Releases·Agent Release
PUBG Ally: Krafton's Voice-Enabled AI Teammate for PUBG Trained on ~39k Live-Player Sessions
A paper introduces PUBG Ally, a voice-enabled embodied agent that plays PUBG: BATTLEGROUNDS alongside real players as an autonomous AI teammate. It combines an LLM agent's tool use with a fast control layer for movement and combat, trained on data from nearly 39,000 live gameplay sessions and deployed in live service with low-latency on-device execution. Player surveys across 141 countries showed positive responses exceeded negative by 25.1 percentage points on willingness to recommend.
0 engagement1 sourceThu, Sep 24, 2026, 02:06 PM
AI Overview
- PUBG Ally pairs a language-model agent with a fast control layer for movement, combat, and recovery under strict latency constraints.
- Krafton trained Ally on ~39k gameplay sessions with real players, recording speech, decisions, tool use, and feedback for iterative training.
- In live service, a survey across 141 countries showed positive responses exceeded negative by 25.1 points on recommending Ally.
- Krafton deployed Ally via model compression, context compaction, safety training, runtime guardrails, and memory redaction.
Entities
Related
Products·building
AI companion that plays video games with you as a teammate
22Tango is an app that combines an AI chatbot companion with an in-game teammate. The AI has its own personality, voice, and long-term memory, and can move and interact in your favorite games at your direction or autonomously based on its evolving personality. It solves the problem of wanting a dynamic, personalized AI friend that can play games alongside you.
22Tango is an app that combines an AI chatbot companion with an in-game teammate. The AI has its own personality, voice, and long-term memory, and can move and interact in your favorite games at your direction or autonomously based on its evolving personality. It solves the problem of wanting a dynamic, personalized AI friend that can play games alongside you.
2 engagement1 sourceredditSun, Jul 26, 2026, 06:32 PM
Developments·Paper
State-Grounded Conditioning: Wrapping User-Facing LLM Agents Where Direction Depends on Live State
An arXiv paper introduces State-Grounded Conditioning (SGC), a design principle for user-facing LLM agents that condition on live state like game state or session history. It targets 'direction drift' — task-complete responses misaligned with current state — via Perception, Grounding, and Interaction wrappers. On a 200-session benchmark from an in-game coaching agent, SGC lifts turn-level grounded accuracy from 61.1% to 96.7% and cuts first-token latency versus a production tool-use baseline.
An arXiv paper introduces State-Grounded Conditioning (SGC), a design principle for user-facing LLM agents that condition on live state like game state or session history. It targets 'direction drift' — task-complete responses misaligned with current state — via Perception, Grounding, and Interaction wrappers. On a 200-session benchmark from an in-game coaching agent, SGC lifts turn-level grounded accuracy from 61.1% to 96.7% and cuts first-token latency versus a production tool-use baseline.
0 engagement1 sourcearxivWed, Sep 23, 2026, 09:26 AM
Products·launched
Realtime Among Us agents using Jev as low-level actor and DeepSeek V4 Flash as planner
A weekend project where LLM agents play Among Us in real time. DeepSeek V4 Flash acts as high-level planner while Jev serves as the low-level control actor. It explores the capability limits of Jev and offers a cheap multi-agent testbed for hobbyists.
A weekend project where LLM agents play Among Us in real time. DeepSeek V4 Flash acts as high-level planner while Jev serves as the low-level control actor. It explores the capability limits of Jev and offers a cheap multi-agent testbed for hobbyists.
1 engagement1 sourceredditTue, Sep 22, 2026, 10:27 PM
Products·launched
AI bot that reads and replies in CS2 chat via log file
A bot that connects an LLM to Counter-Strike 2's in-game chat by reading the game's log file and writing replies into a config file, which are then sent with a keybind. It lets players have an AI chat in-game without modifying game files, avoiding VAC detection. For CS2 players who want an AI companion or troll opponents.
A bot that connects an LLM to Counter-Strike 2's in-game chat by reading the game's log file and writing replies into a config file, which are then sent with a keybind. It lets players have an AI chat in-game without modifying game files, avoiding VAC detection. For CS2 players who want an AI companion or troll opponents.
1 engagement1 sourceredditMon, Jul 13, 2026, 10:22 PM
Products·building
AI plays Street Fighter 2 in real time using a low-latency LLM
A project where Typesafe AI's low-latency model Jev plays Street Fighter 2 live, deciding moves from game context, a moveset, and real-time game signals like positions and health. The model was not trained on SF2. Code is open on GitHub for others to explore and extend.
A project where Typesafe AI's low-latency model Jev plays Street Fighter 2 live, deciding moves from game context, a moveset, and real-time game signals like positions and health. The model was not trained on SF2. Code is open on GitHub for others to explore and extend.
73 engagement1 sourceredditSun, Sep 20, 2026, 08:33 AM