What do you (Gary) expect if you bring up a text-to-image generated mess of a bicycle and shoot that down as if it's claiming to be AI? Maybe a bit more intellectual honesty about what these systems are and claim to be (or don't claim to be) would result in less criticism.
Symbolic systems like CYC and SOAR were always going to be brittle, limited by what was programmed into them, and perhaps equally fundamentally seem to have been addressing the wrong problems (knowledge acquisition and how to solve problems).
With today's large scale neural nets finally we have an approach that shows far more promise. Gone is the brittleness since they are based on learning rather than canned data and algorithms. Tomorrow's systems will learn continuously. Gone is a dependency on hand coded algorithms, replaced by emergent capabilities derived from their ability to learn in self-supervised fashion from data and experience.
Perhaps most promising is that these LLMs derive their emergent capabilities from addressing the problem of prediction, which appears to be the same way our cortex functions. In fact I'd suggest it's fair to define intelligence as nothing more than the ability to predict based on experience, including all that implies such planning based on ability on ability to predict outcomes.
Maybe we shouldn't be so surprised at the emergent capabilities of an LLM such as GPT-3. Trained to predict next word, yet capable of predicting the computer code that will follow a comment describing what it's about to do ...
Yes, we're not there yet in terms of AGI, nor is anyone claiming that we are, but the progress made in last 10 years has been nothing short of staggering, and there is no sign of it slowing down. We are finally making progress, and it is progress in the right direction unlike symbolic AI attempts of past that were flawed by conception.
Symbolic systems like CYC and SOAR were always going to be brittle, limited by what was programmed into them, and perhaps equally fundamentally seem to have been addressing the wrong problems (knowledge acquisition and how to solve problems).
With today's large scale neural nets finally we have an approach that shows far more promise. Gone is the brittleness since they are based on learning rather than canned data and algorithms. Tomorrow's systems will learn continuously. Gone is a dependency on hand coded algorithms, replaced by emergent capabilities derived from their ability to learn in self-supervised fashion from data and experience.
Perhaps most promising is that these LLMs derive their emergent capabilities from addressing the problem of prediction, which appears to be the same way our cortex functions. In fact I'd suggest it's fair to define intelligence as nothing more than the ability to predict based on experience, including all that implies such planning based on ability on ability to predict outcomes.
Maybe we shouldn't be so surprised at the emergent capabilities of an LLM such as GPT-3. Trained to predict next word, yet capable of predicting the computer code that will follow a comment describing what it's about to do ...
Yes, we're not there yet in terms of AGI, nor is anyone claiming that we are, but the progress made in last 10 years has been nothing short of staggering, and there is no sign of it slowing down. We are finally making progress, and it is progress in the right direction unlike symbolic AI attempts of past that were flawed by conception.