A major goal of the Fifth Generation Computer Systems project was the development of high-performance multiprocessors for symbolic applications. The key technologies envisioned as critical were fine-grain concurrent languages and custom microprocessors for directly executing those languages. In contrast to this top-down design approach were most other research laboratories, concentrating on further refining scalable architectures by speeding up communication. One main evolutionary trend was towards threaded architectures that supported latency hiding by fast task switching and low-cost message passing. This article attempts to compare these top-down and bottom-up approaches.