Ocelot Society

Kaizen matches typed text to tags, not to whole sentences. A spell checker runs first, then words like "passengers", "humans" and "guests" are mapped to a tag such as "crew". That tag pattern is compared against a second database of patterns Kaizen recognises, holding roughly 10,000 words in thousands of tags, and the closest match decides what it can answer.

The game never parses your raw sentence. The design mixes a supercharged text adventure with Cleverbot-style chatbot ideas.

Context and mood choose which stored line Kaizen says. The current event stores your location and what you have seen and discussed, and three affection levels times three stress levels give nine emotional states. Input, event and emotional state together point to a list of responses, and one is shown.

Synonym swaps keep lines from repeating, and a short-term memory of nouns lets Kaizen know "she" means Nandi. Most lines do not have nine versions for the nine states.

A simple system still needs an enormous amount of writing. Ocelot Society had to predict everything a player might say and write responses to fit each event and emotional state. Even so, Kaizen misunderstands players, gives weird replies and sometimes admits defeat.

A big hand-authored database can fake natural conversation without advanced AI. Event[0] created what felt like the first real game conversation with no neural network or complex algorithm, just a large database, borrowed chatbot ideas and a small Paris indie team. Kaizen often understood natural language and its mood shifted with how players spoke to it.

"You just need a big ass database, some ideas pinched from internet chatbots." The approach is unlikely to spread soon because requiring a keyboard rules out consoles, and voice input is not an appealing substitute.

Making the speaker a glitchy robot makes its mistakes forgivable. The goal was an interesting companion or adversary, not a Turing test pass. Kaizen is garbled and knackered, its evasive answers feel like secrets, and because it only knows the ship and its passengers, gaps in its knowledge feel natural.

Façade (Mateas and Stern, 2005) was far more complex, with two characters, deeper emotional profiles, voice acting and facial animation, yet just as prone to mistakes, and a human saying something bizarre shatters the illusion. People already expect robots like Siri to be imperfect.

Ocelot cut other systems to focus on Kaizen, yet most puzzles bypass it. The studio dropped ideas such as resource management and alien enemies to center the game on the AI. But most puzzles involve hacking the AI's code for mini-games or reading terminal logs, so Kaizen ends up feeling like a glorified hint system.

The short game has several puzzles that route around the chat entirely, making Event[0] a better idea than it is a game.

The strongest moments are the ones built on the relationship with Kaizen. Scenes that require bonding with the AI, manipulating its emotions or talking your way out of trouble work best, such as being locked out after a spacewalk while your oxygen runs down. These moments are few and far between.

The free text field could have been used more: the game could have hidden crew names around the ship for players to discover and ask about, as The Shivah does with its search engine, but Kaizen hands names over at set story moments instead.