The benefits and limitations of generative AI have been highlighted by researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL). The main topics of conversation were the language, imagery, and source code of the new technology and its effects on society and industry.
The system can generate new data or material comparable to what the generative AI was trained on. Generative AI models, such as generative adversarial networks (GANs), variational autoencoders (VAEs), and diffusion models, have shown extraordinary efficacy across various domains, from the creative to performing arts to science and health.
Fake news, deep fakes, and misinformation are only some of the ethical and social concerns brought up by these models. The researchers stressed the need to consider these concerns when examining the strengths and weaknesses of generative AI and making decisions on how to apply it responsibly and ethically.
EECS Professor and CSAIL Director Daniela Rus of MIT shared a collage of artificial intelligence pictures her students had generated using a computer. These pictures show the models’ visual capabilities, as they were trained using many publicly available images from the internet to produce an image that looks like the training data.
One common approach to training a neural network generator was recently addressed by Phillip Isola, Associate Professor of EECS and CSAIL Principal Investigator at MIT. The most exciting feature of generative data, according to Isola, is not the power to generate photorealistic visuals but rather the unparalleled degree of control it grants us.
As a result, we can now input a description like “Van Gogh style” and have the model generate an image that matches that description, demonstrating the power of language as an interface for image generation. In cases where words alone are inadequate, visual aids such as sketches can be employed to give the model more precise input and produce the required results.
“It could be difficult, for instance, to convey the precise location of a mountain in the background of a portrait,” Isola explains.
The significance of evaluating the reliability of generating content was also highlighted. Numerous benchmarks have been developed to demonstrate that models can perform at or beyond human levels on specific examinations or tasks requiring sophisticated linguistic skills.
CSAIL Principal Investigator and EECS Assistant Professor at MIT Jacob Andreas stated that models can now create poetry, hold discussions and generate targeted papers to represent concepts that seem to be wishes and beliefs.
In word embeddings, words with similar meanings are given numerical values (vectors) and arranged in a space with many dimensions to harness the power of these models. To facilitate several rounds of dynamic interactions between various elements, transformer models employ an “attention mechanism” that selectively focuses on selected parts of the input sequence.
Further, as Andreas pointed out, users can push the envelope by providing sources and citations to other users as they build an argument. However, he emphasised that even the most advanced models available today are not yet capable of doing so in a trustworthy manner and that much more effort is needed to make these sources credible and objective.
While Armando Solar-Lezama, EECS Professor at MIT and CSAIL primary investigator, investigated elaborate language models – the kind that sometimes feels profoundly “meta” – the type that produces code. They are like tiny magic wands, except instead of casting spells, they create lines of code and make (some) software developers’ fantasies come true.
Lezama also discussed the fight against the industry’s enormous and heavy pockets and the success of establishing large language models. Academic models “need huge computers” to develop the technologies above without excessive reliance on industrial backing.
Andreas concludes that while being modest about one’s talents is necessary, the risks of underestimating one’s impact should not be disregarded. “We still don’t know what these models can and can’t achieve, so we should approach them humbly.”