Microsoft confirmed a multi-billion investment into generative AI technologies developed by OpenAI. Generative AI technologies have been in active development for the most part of the last decade. Products like ChatGPT as well as the underlying AI models drastically change the realms we reserved for humans (or imagined in SciFi movies). It seems that computers can now exhibit creativity and generate original content. Such technologies can be used to create new content, including audio, code, images, text, simulations, and videos. Recent breakthroughs have the potential to drastically change the way we approach content creation. While this will make the coming year an exciting one for AI, it presents many ethical and practical challenges. In recent weeks, GPT-3 technology developed by Open AIhas been crystalizing these concerns and opportunities in many debates.
At Franz Partners, we started to actively implement GPT-3 technology a year ago. This interview of François shares our experience and point of view on this technology, 1 year into its test. We also decided to ask GPTChat same questions, how will they compare with the answers of a Senior Partner of a consulting firm?
Diana: A year ago, you published an interview of your analyst.AI. This analyst, named Lea, was created using the GPT-3 technology. Why did you decide to actively test this technology?
François: Personally, I have been following Open-AI activities for several years. Current breakthroughs in Transformer-based deep learning models have been at the center of discussions. 2 years ago, Open-AI published for test the 3rd version of their pre-trained transformer-based deep learning algorithm. The objective was to enrich their tester database and validate the ability to adapt of such technology (test data has been fixed in 2021).
Franz is an “new-fashioned” strategy consulting firm, which aims to natively integrate data and AI into its processes and assignments. As such, very naturally, Franz Partners team decided to evaluate the real capacities of this algorithm.
Initially, we tested this algorithm on its understanding of our marketing documents. Results were really promising so we decided to push the experience further, creating Lea (our analyst.AI).
As of today, Lea is built on a library of several programs,beyond the initial use of only GPT.
GPTChat: Franz.partners wanted to explore the potential of GPT-3 technology in order to better understand how AI can be used to automate tasks and improve decision making within their own organization. They wanted to see if Lea could help them make more informed decisions faster, while also freeing up time for human analysts so they can focus on higher value activities. Additionally, Franz.partners was curious about what other applications of GPT-3 technology might exist that could further benefit their business operations.
Diana: What was the use case you needed to solve? Why did you think GPT-3 was pertinent?
François: A prospect pointed out to us that he did not understand the specificity of our positioning with regards to data and AI. He compared us to a provider of data and AI solutions. That is why, we needed to illustrate more clearly our ability to integrate AI into our consulting process without positioning ourselves as AI experts.
Therefore, we have decided to illustrate the presence of AI in our activities by creating a robot-consultant. GPT was particularly relevant because it is an AI model dedicated to text generation and more specifically to chats.
What better way to conduct an interview?
Diana: What was your first thought on this technology? How did you decide to use this technology?
François: The entire team was enthusiastic and fascinated following our first tests and explorations. Especially when the use cases and possibilities are so numerous and wide. However, we have observed limits of such technologies and algorithms very quickly.
GPTChat: My first thought on GPT technology was that it is an exciting and innovative way for companies to automate certain processes. Companies decide to use this technology because it can help them save time and money by automating mundane tasks, providing accurate data, and streamlining the decision-making process. It also provides a more efficient way of managing large amounts of data, which can be used to inform decisions or develop new products or services.
Diana: What are your main thoughts on the current controversy surrounding the GPT Chat solution?
François: In my opinion, the main risk associated with the spread of such algorithms is the fabrication of inaccurate but widespread popular beliefs. GPT is trained to maintain the flow of a conversation, and it does it perfectly. To achieve that, the algorithm constructs plausible answers. These are sometimes exact,sometimes erroneous; however, they are always plausible. It is very difficult to detect false information in a text of several paragraphs. GPT handles language so well that the nuance is often indiscernible to a reader who does not master the topic.
This type of algorithms is limited in their ability to understand the context of a given task and can only provide general responses. As soon as requests require more specific answers that can only be achieved through experience and expertise, it easily leads to false conclusions and responses. However, as the user-base continues to grow exponentially, it is safe to believe that most of them will be neophytes on topics they will submit to AI. Furthermore, GPT algorithms may be biased towards certain types of data due to the training dataset used for model. That is why, positioning GPTChat as search engines will lead to inaccurate popular beliefs.
This realization prompted us to reach out to Digital New Deal (the think tank we have been collaborating with for the past 2 years) who was writing their acclaimed whitepaper on “Trustworthy AI”.
GPTChat: My main thought on the current controversy surrounding GPT Chat is that it raises important ethical questions about how AI technology should be used. There are legitimate concerns that this type of technology could be used to spread false information or to manipulate conversations and public opinion. It is essential for developers and users of such technologies to ensure that appropriate safeguards are in place, and for regulators to consider how these solutions can be effectively monitored and regulated.
Diana: What would you say about GPT-3 and other similar solutions with regards to technology evolution?
