The answer to this question is highly dependent what you desire as far as gameplay. There are pros and cons to both systems.
Are your clients constantly moving around?
Will other connected clients need to see them in their exact positions or can they be fudged a little? (think WoW where everyone sees each other in slightly different positions)
In an fast paced game (think first person shooter), the client should be sending only their inputs to the server. The server receives the inputs, validates them, applies them to its version of the game world, and ultimately sends the client back a snapshot that contains their changes. The client is never trusted to determine its own position (or health, money, etc.). Pros to this system is it's not possible to cheat the server by telling it impossible things. The server will determine the outcome of all actions so they are always correct. Cons are increased communication necessary between client and server and having to implement more client side prediction and recovery code to make inputs
appear instantaneous. It simply wouldn't do for clients to wait 150ms round trip before their actions took effect, so you predict what will happen on the client side and make corrections based on what the server tells you later.
If your game is twitch based and your players are taking aim at each other, then it makes sense to implement a snapshot based system so everyone is seeing the same thing. If your game is more target based, like you select someone and then cast a spell but never have to aim the spell, you might be able to get away with sending just positions (and doing some checks for impossibility). This is what many MMOs do because it allows them to lower the amount of traffic to the server for basic movements.
Edit: As far as collision detection goes, this is just another thing that the server can do on its end and inform the clients of. Here's sort of an outline of what a server might do in a FPS:
every 15 millseconds:
1. check network for updates clients have sent me, store the unprocessed input for consumption
2. for each player in the world, apply any unprocessed input gathered in step 1
3. run a simulation step and update the game world based on the time since last update (ex. move any projectiles forward)
4. check all game objects for any collisions, perform appropriate actions (ex. move players back so they aren't colliding, apply damage for any projectile hits)
5. clean up the world, removing dead entities
We have a world that is completely up to date. We could send a snapshot of this world to all players, but it's usually not necessary to do so every tick.
6. if it's time to send a snapshot, create one and send it all connected clients
More information:
https://developer.valvesoftware.com/wiki/Source_Multiplayer_Networking