Relevant discussion on the uptick blog: http://wagerlabs.com/uptick/2005/07/very-frustrating-saturday.html The author has rival concerns about speed and memory-efficiency of threads.
My take:
The idea is that there should be a standard API, and then multiple ways of implementing it. Providing a light-weight implementation in terms of green threads (Eric Lavigne just ported erlisp to cmucl), or of a continuation passing transform (Screamer style), are definitely implementation strategies for Erlisp. And if all goes well, we can mix and match implementation strategies within a same Erlisp universe: gimme slower memory-efficient threads for this, faster memory-wasting threads for that, migratable (Tube-like) threads for that other thing, etc.
[ François-René ÐVB Rideau | Reflection&Cybernethics | http://fare.tunes.org ] Proprietary software is dead software. Using it is a form of necrophilia. -- Faré