Will We Get AGI In 2025? (youtube.com)
It's doing a bit better on extremely complex (for us) math. Most people can solve odd problems without having seen anything like them before but LLMs are a long way off, she said.
Many say LLMs will ‘scale’ so the more data you put into them, the better they'll perform.
Because it has cost millions to train the current LLMs, it will be too expensive to just start again and train new ones, she speculates, and that makes it ‘easy to predict’ what they'll do, that the companies will drop the AGI idea and find niches that will be profitable. They might need to spam their customers with ads. (I don't mind ads as long as they don't use my personal data, even my browsing habits, and don't put too many or make my browser vulnerable. Ads are the way free stuff happens.)
For some reason AI does bad at this.