9guesses for AI

 Let's see whether these pan out -

- Large language models based on transformers are going to reach human-level performance as a commodity in ~2 years. However, they will hit various roadblocks in making super-human intelligence in terms of reasoning. (philosophically, because you can't just feed a neural network data and expect it to be more intelligent than that data - and even if you accidentally create a smarter AI, you just won't recognize it as smarter.)
- Armed with knowledge that once you get enough data anything is learnable by machines using simple architectures, smaller models in various niches will become the new thing. They won't be on the headlines, but they will contribute to technological progress.
- Regulation frameworks on larger, general purpose models will be crazy. It will be like GDPR -- bigger players will just pretend to comply and ignore the actual law when it's inconvenient for them to do so.