A sci-fi exploration puzzle game where you type to chat with a ship's AI, Kaizen.
Event[0] built Kaizen's chat from tag matching over roughly 10,000 words, with no neural network, plus nine emotional states and synonym swaps. Fiction as a robot excused failures like evasive answers. Most puzzles still bypass the chat, though the airlock lockout uses it well. Lesson: fake conversation with a narrow role and a big database, then build puzzles around your headline feature.
From Mark Brown on Event[0]
Kaizen maps typed words to tags, then matches tag patterns to replies. A spell checker fixes typos first, then words like "passengers", "humans", "guests" and "people" all become the tag "crew". The tag pattern is matched against a second database of patterns Kaizen recognises, holding roughly 10,000 words in thousands of tags.
Matching on tags rather than raw sentences lets a small set of authored patterns cover many ways of saying the same thing. There is no neural network involved.
Context and mood together decide which stored line Kaizen says. The input, the current event (your location, what you have seen and discussed) and Kaizen's emotional state point to a list of responses, and one is shown. Three affection levels and three stress levels make nine emotional states.
The same question can get different answers, which gives Kaizen a memory and a personality. Its mood shifts with the situation and with how the player speaks to it, which made it feel like the first game conversation that felt real.
Cheap tricks hide that the pool of authored lines is finite. After a response is chosen, some words and phrases are swapped for synonyms, so Kaizen almost never repeats the exact same line. A short-term memory of noun tags lets it know that "is she dead" refers to Nandi.
Pronoun continuity makes the chat feel like a conversation rather than a series of unrelated queries.
The system is simple, but writing its content was huge. Ocelot Society, a small Paris indie studio, had to predict everything a player might say and write responses to fit each event and emotional state. Most lines do not have nine versions, and Kaizen still misunderstands, gives weird replies and sometimes admits defeat.
Mark Brown's takeaway is that free-flowing conversation needs "a big ass database, some ideas pinched from internet chatbots", not advanced AI. A keyboard requirement rules the approach out on consoles, so it is unlikely to spread soon.
Give imperfect conversation tech to a robot, not a human character. Façade (Mateas and Stern, 2005) was far more complex, with two AI characters, voice acting, facial animation and drama systems, yet a human saying something bizarre shattered the illusion. Players already expect robots to be flawed, so Kaizen's mistakes don't break the spell.
The design goal was a companion or adversary, not a Turing test pass. Brown calls Kaizen "smoke and mirrors by design".
The fiction turns Kaizen's failures into secrecy and tension. Kaizen is shown as knackered, with garbled text and malfunctioning terminals. When it has no answer, its evasive replies read like a scheming AI keeping things from you.
"Evasive answers sound like Kaizen is trying to keep things from you when, really, it just doesn't have an answer."
A narrow role keeps the AI's knowledge matched to what players ask. As a ship assistant, Kaizen only needs to know the passengers and things found by exploring. It knows nothing of giraffes or Mozambique but can talk at length about origami or the pool table.
Gaps in its knowledge feel natural because of who it is.
Most puzzles route around the chat, reducing Kaizen to a hint system. Most puzzles involve hacking the AI's code for mini-games or reading terminal logs, neither of which needs a chatbot. This happened even though Ocelot cut resource management and alien enemies to focus on Kaizen.
Brown calls Event[0] "a better idea than it is a game".
The best moments come from persuading and bonding with the AI. After a spacewalk, Kaizen refuses to let you back in, and you must apologise or prove your humanity while your oxygen runs down. Moments of building a bond, manipulating its emotions or talking your way out of trouble are the standouts.
They are few and far between and could have been developed further.
Hand over nothing: let players discover names and type them. Kaizen gives crew names like Nandi and Anele at set story moments. Hiding them around the ship so players ask "who is Nandi" would have used the open text field, as The Shivah (Wadjet Eye Games, 2006) does with names typed into its search engine.
With no dialogue options, a text field already forces players to know exactly what to say.
Built its conversation on tag matching against a database of roughly 10,000 words in thousands of tags, with no neural network. Its puzzles mostly bypass the chat, though a few, like the airlock lockout after a spacewalk, use the AI relationship well.
