If you're willing to "drink from the fire hose", it's probably hard to beat just browsing recent arXiv submissions when it comes to seeing the most novel results in AI/ML.
For the unfamiliar, arXiv is a preprint repository where academic articles are published before/as they are submitted to an academic journal. It also contains work which is not intended for later publication, such as replicating someone else's results.
The two major caveats are 1) these articles are written with an expert audience in mind, so may not feel very inviting and 2) there is no barrier to entry to publishing to arXiv, so there is no guarantee of quality for any individual article.
ArXiv has an AI category, as well as a CS/ML and Stats/ML category.