Sequence to sequence learning with neural networks

In his book "Thinking, Fast and Slow", Daniel Kahneman divides consciousness into two modes of operation: the reflective mode, where we act deliberately, learn, analyze new information, and the autopilot mode, where we rely on accumulated knowledge. The reflective mode requires significant effort, and we can stay in it only briefly before quickly becoming fatigued. In autopilot mode, we operate the rest of the time—it is effortless for us, but in this mode, we can make mistakes if we fail to recognize in time that the situation demands deliberate decision-making.

Large language models exhibit a similar property: they also operate in two modes—training mode and automatic response mode. In training mode, we train the model on a dataset or fine-tune it on a custom dataset. In question-answer mode, the neural network no longer adjusts its parameters; it uses accumulated experience to generate the expected response, just like a person in autopilot mode.

What if, in the future, neural networks learn to independently switch to training mode and train themselves? This would give rise to fully-fledged intelligences capable of adapting to new conditions and evolving. To achieve this, a large language model would need to be equipped with "senses" to process the surrounding reality and be allowed to expand and modify its own weight parameters. This is precisely what Ilya Sutskever speculated about in his talk. And this is exactly what we are now observing as we shift our focus from LLMs to AI agents. It seems this is currently the main driving force and the primary direction in the development of artificial intelligence.