To me, as a researcher in the field, the limitations that ChatGPT exposes about LLMs are much more interesting than what it actually can do. And what ChatGPT and its cousins can do is actually quite technically impressive.
The two big limitations of these LLMs seems to be:
- There is no good way to train an LLM to give "I don't know" as an answer. So if you ask it a question even a tiny bit outside of its knowledge, you will certainly still get an answer back -- an answer that often makes very little sense or is even outrageously wrong.
- They don't know when to shut up. You need to have some way to "trim" the answers you get back. And the longer the answer is, the more likely the LLM is likely to be bullsh!+ing you.
More generally, neural networks learn a representation of their training data. However, they don't learn an
efficient representation and aren't very efficient learners. What that means is best explained by example: you can show a four-year-old a handful of pictures of an elephant, and with a little bit of coaching that four-year-old will be pretty good at identifying elephants in other pictures. For an image classification AI to get the same level of performance, it will need to see
thousands of pictures of elephants. Another example: in most trained AIs, the vast majority of "parameters" (like ninety to ninety-nine percent of them) are zero or very close to zero. The parameters that are very close to zero probably would converge to zero with infinite training time, which is another source of inefficiency in their learning capabilities.
Having said all that, I think there are promising applications for neural networks in e-bikes. I think of neural networks as universal function approximators. Used in this way, it would be very promising to apply neural networks to both battery management systems and to modulating pedal assist. I suspect (but don't know for sure) that you could probably double battery life and effective range with a clever application of neural networks to the controller and charger of an e-bike. Doing so well would likely require a couple of years of data collection and testing, and today's impatient capital isn't really interested in doing something on that time scale. So we'll have to wait until somebody who doesn't know any better just goes out and does it.