Depth MoSeq is great, but it has one major limitation – it depends upon using videos taken by somewhat janky, slow and often noisy 3D cameras, like the discontinued Microsoft Kinect. What would be ideal is to instead use fast and standard 2D cameras (either alone or in combination) to track keypoints using standard algorithms like DeepLabCut or SLEAP, and then to feed those keypoints to MoSeq to identify the behavioral motifs or “syllables” our of which behavior is organized. There is a problem, though – unlike most unsupervised behavioral clustering algorithms, by design MoSeq tries to identify the exact moment when one behavioral syllable switches to the next. It does so in part by looking for discontinuities in behavior, moments when behavior abruptly changes from one mode to another. Because the output of keypoint tracking code is often pretty flickery (with individual tracked keypoints bouncing from one location to another), MoSeq breaks when fed raw keypoints – each flicker becomes a new behavioral transition. In this new paper just published in Nature Methods, the amazing Caleb Weinreb (together with our longtime friend and collaborator Scott Linderman) figured out a solution to this problem — have MoSeq infer keypoint positions based upon the premise they participate in syllable-like dynamics. Modifying MoSeq in this way – to create what we are calling keypoint-MoSeq – enables any keypoint data from any camera type to be fed to MoSeq. There is *so* much validation in this paper – we compare depth to 2D and 3D keypoints, show that this works on both mice and rats, demonstrate it works across environments (e.g., in the open field and in a complex home cage), and more! One important point – we think specifically modifying MoSeq to accept keypoints was especially worth it for folks interested in moment-by-moment relationships between neural activity and behavior. We compare depth and keypoint MoSeq at their ability to detect the systematic fluctuation in dopamine associated with syllable transitions described in this paper from the lab published last year – we find that both methods work about as well at detecting this fast rhythm in dopamine associated with syllable switching. BUT…when we used alternative behavioral clustering methods that don’t privilege time in the same way as MoSeq, this well-established (and causal) neurobehavioral relationship disappeared. There are likely many settings in which these alternatives are superior to MoSeq (largely for practical reasons), but if ultimately your goal is to ask how brain activity tracks (not on average, but in each instance) with behavior, time-sensitive methods like MoSeq remain the best current bet. Congrats to Caleb (and our many collaborators!) on this big step forward to behavioral characterization.