Out of the box, Neuropixels are sharp in the dimension you see when looking at the face of the probe with the sites.
However, they are not sharp in the dimension that you see when looking at the edge, and in this dimension have a square-shaped tip that is the thickness of the probe (~25 µm). The probe tip out of the box is therefore a line of length ~25 µm and width ~1 µm.
The edge dimension can be sharpened using procedures described below, which can result in the probe tip being a point ~1 µm in size. See photos in the protocols linked for more idea of what this looks like.
It is evident from unpublished tests (by this author, Nick Steinmetz, and others), as well as from logic, that a probe sharpened in this way can more easily enter the brain, e.g. by puncturing dura. If you have difficulty inserting the probe into the brain in your preparation, sharpening can be a way to achieve this.
More speculatively, we infer that a sharp probe will pull/tear the brain tissue less after entering, and will instead do a better job sweeping neurons and processes out of the way, much like a "cowcatcher". Moreover, tissue compression, as when pushing on the surface of the dura before the probe passes through it, may negatively impact recording quality - so we additionally infer that recording quality (especially near the surface of the brain, e.g. in cortex) may be improved by compressionless insertions achieved with sharpening. So far as I know these observations have not been quantified (let me know if I'm wrong).
How to sharpen?
Two protocols have been developed:
Using a Narishige micropipette grinder, from Adam Davis and Nick Steinmetz at the University of Washington (following the approach developed by Eric Trautmann).
A comparison of methods
|Hard-drive method||Micropipette grinder method|
|Setup time||~2 hours ^||~0.5 hours|
|Sharpening time||~5 sec on the drive||~15 minutes on the grinder|
|Success rate||"100% out of hundreds" at Janelia. 1/2 at UW †||6/6 at UW|
* n.b. the exact parts list in the hard-drive method protocol actually costs ~$14000. Here I am just betting that you could do it for much cheaper with similar alternative parts, e.g. a manual manipulator instead of the Sutter MP-285, and somewhat less expensive optics.
^ an estimate: the hard drive "re-purposing" procedure is not described. It will also be longer if you're trying to figure out cheaper substitute parts relative to what is given in the protocol.
† The probe broken at UW with this method was the motivation for developing a different approach that did not involve grind speeds so high that the probe "flutters" and vibrates rapidly.