Zork

Why MIT built a new parser

Development began in May 1977 at MIT's Laboratory for Computer Science. Behind it were Tim Anderson, Marc Blank, Bruce Daniels and Dave Lebling. According to Wikipedia, they were inspired by Colossal Cave Adventure, but wanted to allow more complex input than that game's two words. Lebling's first version was still limited, close in level to Adventure. The group went on to develop a parser in MDL (a Lisp-like language that ran on a PDP-10 computer). This parser understood adverbs, conjunctions, prepositions and compound verbs, so it could accept a command like "put the lamp and sword in the case". Features were added and the parser was improved through June 1977. The last puzzle was added in February 1979. Dave Lebling also solved a design problem from Colossal Cave Adventure. There, walking in the dark ended with a fall into a pit and death. That solution didn't suit Zork's more varied spaces. Lebling liked the Dying Earth stories by the writer Jack Vance, and took from them the name of a monster that is not described in the stories, the "grue". He turned it into a general danger in any dark place. That is where the famous line comes from: "It is pitch black. You are likely to be eaten by a grue."

From MIT to a company: Infocom

In 1979 Anderson, Blank and Lebling founded Infocom, together with other members of MIT's Dynamic Modelling Group. Joel Berez joined to design the versions for microcomputers, and later became the company's president. The commercial Zork I came out in 1980 for the TRS-80, and soon also for the Apple II. In this way, a parser built in an academic research lab became an off-the-shelf product you could buy. According to Wikipedia, Infocom was officially founded on 22 June 1979 by MIT staff and students, including the senior computing researcher J. C. R. Licklider. On 13 June 1986 Activision bought Infocom for about $7.5 million, and in 1989 Activision shut down the Infocom division entirely.

Still rules, not learning

To be clear: Zork's ability to "understand" a whole sentence was an impressive engineering achievement for its time, but it is still pattern matching, not artificial intelligence in today's sense. The parser breaks the input down according to a rigid grammar of predefined rules. It knows a limited vocabulary, assigns each word to a grammatical category (verb, noun, preposition), and checks whether the sentence fits a known action pattern. It is a distant ancestor of today's natural language processing (NLP) and intent recognition, and it is still rule-based. It is not an early version of them.

2025: Microsoft releases the code

On 20 November 2025, during the week of the Microsoft Ignite conference, two people at Microsoft announced the release of the source code of Zork I, II and III under the MIT license: Scott Hanselman, a senior figure in the company's developer community, and Stacey Haffner, director of the company's Open Source Programs Office (OSPO). According to The Register, the code is written in ZIL (Zork Implementation Language). It can be compiled with a tool called ZILF, and the resulting file can be run in a program such as Windows Frotz. Hanselman said: "I wanted to make the ZIL OSS so students and teachers could use them in classes without concern." The official goal of the announcement was to place "historically important code" in the hands of students, teachers and developers.

Zork as a test for today's AI agents

Back in 2019, Microsoft Research researchers (Hausknecht, Côté, Yuan and Ammanabrolu) built Jericho. It is an open-source environment in which reinforcement learning agents can play 32 text adventures, including Zork I. According to the official Microsoft Research blog, an agent choosing a command of up to four words from a vocabulary of about 700 words faces the task of "choosing from billions of possible actions". Jericho narrows this action space using the game's own internal list of templates: in Zork I it drops from about 240 billion to about 98 million possibilities. Even after narrowing, the results are weak. Averaged over the 32 games, a random agent completed about 1.8% of the game, and the leading algorithm in that work (DRRN) reached only about 10.7%. The researchers wrote explicitly that "much improvement is needed before these algorithms and others can rival skilled humans at these games". In this way the rule-based parser from 1977 became a reference point, showing how far today's learning language agents still have to go. Even earlier, in 2018, Microsoft Research researchers in Montreal built TextWorld, a similar open-source library. Besides running existing text games, it also makes it possible to build new text games by hand, and to generate them automatically, in order to train agents.