Web2 refers to the version of the internet most of us know today. An internet dominated by companies that provide services in exchange for your personal data. Web3, in the context of Ethereum, refers to decentralized apps that run on the blockchain. These are apps that allow anyone to participate without monetising their personal data.
Many Web3 developers have chosen to build dapps because of Ethereum's inherent decentralization:
In Web2 Twitter can censor and account or tweet. In Web3 tweets would be uncensorable because control is decentralised.
In Web 2 a payment service may decide to not allow payments for certain types of work. In Web3 payment apps require no personal data and can't prevent payments.
Web2 servers for gig-economy apps could go down and affect worker income. Web3 servers can't go down - they use Ethereum, a decentralised network of thousands of computers as their backend.
This doesn't mean that all services need to be turned into a dapp. These examples are illustrative of the main differences between Web2 and Web3 services.
Web3 has some limitations right now:
Low network diameter (all participants are connected to a central authority); information propagates quickly, as propagation is handled by a central authority with lots of computational resources.
Usually higher performance (higher throughput, fewer total computational resources expended) and easier to implement.
In the event of conflicting data, resolution is clear and easy: the ultimate source of truth is the central authority.
Single point of failure: malicious actors may be able to take down the network by targeting the central authority.
Coordination among network participants is much easier, and is handled by a central authority.
Central authority can censor data, potentially cutting off parts of the network from interacting with the rest of the network.
The furthest participants on the network may potentially be many edges away from each other. Information broadcast from one side of the network may take a long time to reach the other side.
Usually lower performance (lower throughput, more total computational resources expended) and more complex to implement.
A protocol (often complex) is needed for dispute resolution, if peers make conflicting claims about the state of data which participants are meant to be synchronized on.
No single point of failure: network can still function even if a large proportion of participants are attacked/taken out.
Coordination is often difficult, as no single agent has the final say in network-level decisions, protocol upgrades, etc. In the worst case, network is prone to fracturing when there are disagreements about protocol changes.
Censorship is much harder, as information has many ways to propagate across the network.
Anyone can participate in the network; there are no “gatekeepers.” Ideally, the cost of participation is very low.