Speaking to retail executives in 2010, Rama Ramakrishnan came to two realizations. First, although retail systems that offered customers personalized recommendations attracted a lot of attention, these systems often provided little reward to retailers. Second, for many of the companies, most customers only shopped once or twice a year, so the companies didn’t really know much about them.
“But by being very diligent about recording the interactions a customer has with a retailer or an e-commerce site, we can build a very nice and detailed composite picture of what that person is doing and what they’re interested in,” says Ramakrishnan, professor. of Medicine at the MIT Sloan School of Management. “Once you get that, then you can apply proven algorithms from machine learning.”
These realizations led Ramakrishnan to found CQuotient, a startup whose software has now become the foundation for Salesforce’s widespread AI e-commerce platform. “On Black Friday alone, CQuotient technology probably sees and interacts with more than a billion shoppers in a single day,” he says.
After a highly successful business career, in 2019 Ramakrishnan returned to MIT Sloan, where he earned an MA and PhD in business research in the 1990s. He teaches students “not just how these amazing technologies work, but how to get them these technologies and use them realistically in the real world,” he says.
In addition, Ramakrishnan enjoys participating in MIT executive education. “This is a great opportunity for me to pass on the things I’ve learned, but also just as importantly, to find out what’s on the minds of these senior executives and guide them and push them in the right direction,” he says.
For example, executives are understandably concerned about the need for massive amounts of data to train machine learning systems. Now it can guide them through a multitude of models that are pre-trained for specific tasks. “The ability to take these pre-trained AI models and adapt them very quickly to your specific business problem is an incredible advance,” says Ramakrishnan.
Understanding AI categories
“AI is the quest to imbue computers with the ability to do cognitive tasks that usually only humans can do,” he says. Understanding the history of this complex, supercharged landscape helps exploit the technologies.
The traditional AI approach, which basically solved problems by applying rules if/then we were taught by humans, proved useful for relatively few tasks. “One reason is that we can do many things effortlessly, but if asked to explain how we do them, we can’t really articulate how we do them,” comments Ramakrishnan. Also, these systems may be confused by new situations that do not match the rules included in the software.
Machine learning takes a dramatically different approach, with software basically learning by example. “You give it lots of examples of inputs and outputs, questions and answers, tasks and answers, and you have the computer automatically learn how to go from input to output,” he says. Credit scoring, loan decision making, disease forecasting and demand forecasting are among the many tasks mastered by machine learning.
But machine learning only worked well when the input data was structured, for example in a spreadsheet. “If the input data was unstructured, like images, video, audio, EKGs or X-rays, it wasn’t very good to convert from that to a predicted output,” says Ramakrishnan. This means that people had to manually structure the unstructured data to train the system.
Around 2010 deep learning began to overcome this limitation, providing the ability to work directly with unstructured input data, he says. Based on a long-standing artificial intelligence strategy known as neural networks, deep learning became practical due to the global flood of data, the availability of extremely powerful parallel processing hardware called graphics processing units (originally invented for video games), and advances in algorithms and mathematics.
Finally, in the context of deep learning, the generative AI software packages that emerged last year can generate unstructured output such as human-audible text, images of dogs, and 3D models. Large language models (LLMs) like OpenAI’s ChatGPT go from text input to text outputs, while text-to-image models like OpenAI’s DALL-E can produce realistic images.
Rama Ramakrishnan – Small data note to improve customer service
Video: MIT Industrial Liaison Program
What genetic AI can (and can’t) do
Trained on the Internet’s unimaginably vast text resources, an LLM’s “fundamental skill is to predict the next most likely, most plausible word,” says Ramakrishnan. “It then attaches the word to the original sentence, predicts the next word again, and keeps doing it.”
“To the surprise of many, including many researchers, an LLM can do some very complex things,” he says. “He can compose beautifully coherent poetry, write episodes of Seinfeld, and solve certain kinds of reasoning problems. It’s really quite remarkable how predicting the next word can lead to these amazing possibilities.”
“But you always have to keep in mind that what it’s doing is not so much finding the right answer to your question as finding a reasonable answer to your question,” Ramakrishnan emphasizes. Its content may be factually inaccurate, irrelevant, toxic, discriminatory or offensive.
This puts the onus on users to make sure the output is correct, relevant and useful for the task they are performing. “You have to make sure there’s some way to check its output for errors and fix them before it goes out,” he says.
Intensive research is underway to find techniques to address these shortcomings, adds Ramakrishnan, who expects many innovative tools to do so.
Finding the right corporate roles for LLMs
Given the amazing progress in LLMs, how should industry think about applying software to tasks like content production?
First, Ramakrishnan advises, consider the cost: “Is it a much less expensive effort to have a draft that you fix than to create the whole thing?” Second, if the LLM makes a slip-up and the incorrect content is released to the outside world, can you live with the consequences?
“If you have an application that satisfies both considerations, then it’s good to do a pilot project to see if these technologies can really help you with that particular task,” says Ramakrishnan. He emphasizes the need to treat the pilot as an experiment rather than a regular IT project.
Currently, software development is the most mature corporate LLM application. “ChatGPT and other LLMs are text-in, text-out, and a software program is just text output,” he says. “Developers can switch from English text to Python text output, just as you can switch from English to English or English to German. There are many tools that help you write code using these technologies.”
Of course, developers need to make sure that the output does the job correctly. Fortunately, software development already offers infrastructure for code testing and verification. “That’s a pretty sweet spot,” he says, “where it’s a lot cheaper to have the technology write code for you because you can very quickly test it and verify it.”
Another important use of the LLM is content production, such as writing marketing copy or e-commerce product descriptions. “Again, it can be much cheaper to fix the draft of ChatGPT than to write the whole thing,” says Ramakrishnan. “However, companies have to be very careful to make sure there is a human in the loop.”
LLMs are also spreading rapidly as internal tools for searching corporate documents. Unlike conventional search algorithms, an LLM chatbot can provide a conversational search experience because it remembers every question you ask. “But then again, occasionally it will fix things,” he says. “When it comes to chatbots for external customers, these are very early days because of the risk of saying the wrong thing to the customer.”
Overall, notes Ramakrishnan, we live in a remarkable time to confront the rapidly evolving potential and pitfalls of artificial intelligence. “I help companies figure out how to take these very transformative technologies and implement them, make products and services much smarter, employees much more productive, and processes much more efficient,” he says.