generative Pre-trained Transformer 3), an autoregressive language model that uses deep learning AI techniques to produce human-like text from an initial text prompt.

This was followed in 2021 by the release of the DALL-E deep learning models, based on GPT-3, by OpenAI for the purpose of generating digital images from natural language descriptions/prompts. The growth in popularity of digital art in 2021, led partly by the speculative investment nature surrounding NFTs (Non-fungible Tokens), and the continued broad reach of social network platforms, saw a spike in awareness of the pictures that DALL-E could produce by the public across the world in both 2021 and 2022. The barrier to entry for generating pictures, and as some debated art, with DALL-E was very low and did not require a person to have a technical background. All that is needed is an account and the purchase of credits used to generate pictures from any range of text descriptions a person wants to provide.

Most recently, in late 2022, OpenAI launched a free preview of its next AGI innovation ChatGPT based on its GPT-3.5 AGI engine that utilises a 175 billion parameter model to generate its responses. Since its launch, it has become a viral sensation and has gained wide public recognition globally due to its coverage across traditional, social, and online media outlets. OpenAI has been flooded with new account requests and is still unable to keep up with demand, including a reportedly surprising amount believed to be Russian-based hackers. ChatGPT uses a type of generative AI called a large language model (LLM) to create complex responses based on descriptions/prompt requests by humans. Many journalists and industry professionals have described it to be as important an innovation as Google search. However, PAC considers this an excellent example of how sophisticated LLM AGI models can be when they have access to vast petabytes of data but is cautious in several regards to its capabilities being unrealistically hyped beyond its current capabilities. In its current form, ChatGPT can produce a range of outputs from essays to poems, song lyrics, software code (including malware), to instructions on how to build things.

However, despite its ability to provide what seems to be highly detailed and sophisticated responses, it does have limitations behind its chat façade. PAC believes it is essential to understand that whilst to some ChatGPT may seem on the cusp of sentience, its 175 billion parameter model, petabytes of data, years of training, and the processing power of a supercomputer running 24/7 provides what seems as human-like intelligence. In reality, it is an incredibly innovative and sophisticated text predictor that has been ingesting publicly available data to drive the creation of complex responses. By highlighting this PAC is not attempting to demean or lessen the importance of ChatGPT but rather ensure it is clearly understood what it is and how it behaves. For example, it is currently limited by the range and type of data it has access to from across the internet, which means it is fallible, prone to bias, and has ingested enough data to contextually misunderstand how responses may impact an individual regarding challenging societal topics. However, it is important to understand that OpenAI has teams of employees checking the data it is ingesting and the outputs it provides to help adjust its output regarding bias and other sensitive topics.

ChatGPT is only as clever as the range of information it has ingested, which means it will have knowledge gaps that could cause issues in how intelligent its responses are perceived. The hype surrounding its release has raised awareness of this technology and shown its current sophistication. However, everyone should know that OpenAI is getting something precious in return. For every person using ChatGPT, OpenAI is receiving free AI training that further supports their ability to refine and raise the sophistication of the responses from the LLM, but the interactions could also introduce forms of bias. It takes many years to establish an LLM like ChatGPT and requires continued ingestion of data and a large amount of ongoing training.

Whilst PAC considers ChatGPT to be a great technology innovation, there is a range of concerns that must be understood, for example:

• Providers of LLM AGIs like ChatGPT need to state what ranges and types of data sets it has ingested. This limits ChatGPT, so its responses, for example, will seem sophisticated to people in the USA but less for those in other countries, especially as it is currently only English-based.
• The ethical and societal ramifications of how information is provided to humans by LLM AGIs like ChatGPT have yet to be fully considered and understood. PAC also emphasises that the ethical impact of technologies like this on society is typically determined by governments and varies considerably between them. Before the public use of LLM AGIs grows disproportionally to ubiquitous levels, governments must act to develop regulations for private companies to adhere to.
• The compute power required for LLM AGIs at the scale ChatGPT is operating today should be a concern for global, regional, and local sustainability goals. As more organisations and governments look to create and operate their own LLM AGIs, the negative sustainability impact has the potential to increase.
• Whether it is an AI-driven art/picture generator or chat-based content generator there remain questions regarding the legality and copyright of the information sourced by the AGI models. Many companies providing access to such AI models currently assert degrees of ownership, operate as a non-profit, and claim their technology is a tool, not a creator. The last point is particularly interesting from a legal standpoint because these technologies create outputs from an amalgam of inputs sourced from diffusion models based on publicly available data. However, these technologies need to consider copyright when generating a result. Otherwise, misrepresentation, misappropriation, and plagiarism could occur.

All that being said, PAC considers ChatGPT a great example of how sophisticated AGI models have become in a relatively short time. PAC expects the next horizon to be reached for this technology, after pictures and chat, is a voice-based version of ChatGPT that provides a conversational experience rather than text-based. Whether text or voice-based, these technologies offer a new way to interact with the repository of human knowledge, much like internet search did some decades ago. Aside from the levels of compute required, the only thing limiting these types of technology is the amount and variety of data it has access to. Over the coming years, PAC believes some governments and organisations will become wary of what and how much information is publicly available to ChatGPT, and similar technologies, at no cost and no advantage to the source that created the data. This could drive both islands of private data and forms of competition that ultimately limit how sophisticated a technology like ChatGPT can become.