1) Why unguessable? That sounds like something one would naively use to keep secrets, but encryption is the right tool for that. Apart from that, it sounds like nodes participating in the network would inherently have to see names in order to process requests...
Would something like 256 or 512 bit hashes suffice? You can try to guess or enumerate, but the chances of finding anything are slim.
2) What problems of networked data storage would this end? I see lots of problems, such as discoverability, bandwidth, latency, retention, scaling, censorship, etc. Which of these are solved by globality of the netwokr? Which of these are solved by unguessable addresses?
3) How does this end copyright? AIUI copyright is a social problem, not an engineering problem. If you want to work around it, you would need (at least) strong anonymity and censorship resistance. Freenet is the only thing (that I know of) that comes close, and while it has engineering problems, the main issue with any such project is a people problem: you need huge adoption, otherwise it is impossible to resist deanonymization and offer sufficient bandwidth & storage, etc.
1) Unguessability lets us treat the namespace as uniform and opaque. You are right that encryption is required; see the designs of Tahoe-LAFS or Dat for examples of how to blend these concepts. Another advantage is that an unguessable reference is a basic capability, and in fact the most complex capability that can be used to protect mere data.
You are also right that we don't yet have a satisfactory proof that any of this can happen. It does indeed seem like participants in the network must, as a condition of routing, be forced to handle bundles of data to which they do not have keys, and while mixnets are real, mixnets do not completely solve the problem.
Yes, if we get concrete, the typical way of building unguessable names for data is to do something like take a Merkle tree hash, and then use that as a basis for several "exported" names which are made from further hashes.
2) This would form the basis for a data commons. Existing commons are actually centralized platforms supported by small specialized entities, leading to lots of extra details that nobody is incentivized to get right. By contrast, if there were a single namespace that existed beyond corporate or state control, where participation is platform, then people are incentivized to pay their own way on discoverability (using existing social graphs), bandwidth, latency (using existing compute hosts), retention, scaling, and censorship (using low cost of publication plus low cost of maintenance).
The global/universal nature of the network ensures that, if you can reach it, then it can reach you, and you are connected. IP gives us a hint of this; the typical connection comes with an IP address, and that address can be globally routed. For a real example today, look at how Bittorrent is diverging from the need for trackers.
Finally, it's worth pointing out that decentralized designs might be able to shard computational work or otherwise balance resources. Bittorrent famously was designed to go faster as more peers contribute more spare bandwidth, to the point where the early days of Bittorrent were marked by the protocol chewing up and choking residential ISP connections.
3) Copyright is a social solution to a technical problem in most media. The problem is that publication isn't instantaneous and uniform; there is a delay of time while a published work is copied around the world, and that delay introduces opportunity for pirates to make bootlegs and undercut official releases. On the modern Web, though, this is silly. If one wants to make a simultaneous publication to all paying customers, then one can sell customized encrypted copies at each point of sale, and release a master key which decrypts them all, a day later. The entire window of opportunity for pirates can be shrunken to mere milliseconds, which is impossibly small for pirates to make a profit.
Since we don't really need copyright in order to prevent piracy, then it makes sense to raise the eyebrow and look more cynically at such an intrusive and artificial right. In particular, the proposed namespace grants massive power to publishing artists first.
1) Why unguessable? That sounds like something one would naively use to keep secrets, but encryption is the right tool for that. Apart from that, it sounds like nodes participating in the network would inherently have to see names in order to process requests...
Would something like 256 or 512 bit hashes suffice? You can try to guess or enumerate, but the chances of finding anything are slim.
2) What problems of networked data storage would this end? I see lots of problems, such as discoverability, bandwidth, latency, retention, scaling, censorship, etc. Which of these are solved by globality of the netwokr? Which of these are solved by unguessable addresses?
3) How does this end copyright? AIUI copyright is a social problem, not an engineering problem. If you want to work around it, you would need (at least) strong anonymity and censorship resistance. Freenet is the only thing (that I know of) that comes close, and while it has engineering problems, the main issue with any such project is a people problem: you need huge adoption, otherwise it is impossible to resist deanonymization and offer sufficient bandwidth & storage, etc.