BBN Technologies

Contribution to AI

Two things the field now takes for granted began as government contract work in Cambridge. The Interface Message Processor, a ruggedised Honeywell 516 delivered to UCLA in 1969, existed to take packets from a host and pass them on, which separated the network from the machines hanging off it and let laboratories with wholly incompatible computers share one fabric. That fabric became the medium through which the AI groups of the following two decades exchanged code, corpora and argument, and Ray Tomlinson's network mail in 1971 — with the @ convention dividing user from host — gave them the habit of daily correspondence at a distance. The language work is less remembered and matters as much. William Woods's augmented transition networks made a grammar into something executable, with registers and tests rather than static rewrite rules, and LUNAR used one to answer geologists' questions about the Apollo 11 samples in ordinary English; the formalism carried most serious parsing work through the 1970s. HWIM, built for ARPA's speech understanding programme, stacked acoustic, phonological, syntactic and discourse knowledge by hand and reached 44 per cent sentence accuracy in 1976 — short of target, and read afterwards as evidence that hand-built rules would not scale. BYBLOS answered that with context-dependent hidden Markov models of phonemes, and IdentiFinder later applied the same statistical logic to finding names in text.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).