Written By Audrey Woods
CSAIL Director Professor Daniela Rus introduced the panel by discussing the exciting potential of artificial intelligence, saying that these technologies “have the potential to transform not only academic disciplines but entire economies,” especially with the arrival of “extraordinary tools” like ChatGPT, Bing Chat, Dall-E, and Stable Diffusion. Roughly defined as a system which can create new data similar to but different from the data the model was trained on, generative AI has enormous implications for how we interact with technology and offers potential technical applications across various industry sectors. However, these advancements have sparked debate about how generative AI should be implemented, regulated, understood, and approached.
Here are the key takeaways from CSAIL Professors Rus, Armando Solar-Lezama, Phillip Isola, and Jacob Andreas on the topic of generative AI:
How These Models Work
- Professor Isola spoke about image generators, which create images by adding an increasing amount of noise, or randomness, to images in the training dataset and then reverse that process to “de-noise” based on adjustable vectors, which he compared to knobs, controlling aspects like color, angle, size, etc. A typical image generator might have hundreds of knobs.
- Professor Andreas explained how large language models are fundamentally predictive text models that map language in multiple dimensions, selecting the next word based on what is dimensionally closest to the one before it. This requires a basic understanding of grammar, facts, physical common sense, and training on classical, well-known texts such as Shakespeare.
- Professor Solar-Lezama then discussed code generation and how key breakthroughs in transformers and language models have led to the possibility of generating functional code with AI. There are key challenges to this process—code is brittle, large, and diverse, requiring more precise outputs than in natural language—but there are benefits too, since code is verifiable and testable and thus such models can sustain a higher bar for accuracy.
Concerns & Considerations
- Many audience questions had to do with the fear of generative AI replacing human work entirely, making programmers, artists, and other knowledge workers obsolete. The panelists eased this concern by pointing out that there is arguably more value than ever in “algorithmic thinking,” as Professor Solar-Lezama said, and that the emotional value of human-created work will likely outweigh whatever impact AI generated art will have.
- Several concerns were brought up around safety, especially the potential malicious use of deepfakes and the danger of perpetuating bias and harm with bad datasets. The researchers discussed how such issues should be approached in a cohesive way, with not only technologists but also policymakers and the general public participating in a combination of laws, social norms, and future technical solutions.
- A current drawback with these large AI models is that many of them lack controllability and explainability. Professor Andreas touched upon this when asked about a specific failure point of ChatGPT, saying that researchers don’t know enough about what’s going on inside these models to confidently point to a reason for a given error. It’s also difficult to, for example, remove specific data points from models after they’re trained on it, leading to privacy concerns.
- Creating generative AI models can require enormous resources, specifically computing power. While it’s tempting to scale up with brute force, Professor Solar-Lezama pointed out that “scale buys you a lot, but it’s not going to buy you everything.” Professor Isola agreed, highlighting how algorithmic innovation could make the learning process more efficient by, for example, maximizing each data point.
- On the philosophical side, several questions brought up the ethical considerations of generative AI models, particularly the inclination to anthropomorphize them and whether or not they will ever reach true human intelligence. While Professor Andreas said that these models are currently a long way from a human-like understanding of the world, he advised “humility” when considering what future models might be able to conceptualize.
Future Opportunities for Work & Research
- It’s important for generative AI models to be certifiable, which means there is work to be done creating evaluations and benchmarks to test what these models can do and identify their limitations.
- Professor Solar-Lezama talked about how important it is for research institutions like MIT CSAIL to build large generative models from the ground up rather than just experiment with available models. He said that while this will require significant funding, it’s necessary to help academics understand how the technology works at a fundamental level.
- As mentioned above, there are technical approaches to many key concerns around generative AI, such as computational ways to identify deepfakes, more efficient architecture structures, controllability solutions, privacy measures, etc.
- Professor Isola highlighted the exciting potential of using generative AI to create tangible products such as drugs, architecture designs, fabrication ideas, and more. Professor Rus added that summarizing corpuses of data could also prove useful in areas such as law.
It’s important to keep in mind—as all four professors emphasized at various points in the conversation—that generative AI is, first and foremost, a tool. This technology is neither inherently benevolent or malicious and depends entirely on how it is used by researchers, business leaders, and the general public. Professor Rus concluded, “we have questions about intelligence and questions about ourselves that we need to answer sooner rather than later before AI plays out without a careful hand.” Thankfully, these MIT researchers are hard at work finding ethical, functional, and inventive ways to apply generative AI.
As Professor Solar-Lezama said when asked whether these models will enhance or stagnate innovation, “I would never bet against human creativity.”
On March 20th, MIT CSAIL researchers Professor Daniela Rus, Professor Jacob Andreas, Professor Armando Solar-Lezama, and Professor Phillip Isola gathered to look under the hood of generative AI and its various subfields. Learn more about the event here.