AI came to general public knowledge with huge success by beating the best humans at games like Chess and Go much sooner than was expected. Here is a good movie about it:
Another big success, the use of AI for voice recognition, is now nearly ubiquitous.
Decades of not much success in the 20th century led to predictions of a continued lack of success. Those predictions failed with the stunning breakthrough of machines like AlphaZero which can play many different games far better than humans
without being taught by humans. We just need to give the machine the rules and let it play against itself to become a world champ in a matter of days. Truly breathtaking.
This rapid success led to predictions of similar rapid AI success in other areas. Alas, these prediction were also wrong although we are now starting to see some pretty spectacular early results from things like ChatGPT and AI art.
In some domains AI totally dominates its human competition. In other domains, where the problem space is less clearly defined, researchers are still struggling to surpass what humans can do.
There is a cycle:
- Predictions of mind boggling success
- Failure to meet those predictions
- Predictions of doom and gloom
- Unexpected massive success
- rinse, repeat
We seem to currently be in phase (3), perhaps close to rounding the corner into another phase (4) spate of huge successes.
It's true, humans are probably still better than AIs at most things. But AIs are advancing rapidly while humans are almost standing still. The fact that we can get AIs to solve very hard problems merely by training them (not programming them or teaching them) is mind-boggling.
Another thing to keep in mind is the wide-spread "failures" of AIs to do exactly the right thing are a part of the plan. AIs need massive amounts of training data. Many orders of magnitude more than humans need. Widely exposing "failures" to the public is giving AI teams free training. Likewise, Elon is counting on the huge fleet of AI equipped cars to provide the training needed to solve the FSD problem. All the cars they sell now are equipped and help with the training even if FSD is not enabled.
AFAIK, we have not yet found a limit to what AIs can learn given a large enough neural net and enough training data. When I look back over the past 3 or 4 decades of AI research, it sure looks like we should expect some astounding breakthroughs in the next 5 (maybe 10) years. But it's devilishly hard to predict exactly when those breakthroughs will happen. I would not be at all surprised if the tables flipped in the next 10 years and AIs became far better than humans at most things.