In 2013 a science fiction movie called Her captured the public imagination. The film followed Theodore Twombly, played by Joaquin Phoenix, a man who develops a romantic relationship with Samantha, played by Scarlett Johannson, an artificially intelligent virtual assistant interacting through a female voice.
In May of this year OpenAI, the creators of the well-known AI program ChatGPT demonstrated functionality which mimicked the female character from Her. The program responded intelligently with a range of emotions and a voice that sounded a lot like Johannson (in fact she threatened to sue OpenAI). This is one example of how far AI has come in a few short years. It has been less than two years since ChatGPT sprang into the public consciousness and already there are many new capabilities that were difficult to imagine in 2022. The AI revolution has arrived!
ChatGPT is based on a Large Language Model (LLM) and the field is called Generative AI. An LLM is trained on a massive library of text which is used to teach the chatbot about language. The tool learns patterns, grammar, facts and context from this library. An important technical term to be familiar with is the “context window”. Currently ChatGPT has a 128k context window. This means that it can process books of up to 300+ pages at one time and create summaries and analyses of that content.
Today many LLMs are “multimodal”. This means they can process audio and have vision capability in addition to text. Here is an example: I asked ChatGPT to describe the picture on the cover of the OLLI-USF Summer 2024 catalog. This is what it said:
“The cover of this document features an image of two women exploring Gainesville’s Rainforest Butterfly Garden. The women, identified as Diane White and Nancy Shivers, are part of OLLI’s Outdoor Special Interest Group.
The photograph, taken by Mike Hayes, captures them amidst lush greenery and vibrant butterflies, highlighting the engaging and adventurous spirit of the OLLI Community.”
Note: ChatGPT read the names off the image, but the rest Is original. It had to analyze and understand the image.
In the latter part of 2022 when ChatGPT was announced, there were a limited number of LLMs. Today there are many thousands of these models! Most of the largest tech companies have them – e.g., Google has Gemini, Meta (Facebook) has Llama. Some tech companies are partnering with a company that has one (e.g., Microsoft and Apple are using OpenAI’s model although these companies also have some internal development). Some of the recent models can run on a home computer or laptop. The ones developed by Meta and several from Google can run on a home machine and are as powerful as earlier versions of ChatGPT. As of this writing (July 2024), Meta just announced a new model that rivals the latest version of ChatGPT. There are many more like these and new ones are being announced every week. Fairly soon, versions of these models will be running on our phones.
I would be diminishing the importance of Generative AI if I implied that its only function was to create chatbots. Here are some other applications and recent trends:
• Content Creation: LLMs can help write articles, stories, and even poetry.
• Language Translation: LLMs can help bridge language barriers. ChatGPT supports over 50 languages.
• Code Generation: Some LLMs can write computer code, making programming accessible to non-experts. For example, I asked one of these models to create a Tetris game using plain English and it did so in seconds, generating several hundred lines of code. The game was fully functional.
• Educational Tools: LLMs can summarize documents for study, create questions for class participation and even role play as a means of instruction. Textbooks generated in part by LLMs are rolling out to over one thousand universities this fall.
• AI Agents: These are LLMs that can autonomously make decisions and work cooperatively to achieve a goal. Together they can achieve better results than any single model. The vision is that eventually these agents can execute a large array of tasks, functioning like a capable human assistant. AI Agents are in an early stage of development.
• Application Integration: Currently Google has integrated its Gemini model into Gmail. I can access Gemini on the same screen as my mail and ask it relevant questions, (e.g. show me the status of my latest orders from Amazon). It can also summarize long emails. Another example: ChatGPT has integrated Google Drive and Microsoft OneDrive into its application, and it can access documents from those sources.
AI text-to-image generators have been around for some time, and the image quality has steadily increased to the point where many of the images are indistinguishable from high-quality photos as can be seen with the AI-generated image on the right. An emerging trend in this area is text-to-video. OpenAI has a tool called Sora which can create high quality short videos. Soon it may be possible to make commercial movies with this technique. No wonder Hollywood is concerned!
Most of us remember Data from Star Trek: The Next Generation who was a humanoid robot (a robot shaped as a human). The Generative AI revolution has accelerated the development of these robots and is enhancing their capabilities by improving their ability to understand and interact with humans naturally. This synergy is pushing robots closer to general purpose functionality, where they can adapt to various tasks in different environments. The financial firm Morgan Stanley has estimated that there will be 8 million humanoid robots by 2040. Elon Musk announced this July that he will be selling his Optimus humanoid robot to the public in 2026. His robots are expected to have conversational capabilities.
As with any new technology, AI presents challenges that need to be addressed. Since AI systems possess huge amounts of data, ensuring the privacy and security of personal information is crucial. There is also concern about job displacement. While AI will create new job opportunities, some existing roles may become automated. The prevailing wisdom for a long time was that AI would mainly displace lower-skilled jobs, but that is no longer the case. An example is software coding where AI can already augment the skills of current programmers enabling them to be more productive. It’s possible that we will need fewer of them in the long run.
An ongoing problem with these models is that they are prone to making up convincing answers even when they have no knowledge. The term that is usually used is “hallucination”. This limits their effectiveness in environments where perfection is key.
Ensuring AI is used responsibly and doesn’t perpetuate biases or cause harm is an ongoing challenge. There is a great deal of concern about our upcoming U.S. election as it’s clear that AI can be used to disrupt it when employed by malicious actors. Finally, as AI systems become more autonomous, ensuring they operate safely and remain under human control is extremely important. There is ongoing work in some countries to create AI based weapons which should be a cause for concern for everyone.
So where is this going? The latest buzzword in the industry is AGI, or Artificial General Intelligence. This is when AI can outperform humans in general, not just on specific tasks. OpenAI has defined a five-level classification system to define where we are on this spectrum. Currently they say we are at Level 1 which designates AI that can interact conversationally. Level 5 refers to AI systems that can do the work of entire organizations. They believe we can achieve Level 5 within ten years. Some experts predict that we will be there within five years or even less. One mitigating factor is cost. These models require enormous investments in computing hardware to develop and it’s not clear yet that the companies developing them have figured out how to monetize AI effectively. In any case, this is an exciting time to be alive!
Bruce Gobioff is a graduate of MIT and spent 35 years in the computer industry with IBM as a software developer and manager. He has been an OLLI-USF member for nine years as well as a being a long- time OLLI instructor. He is currently teaching OLLI classes in both liberal arts and technology subjects including Artificial Intelligence. His next class on AI is “Understanding AI Today” and will take place in November. You can learn more about it and register for it now by clicking here. Click on his photo for an invitation.
Leave a Reply