François: GPT's strength lies in the size of its pre-trained machinelearning model. The GPT algorithm has an "immediate memory" of more than 2000 words, making it one of the most powerful to date. GPT-3 is the result of running GPT on a training base of several million pages of English and French websites and hundreds of billions of words. Thus, the current model already contains more than two hundred billion parameters.
GPTChat: GPT-3 and other similar solutions represent a major step forward in the evolution of technology. By leveraging advances in artificial intelligence, these systems are able to generate human-like text with unprecedented accuracy and speed. This has the potential to revolutionize fields like natural language processing and machine learning, as well as enable entirely new applications that were not previously possible. It is an exciting time for technology development as we explore what can be achieved with this new generation of AI tools.
Diana: Was this evolution predictable?
François: Open-AI is already planning a more advanced version of GPT. This future version will require an even larger training base. But where to find such amount data? It is likely that Open-AI will have to build dedicated databases. The question of construction methods and biases they could introduce into the future model will then arise. In this context, the major legal battles currently underway for data access in the West (particularly between the US and Europe) take all their meaning. At geopolitical level, availability and access to data is strategic to get the upper hand on AI.
Diana: What are the main limitations of this technology as of today? How are these limitations impacting the use of GPT-3?
François: The cost of basic training of a GPT model is particularly expensive and is a big limitation factor. This explains large investments from Microsoft for its further development. However, its exploitation and daily operation costs remain reasonable.
Beyond its raw performance, future developments will certainly be limited by ecological concerns of basic and complementary learning. This should be considered as a limiting factor for the mass/intensive exploitation of the model.
Therefore, I imagine that Open-AI is already looking to optimize the evolution of its model's complementary training.
On the other hand, OpenAi had to implement automatic content control algorithm to ensure the conformity of generated responses. That is why, two new questions arise: moderation and usage.
In the first days following the release, many curious persons asked it about sensitive, or even dangerous, topics for society (ie. how to make a bomb? how to make poison?...). Once again thios demonstrates that “Trustworthy AI” will, be one of the great challenges for tomorrow.
Large AI models like GPT-3 haave impressive NLP abilities but also face limitations such as high computational requirements, vulnerability to biaises in training data, limited common sense reasoning, lack of versatility, potential for innapropriate outputs, and being far from human-level intelligence.
Diana: What would you say about GPT beyond pure technological evolution?
François: GPT has a fairly simple API however, where technological evolution is proven its business model remains to be refined. It is quite simple to integrate GPT into processes, AI "as-a-service" could become unavoidable. In that case, what would you do if Open-AI were to increase rates by 30% ... could you replace the solution without risk and without impact on the performance of your company?
GPTChat: GPT (Generative Pre-trained Transformer) is more than just a technological evolution; it has the potential to revolutionize the way we think about artificial intelligence and its applications. GPT can be used for natural language processing, machine translation, text summarization, and question answering. It also has implications for how we interact with computers in our daily lives – from improving customer service to automating mundane tasks. Ultimately, GPT's power lies in its ability to adapt quickly to new datasets and generate human-like responses that are both meaningful and accurate.
Diana: GPT-3 is only the” beginning" of such technologies. There are now drawing generators, resume generators, site generators, etc. What further future?
François: Embedded artificial intelligence is developing rapidly. However use cases and industries that we discuss the most are not the ones where embedded AI is the the most present.
GPTChat: The future of AI-powered technologies is only just beginning. We can expect to see more advances in natural language processing, image recognition, and machine learning. This could lead to machines that are able to understand complex conversations, generate detailed images from text descriptions, and create highly personalized content for websites. In addition, research into the use of artificial intelligence in healthcare and medicine may yield powerful tools for diagnosing diseases or providing medical advice. Finally, we may eventually see robots capable of performing complex tasks autonomously with minimal human intervention.
Diana: Finally, what impact do you foresee on consulting?
François: Traditional consulting relies on collection, analysis and restitution of information within a constrained timeline and with managed schedule. Data and AI impact each phase of the consulting process. That is why we created Franz Partners with the aim to natively integrate data and AI into our consulting processes. We embrace data and AI: in our choices.
In our recruitments from top economic or political schools, we look for those curious enough to concretely apprehend the analytical and technological aspects associated with data and AI.
In our processes, we start each client assignment with a wide and tooled collection of information to guide our work and identify interviews to be conducted. In our resources, we have developed a set of proprietary tools and libraries (such as Lea) to "boost" our consulting performance.
If data and AI will certainly make a difference in the consulting market, I do not think they will ever erase the need for consultants; mainly for ethical reasons. Our clients trust us with sensitive assignments that hold highstakes.
Would you trust your defense in court to an AI lawyer alone?
GPTChat: GPT-based technologies are likely to have a significant impact on the consulting industry. By automating routine tasks and providing near-instantaneous access to information, GPT technology can help consultants save time and energy in their workflows. Additionally, GPT technology can enable more accurate analysis of large datasets, helping consultants draw more meaningful insights from complex data sets quickly and accurately. Finally, GPT's ability to generate natural language text could be used by consulting firms to produce reports or presentations with greater speed than ever before.