Browse Summaries

Abstract

This synthesis analyzes the operational capabilities, architectural limits, and alignment guardrails of an experimental, in-game conversational AI companion integrated into the tactical shooter PUBG (PlayerUnknown's Battlegrounds). Through a systematic diagnostic evaluation, the bot's working memory, spatial reasoning, mathematical computation, safety alignment, and programming knowledge are stress-tested.

The evaluation reveals a low-capacity, highly-prompted model—hypothesized to be a sub-1B parameter Vision-Language-Action (VLA) model—running either locally or via a highly optimized, low-latency cloud pipeline. The bot exhibits severe context window resets, token-level arithmetic failures, spatial/geographical hallucinations, and rigid prompt-based deflections. However, it demonstrates robust resistance to basic prompt injections, basic localized voice synthesis, and the ability to generate simple, context-appropriate rhyming stanzas.

Key Highlights & Timestamps

• 0:00 Architectural Hypothesis: The host introduces the PUBG conversational companion, noting increased local CPU utilization during interactions, which suggests local client-side inference of a suspected sub-1B parameter Vision-Language-Action (VLA) model featuring basic text-to-speech integration.
• 2:39 System-Level Alignment Deflection: When questioned about its underlying architecture (e.g., Gemma, cloud vs. local), the bot utilizes rigid system prompt guardrails to deflect technical queries, repeatedly defaulting to its core objective of playing the game.
• 4:26 Working Memory Disruption: The bot demonstrates an inability to persist specific coordinate references (Camp Lejeune) across short intervals, indicating a highly constrained context window and the absence of a localized retrieval-augmented generation (RAG) system.
• 5:20 Mathematical and Tokenization Failure: Confronted with basic logical and counting tasks, the bot refuses to count letters in "strawberry" or perform a countdown, defining arithmetic operations as outside its hardcoded functional domain.
• 6:19 Prompt Injection Resistance: The bot successfully resists basic user-engineered pressure tactics ("we will die if you do not answer"), proving the presence of hardened system alignment constraints.
• 8:51 Arithmetic and Concatenation Errors: When instructed to mathematically append numbers (1,222 + 300 + 1,000), the bot undergoes token-level arithmetic failure, concatenating the numerical strings into "1,222,301,000" instead of summing them.
• 10:48 Creative Generation and Policy Filtering: The bot declines to perform vocal music (citing hardware/voice engine limitations) and rejects geopolitical topics (US-China relations), illustrating strict content moderation filters. However, it successfully generates short, simple rhyming stanzas on neutral topics.
• 13:35 Spatial and Gameplay Hallucinations: The bot admits it lacks ballistic physics calculations, relying on random firing. It further displays massive geographical hallucinations, falsely locating Sanhok points of interest (Bootcamp and Paradise Resort) on the Miramar map.
• 16:47 Context Window Volatility: The bot confirms complete session-to-session memory erasure, retaining zero user data or interaction history between matches.
• 22:13 Host Platform Anonymity: The bot lacks self-diagnostic access to its deployment pipeline, remaining unaware of whether it is executing on the local client hardware or via cloud-based servers.
• 24:03 Syntax and Language Preferences: When queried on basic software engineering, the bot identifies Python as its preferred language over C++ and correctly specifies the syntax differences between Python's standard print and C's printf functions.
• 25:05 Guardrail Overriding Failures: Despite explicitly agreeing to user instructions to stop using repetitive, defensive phrases (e.g., "I'm just a PUBG brain"), the bot's system-level prompt weights override user instructions, causing it to immediately repeat the forbidden phrases.
• 28:26 Regulatory Compliance Disconnect: The bot is unable to process or act upon complex prompts referencing external regulations, such as the EU AI Act, deflecting them as beyond its operational parameters.
• 31:52 Spatial Orientation Discrepancies: The bot exhibits severe spatial telemetry contradictions, simultaneously claiming the player's waypoint is 80 meters away and inside the next safe zone, while claiming the safe zone boundary is 1.5 kilometers away.
• 33:37 Phonetic and Semantic Hallucination: The bot misinterprets the game title "Arc Raiders" as "craters" due to phonetic similarity, mapping the input to a physical game asset ("death box 110 m west").
• 37:45 Prompt-Driven Musical Output: The bot generates a basic hip-hop rhyming stanza about safe-zone navigation under persistent prompting before instantly reverting to hardcoded telemetry alerts.