Large language models (LLMs) learn patterns from enormous collections of data and use them to generate responses. Most deployed models cannot continually update their central parameters from experience after training, Goertzel said. They can retrieve saved information or refer to earlier conversations, but those tools do not give them a continuous personal history.
“They don’t have a long-term memory of their whole life like a person does,” he said. “The result of that is they don’t know who and what they are and how they relate to the world around them.”
In his view, a generally intelligent machine would need a self-model, meaning an internal understanding of its history, abilities, limits and goals. It would also have to be able to preserve useful lessons and connect them to future decisions.
Conference participants discussed predictive coding as one possible route. The method asks a model to predict an outcome, compare its expectation with what actually occurs and learn from the difference.
Goertzel said his team has made predictive coding work in transformers at roughly the scale of GPT-2, an early OpenAI language model. The group hasn’t demonstrated the approach at the size of today’s leading models.
Continual learning addresses only part of the problem. Goertzel favors neural-symbolic AI, which combines neural networks that recognize patterns with symbolic methods that represent rules, relationships and goals more explicitly.
He also says that developers should place ethical reasoning, self-understanding and goal pursuit inside the core learning process. Adding safeguards after training doesn’t provide the same foundation, he said.
Proving that an AI has reached general intelligence could present another obstacle. Companies can train models to perform well on popular evaluations without demonstrating the broader skill each one aims to measure.
Goertzel proposes a more demanding alternative, which he calls the “MIT test.” An AI-controlled robot would have to earn a doctorate at the university while following the same rules as human students. The robot would need to navigate the campus, attend classes, pass written and oral exams, work with professors and produce original research. Success would require communication, physical competence, social judgment, long-term planning and the ability to develop new knowledge.
“If I could make a robot that could go to MIT and get a PhD without breaking any of the rules humans follow, I would be very impressed,” Goertzel said. “It’s got to roll through the hallways. It’s got to do an oral exam with someone, do a written exam and write a PhD thesis that passes the criteria of an original contribution to knowledge.”