Kevin Scott

One of the things that folks talk a lot about with modern AI systems are hallucinations: instances where an LLM “makes up“ a response to a prompt when “made up“ is not desired. (For some prompts, e.g., “write me a nursery rhyme about walruses eating cake“, made up is exactly what you want!) My colleagues Steve Lucco and Anders Hejlsberg have just recently open sourced a system called TypeChat, that I think is one of the most powerful tools in the AI development arsenal for getting rid of undesired hallucinations. If you're building an AI application, you should check it out!

GPT-4 isn't an AGI (an artificial general intelligence), and it's silly to frame your thinking about it as a general stand in for human intelligence. The way that I approach these technologies is imagining what sorts of tasks they might be useful for, and then testing them out to see how well they work. What we see in each generation of the technology is that they get better at some obvious things, and some non-obvious things go from impossible to possible. Here's an example of what that progress can look like:

Sal Khan at Khan Academy has been on a mission for many years to provide a free, world-class education for anyone, anywhere. They've done some really interesting work with GPT-4 on tutoring that could mean big progress toward that mission.