Interactive AI Exposed: Real-time LLMs’ actual capabilities

Last year, generative AI went mainstream when OpenAI’s ChatGPT was introduced to the public. The chatbot instantly became a success inspiring widespread conversations and interest into the possibilities of AI and the different ways that it can be used for good. Since then, several generative AI applications and platforms such as Stable Diffusion, DALl-E, and Bard among others, have seen similar success with the general public. As with every technology before it, while generative AI is currently on everyone’s radar, it might not be for long. Researchers in companies like Deepmind, Google, and Amazon are already working on the next iteration of AI, interactive AI.

While generative AI requires a user to act as a form of partner, inputting questions and prompts from which the model forms its answers, interactive AI is expected to be more hands off. The newest ChatGPT features are an example of the early stages of interactive AI. The addition of voice and video components allows users to prompt the model by speaking aloud or uploading pictures in addition to text prompts. This move removes a barrier that is automatically in place when sending prompts through text allowing users to talk to the model as they would another human.

The Evolution of Interactive AI 

Engineers, scientists, and pop culture fans have been discussing the possibilities of a more interactive AI with anthropomorphic capabilities in fiction and real life for almost as long as science fiction and technology has existed. The first mention of the idea of what we now know as interactive AI was in Samuel Butler’s 1872 novel, Erewhon but the first real life steps towards actually creating one was not taken until 1950 when Alan Turin published his now famous paper, Computing Machinery and Intelligence, amidst a growing interest among scientists in the United Kingdom in the possibility of “machine intelligence.” 

Getting to where we are from there has taken billions of dollars and hundreds of scientists. Marvin Minsky and Dean Edmonds developed the first artificial neural network in 1951 which laid the foundation for the creation of STUDENT, an early natural language processing program developed by Daniel Bobrow, a doctoral candidate at MIT. In 1966, Joseph Weizenbaum created Eliza, a chatbot that was able to engage in conversations and was considered a feat at the time. 

By the 2000s, the idea of using GPUs to train large neural networks was being floated around as more effort and money was put into Artificial Intelligence research. Work had also started on the ImageNet visual database, one of the motivators for the current AI boom. In 2011, Apple released Siri, a voice-powered digital assistant that used voice queries to answer questions. In 2018, OpenAI released GPT (Generative Pre-trained Transformer) and the following year, Microsoft launched the Turing Natural Language Generation generative language model. Fast forward to 2023, we now have ChatGPT, DALL-E as well as several other generative models with the promise of better, more interactive models on the way.

How it would work

For as long as there have been talks about interactive AI, there has also been debates about what it would look like. At its core, Interactive AI refers to systems that can dynamically engage in human-like conversations and prompts enabling users to have more natural and instinctive interactions with the machine. This requires real-time feedback including the ability to adapt to the user’s preferences and tailor its output to the specific user. To achieve this, interactive AI systems are built using a diverse set of features including;

1. Sentiment Analysis - Sentiment analysis is a technique used to determine the overall mood or sentiment of a particular text. It is useful in determining the underlying meaning behind a particular input.

2. Feedback Mechanism - This is a loop that allows a model to learn from real time feedback, making it better and more personalized over time.

3. Ensemble learning - Ensemble learning is the process of combining predictions from multiple models in order to get a more accurate predictive performance.

4. Natural Language Processing - This is the field of artificial intelligence that is concerned with teaching models how to understand and process human speech. This is important when building interactive AI models because understanding human language is key in ensuring seamless interactions.

5. Deep Learning - Deep learning enables the system to learn from huge amounts of data using multiple layers.

6. Multimodal Interaction - This helps users to interact with the machine in various ways. With ChatGPT for example, users are able to interact with the model through text and speech.

Apart from chatbots, the potential applications of interactive AI are endless. It would mean better and more personalized virtual assistants that are able to adapt to the user’s preferences quickly. It could also be a gamechanger in areas like education and learning especially for people with learning disabilities who learn in unconventional ways. Interactive AI could also be used in healthcare to provide personalized care and advice for people with different lifestyles and upbringing. 

The ethics of interactive AI

The possibility of interactive AI is very exciting making it easy to overlook some very real ethical concerns. As with all artificial intelligence systems, there is a responsibility to make sure that interactive AI systems are as ethical and safe as they can be. This means that there should be some degree of transparency and communication about the limitations of these models and their possible impact. This would help ensure that users are knowledgeable and able to make an informed decision when deciding to use them. This requires building trust over time and putting processes in place to make sure that these systems are reliable and can be trusted.

There is also the issue of data privacy and security. Since interactive AI requires (and stores) large amounts of data, there needs to be an answer to the question of how that data is collected and stored. To prevent data breaches, there should be safety guards in place to protect the privacy of users and handling of the data should be in compliance with rules and regulations. Users should also be informed beforehand if their data can be used for surveillance and monitoring purposes such as with law enforcement. 

There is also a possibility that an interactive AI system may be built using biased data leading to  bias and discrimination in the output of the system. This can be tackled by ensuring that a diverse range of training datasets is used and harm prevention guidelines and oversight processes are put in place to limit unwanted harm by the system.

Conclusion

Interactive AI seems like the next logical step in the evolution of artificial intelligence bringing us one step closer to the still far off artificial general intelligence. While this could be a force for good, it is important to make sure that we are building these models as safely as we can for the user. This means putting in place strict guidelines and making sure that the process is transparent and ethical. With all these in place, interactive AI is poised to change the way we interact with machines forever.