I appreciated her closing response and the distinction made between over-hyped LLMs and specific ML models.

AI is such an interesting word because it's sort of like the word transportation and that you have bicycles, you have gas guzzling trucks, you have rocket ships, they're all forms of transportation, but they all serve different purposes and they have different cost benefit trade-offs.

And to me the quest to artificial general intelligence has the worst trade-offs because you are trying to build fundamentally an everything machine, but ultimately it can't actually do all of the things. So not only do you confuse the public about what you can actually do with these technologies, which leads to harm because then people start asking it for things like medical information and instead getting medical misinformation back.

But also it requires all of these things that I described, the colossal resource consumption, the colossal labor exploitation. But there are many, many different types of AI technologies that I think are hugely beneficial. And this is task specific models that are meant to target solving a specific well scoped challenge, something like integrating renewable energy into the grid, weather prediction, drug discovery, health care, where you identify cancer earlier on in an MRI scan.

These are all very task specific. It's very clear what the use case is. It's — you can curate very, very small data sets, train them on very, very small computers. And I think if we want broad based benefit from AI, we need broad based distribution of these types of AI technologies across all different industries.

Excellent interview. Karen Hao is the hero we need but don’t deserve.