It’s one of best designs I have seen, I admit. But for that price you cannot get absolute encoders outside the motor, reliable force/torque sensors (think picking up a strawberry), tendons (thread below). It might be too limited for research and real-world projects unfortunately.
Depends entirely on the application I guess, but you'd get further for sure. Either way you are now confronted with more complex cable management and maintenance, and additional weight. You'd at least get around gravity compensation with that approach. Personally, I think you need a lot more sensors and complex sensor fusion to achieve acceptable dexterity to be useful outside of an industrial context and this comes at high costs
