Hacker News Reader

Ask ourselves? Who is us? Your comment doesn’t apply if everyone has access to their own local AGI.

Right now training is expensive and inference is cheap. There’s no reason to believe that inference cost will suddenly balloon when we cross the line to general intelligence.

View on HN

It's easier to understand if you flip the perspective. The AGI has access to you.

Why are you letting your AGI run wild? Let’s make the assumption that AGI is just a GPT combined with a new training algorithm that allows it to learn rapidly and generalize from small amounts of information. The moment you turn off the training algo your AGI reverts to what’s essentially a next token predictor. It retains everything it learns but is unable to form long term memories.

I think it would be very hard for anyone to be manipulated by an entity without long term memory.

The counter point to that, is that in order to get AGI you will need long term memory.

In which case we will design a long term memory at first to gather the benefits that came come AGI.

Reply on news.ycombinator.com