One major roadblock to something like this is the precision (and invasiveness) of the probes. You can get very imprecise probes that are not invasive and you will get an average reading from a large number of neurons. This may be enough to make predictions but likely much too noisy to actually “read” any thoughts. Alternatively you can get very precise probes that are connected to single neurons (or at least a very small number), but it’s invasive (you literally have a wire connected to the nerve tissue). You can now very accurately read individual values but you probably need a lot of these to actually determine what’s going on since a single neurons state isn’t enough data. Additionally, such invasive probes so far eventually are always rejected by your body.

It is, of course, a spectrum. You can get a middle ground between the two extremes but it’s still a trade off between invasiveness (and the problems and concerns that come with that, such as rejection), precision (how many neurons worth of data is being averaged) and the need to probe many neurons to build an accurate model.