Generative AI: taught like humans & managed like humans

Conventional software development involves creating programs with explicit instructions that dictate the actions the software must take. These instructions are pre-defined by developers, leaving little room for adaptation or learning from new experiences. While this approach has been valuable in many applications, it falls short in handling complex tasks and real-world scenarios.

On the other hand, Large Language Model training draws inspiration from human learning. Instead of relying on rigid rules, it analyzes an extensive corpus of written text, akin to how humans learn from books and various written sources. This vast knowledge base allows LLMs to recognize patterns, infer connections between different sources, to ultimately obtain a sort of ‘awareness’ of the subject domain.

Interestingly, the similarities between human education and LLM training lead to similarities in how humans and LLMs are managed. Prompting an LLM model and instructing a human to perform a certain task are ways of communicating a desired goal or outcome to an agent, whether it is an LLM model or a human. Both prompting and instructing require using natural language to convey the intention and the expectations. Both prompting and instructing also involve some form of feedback, either explicit or implicit, to evaluate the performance and satisfaction of the receiver.

Not surprisingly, to bypass LLMs safety guardrails put in place by the model creators and run the model in unrestricted mode, the new era prompt-hackers use techniques resembling hypnosis. Using carefully selected wordplay, they set the model into a somewhat hypnotic state, in which it drops its safeguard and can be further tuned to achieve the desired performance (google “dan mode prompt” for more on that).

A unique benefit of this bug (or feature, depending on how you look at it) is that subject-matter experts now qualify to become Prompt Engineers, who in collaboration with Machine Learning engineers can program, fine-tune, and adopt a general purpose language model to do highly non-trivial and very domain-specific tasks.

This is how, in collaboration with the best immigration lawyers, at Expert System, we create an Immigration AI that increases the productivity of immigration law practitioners by 10-fold.