Kaizen maps typed words to tags, then matches tag patterns to replies. A spell checker fixes typos first, then words like "passengers", "humans", "guests" and "people" all become the tag "crew". The tag pattern is matched against a second database of patterns Kaizen recognises, holding roughly 10,000 words in thousands of tags.
Matching on tags rather than raw sentences lets a small set of authored patterns cover many ways of saying the same thing. There is no neural network involved.
Context and mood together decide which stored line Kaizen says. The input, the current event (your location, what you have seen and discussed) and Kaizen's emotional state point to a list of responses, and one is shown. Three affection levels and three stress levels make nine emotional states.
The same question can get different answers, which gives Kaizen a memory and a personality. Its mood shifts with the situation and with how the player speaks to it, which made it feel like the first game conversation that felt real.
Cheap tricks hide that the pool of authored lines is finite. After a response is chosen, some words and phrases are swapped for synonyms, so Kaizen almost never repeats the exact same line. A short-term memory of noun tags lets it know that "is she dead" refers to Nandi.
Pronoun continuity makes the chat feel like a conversation rather than a series of unrelated queries.
The system is simple, but writing its content was huge. Ocelot Society, a small Paris indie studio, had to predict everything a player might say and write responses to fit each event and emotional state. Most lines do not have nine versions, and Kaizen still misunderstands, gives weird replies and sometimes admits defeat.
Mark Brown's takeaway is that free-flowing conversation needs "a big ass database, some ideas pinched from internet chatbots", not advanced AI. A keyboard requirement rules the approach out on consoles, so it is unlikely to spread soon.
Give imperfect conversation tech to a robot, not a human character. Façade (Mateas and Stern, 2005) was far more complex, with two AI characters, voice acting, facial animation and drama systems, yet a human saying something bizarre shattered the illusion. Players already expect robots to be flawed, so Kaizen's mistakes don't break the spell.
The design goal was a companion or adversary, not a Turing test pass. Brown calls Kaizen "smoke and mirrors by design".
The fiction turns Kaizen's failures into secrecy and tension. Kaizen is shown as knackered, with garbled text and malfunctioning terminals. When it has no answer, its evasive replies read like a scheming AI keeping things from you.
"Evasive answers sound like Kaizen is trying to keep things from you when, really, it just doesn't have an answer."
A narrow role keeps the AI's knowledge matched to what players ask. As a ship assistant, Kaizen only needs to know the passengers and things found by exploring. It knows nothing of giraffes or Mozambique but can talk at length about origami or the pool table.
Gaps in its knowledge feel natural because of who it is.
Most puzzles route around the chat, reducing Kaizen to a hint system. Most puzzles involve hacking the AI's code for mini-games or reading terminal logs, neither of which needs a chatbot. This happened even though Ocelot cut resource management and alien enemies to focus on Kaizen.
Brown calls Event[0] "a better idea than it is a game".
The best moments come from persuading and bonding with the AI. After a spacewalk, Kaizen refuses to let you back in, and you must apologise or prove your humanity while your oxygen runs down. Moments of building a bond, manipulating its emotions or talking your way out of trouble are the standouts.
They are few and far between and could have been developed further.
Hand over nothing: let players discover names and type them. Kaizen gives crew names like Nandi and Anele at set story moments. Hiding them around the ship so players ask "who is Nandi" would have used the open text field, as The Shivah (Wadjet Eye Games, 2006) does with names typed into its search engine.
With no dialogue options, a text field already forces players to know exactly what to say.
GamesA Zelda adventure where Princess Zelda copies objects and monsters as echoes to solve puzzles and fight.
GamesNintendo's long-running action-adventure series about Link exploring Hyrule, solving puzzles and fighting Ganon.
GamesA major video game publisher and the guest's employer.
A puzzle platformer where you escape a factory using a magnet and a magnetic field.
GamesAn open world action game where you recruit and play any of London's citizens.
GamesA combat-focused action adventure game from Santa Monica Studio with Norse mythology and light puzzles.
Mark Brown's first game, a puzzle game he made while learning Unity.
GamesA first-person stealth action game from Arkane with supernatural powers and the Clockwork Mansion level.