Explainability in AI is really overlooked and often skipped over as there is little progress in this area. GPT-3 is essentially GPT-2 + tons of data, compute and parameters and yet it still cannot explain itself as to why it can generate 'human-level' text, much like how AlphaGo can't explain why it performed move 37. Not discrediting these achievements, but explainability is just as important in these AI models.
Once you have an AI-based 'auto-pilot' in any vehicle, the importance of AI explainability will haunt manufacturers when the regulators would want them to explain why this 'AI' took this decision and they're unable to explain this.
I hope GPT-4 isn't just going to be GPT-3 + 1000x the data. Otherwise nothing would have changed here other than the parameters and data.
The easy solution is to copy what the human brain does: just make up something plausible. There's pretty good evidence that we don't have great introspective access to much of our own internal processing. We just paper it over as "intuition" or "judgement."
We already see this in superhuman stock algorithms. You can "debug" them, in the sense that for a given trade, it can tell you what signals provoked it. But they don't make any sense: it saw rainfall in the Amazon tick up, the price of beef in Russia tick down, and the UK call a snap election, so it bought more GE stock.
You could... theoretically... write a story that connected those dots, but it will either be facile or nonsensical. That's because the model of the market the algo has is bigger and more complete than anything a human can have. It's drawing a straight line through some upper-dimensional manifold that you can't comprehend.
It can't explain what it's doing to you anymore than you can explain "algorithmic stock trading" to a three year old child. You can say what the outcome was, but you can't explain it in such a way that the kid could replicate the performance.
Explainability in AI is really overlooked and often skipped over as there is little progress in this area. GPT-3 is essentially GPT-2 + tons of data, compute and parameters and yet it still cannot explain itself as to why it can generate 'human-level' text, much like how AlphaGo can't explain why it performed move 37. Not discrediting these achievements, but explainability is just as important in these AI models.
Once you have an AI-based 'auto-pilot' in any vehicle, the importance of AI explainability will haunt manufacturers when the regulators would want them to explain why this 'AI' took this decision and they're unable to explain this.
I hope GPT-4 isn't just going to be GPT-3 + 1000x the data. Otherwise nothing would have changed here other than the parameters and data.