You may be right that the majority is probably an exchange. But is there something in the protocol specification that prohibits such majority?
Then it raises a question, can an actor such as a government (which may have unlimited resources) to hold large numbers of ether (> 51%) to add blocks to their advantage?
Perhaps, it's my ignorance about this technology that makes me question and prevent me from adopting this technology.
There's a common misconception that once you have >51% you can do anything. This isn't true. There's plenty of mischief you might get away with (censoring, double spending), but you can't transfer other peoples money without their private key, and you can't change the rules of the protocol. You can probably tank the value of the currency by doing large enough double spends and causing problems, but in PoS importantly you're hurting yourself more than anyone else, while in PoW, you still have a bunch of useful hardware left over after the attack, and with hash power marketplaces you can attack a PoW chain while having more or less no investment in the chain itself.
I keep hearing "you can't do X without getting slashed". What happens if there is a network partition that lasts for longer than 6 minutes? Which two of the diverging blockchains get to slash the other one and take all their stake?
> What happens if there is a network partition that lasts for longer than 6 minutes?
With less than 2/3 of the total stake active on a single partition, that partition stops finalizing transactions, meaning that the chain explicitly stops guaranteeing that it's canonical.
Notably, slashing cannot result from a partition, only from malicious validator behavior.
> Which two of the diverging blockchains get to slash the other one and take all their stake?
For a partition, which is not a slashable offense, there is no slashing. The minority partition stakers suffer inactivity leak on the majority chain, meaning that they very slowly (at first) start losing their stake until the majority partition has 2/3 stake again. It's not a big penalty like slashing, unless the chain remains in a degenerate state for many hours or days.
On the other hand, a slashing rules offender (attacker) gets slashed on all chain forks. The conflicting signed block from one gets included on all others for a bounty. This means that every staker must vote for only one fork at a time, which means the network can eventually determine which fork is canonical because it was voted for the most.
The short answers is yes, there are various safeguards and countermeasures in place, some on the protocol and some on the social/incentive layer. But it would take dozens of pages to explain all of these in detail, so if you are truly interested in this you can search for "proof of stake security" or "proof of stake centralization risk" and you will find a huge number of resources.
Perhaps, it's my ignorance about this technology that makes me question and prevent me from adopting this technology.