LLMs cannot learn in incrementally in real-time
Real AGI needs the ability to learn new knowledge and skills incrementally in real-time. Whether AGIs are employed as researchers, workers, personal assistants or some other function, a common requirement is that they are able to adapt to changing situations and requirements on the fly, and to do so autonomously.
Imagine an AGI researcher discussing an important, highly-referenced paper with someone and being informed that this paper has just been withdrawn. This single fact will immediately invalidate a number of assumptions and raise a number of questions. The AGI can only properly handle additional related information or instructions once it has reasoned through and integrated implications of this change. Its core model or knowledge representation must be updated in real-time.
Similarly, a human-level AGI worker or personal assistant must be able to properly react to important everyday news such as changed relationships, locations, products, business rules or laws, etc. The system needs to immediately think through implications to existing knowledge, values and priorities, and to proactively gather additional clarifying data.
AGIs must be able to work autonomously, knowing when and whom to ask for help or information. A core assumption is that there will not always be a human-in-the-loop to monitor and correct the system, or to prompt-engineer it until it gets it right!
Another reason that real-time, incremental learning is so important is that many specific real-world tasks and problems are dynamic in nature, quite rare, novel, or unique and thus lack (sufficient) prior training data. We should expect AGIs to be able to learn instantly, incrementally with sparse data, just the way we can.
Aigo’s INSA (Integrated Neuro-Symbolic Architecture) facilitates Real-Time Incremental Autonomous Learning.
A recent whitepaper reviewed more than 200 research papers concerned with Incremental Learning in Large Language Models (LLMs). A notable finding was that not a single reference could be found of LLMs learning incrementally in real-time — all of them require batch updates. Furthermore, most of these approaches did not update the core model at all but rather manipulated only the input buffer, the output layer, or an external database or RAG.
Clearly, an advanced dependable AI agent must be able to adjust its core knowledge in real time with limited input. We certainly expect this of human workers.