Suppose you had an operation that was guaranteed to either add your input to some location in memory or fail gracefully in some fixed amount of time. It's first come first served and implemented at a hardware level. Now this fits with the definition of lock free in that a process can't prevent your process from finishing, and works just fine if you want say a counter of number of moves examined by your AI thread as a thread can always keep an internal counter and try again say after the next batch goes though. However, you still get into resource starvation issues depending on how it's used and with enough threads you encounter a wide range of issues.
Lock free programming is not a binary condition so much as hierarchy of conditions. Where various types of resource starvation (ex: at the cache or buss level) are perhaps the most stringent. Scale up to say a billion cores and 100 billion threads and many 'safe' operations create significant issues.
PS: Wikipedia uses Wait-freedom > Lock-freedom > Obstruction-freedom all under http://en.wikipedia.org/wiki/Lock_free
The author's definition does not allow one to use locks to make individual operations atomic. A process could always crash after obtaining the lock, but before releasing it, causing the entire application to become stuck.
With the correct protocol you can even write to that shared memory. For example a sudoku solver where each tread looks for contradictions and then overwrites the shared states with their findings. This works because there is no contradictions, dirty reads are ok, and there is no contention with all threads writing 0's and nobody writing 1's. You can go even further where any thread can flip the error flag if a contradiction shows up or solution flag if a solution is found, but again it works because nothing clears those flags.
PS: The article's author uses a less strict defintion of lock free where it's ok to use locks for individual atomic operations like counters. But, that still creates performance issues at scale.