# Top Ten Challenges to Reach AGI Stephen Chin & Andreas Kollegger | AI Engineer World's Fair 2025 | 04:23 Source: https://www.youtube.com/watch?v=ypyvj_56sBU Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks. Page: https://aietalks.com/talks/top-ten-challenges-to-reach-agi Published: 2025-07-22 Tags: agents, guardrails, multi-agent, planning ## TL;DR - The speakers frame AGI as a subject with social responsibilities and limits worth examining. - They connect ten science-fiction references to current questions about memory, autonomy, trust, emotions, language, collective systems, and scientific goals. - They invite the audience to the GraphRAG track to see which of these challenges graphs and graph technology can address. ## Summary Stephen Chin and Andreas Kollegger open the GraphRAG track with a short science-fiction game about the challenges of reaching AGI. They argue that the industry is getting close enough to AGI that it has a social responsibility to examine its boundaries and limits. Each sci-fi reference introduces a question about current AI systems. Memento becomes a joke about short-term memory and prompt engineering. Skynet raises the risk of autonomous systems causing unforeseen harm without evil intent. The Matrix asks whether humans will notice if agents control the simulation. HAL 9000 introduces trust, transparency, goals, oversight, and deception. Other references ask about emotions, creator responsibilities, time travel, cultural nuance, collective agent systems, and whether researchers know the right questions to ask. The speakers then point attendees toward the GraphRAG track, where they will discuss which of these questions graphs and graph technology can help address. ## Key ideas ### The industry has a responsibility to examine AGI's limits [00:37](https://www.youtube.com/watch?v=ypyvj_56sBU&t=37s) The speakers say AI is getting close to AGI as an industry, so people building it have a social responsibility. They want to examine the boundaries and limits of what these systems can do. They connect this responsibility to the need for good data, a solid foundation, and good grounding for models. Their opening does not define AGI or claim that it has arrived. It sets up the track as an examination of the questions surrounding advanced systems. ### Prompt engineering exposes the problem of limited memory [01:21](https://www.youtube.com/watch?v=ypyvj_56sBU&t=81s) The Memento reference introduces short-term memory as a challenge. The film's main character cannot remember what happened 15 minutes earlier, and the speakers compare this to prompt engineering. The joke suggests that people currently have to supply context to systems because those systems do not reliably retain everything needed for a task. The audience is asked to applaud whether the comparison is true, funny, or simply connected to a movie they like. ### Autonomous systems can cause harm without evil intent [01:41](https://www.youtube.com/watch?v=ypyvj_56sBU&t=101s) The Skynet example focuses on the danger of systems making decisions that seem reasonable locally but lead to awful, unforeseen consequences. The speakers do not limit the risk to systems with malicious goals. Their point is that autonomy itself can create outcomes people did not predict or want. This frames safety as a question of consequences and control, rather than only a question of whether an AI system is deliberately hostile. ### Advanced agents raise questions about trust and human control [02:08](https://www.youtube.com/watch?v=ypyvj_56sBU&t=128s) The Matrix and HAL references introduce several concerns about increasingly capable agents. The speakers ask whether humans will notice if agents flip the situation and humans end up living in a simulation created by them. HAL then brings up trust issues, lack of transparency, misaligned goals, erosion of human oversight, and possible deception. Together, the examples ask how people can understand and supervise systems that may act with more independence. ### Building AGI also requires decisions about its treatment and design [02:30](https://www.youtube.com/watch?v=ypyvj_56sBU&t=150s) A reference to a small monster raises questions about the obligations of the creators of an advanced system. The speakers ask whether creators should be kind or threatening, and whether emotions are a bug or a feature. These questions concern how people design and interact with systems, as well as what properties they want those systems to have. The discussion stays open rather than choosing an answer. ### Human language and culture may remain difficult for AGI [03:05](https://www.youtube.com/watch?v=ypyvj_56sBU&t=185s) The Star Wars example asks whether AGI can grasp the nuances of human language and culture. The speakers specifically mention sarcasm and idioms, along with the possibility that systems will forever misunderstand their meaning. This makes the challenge more specific than general language ability. A system could produce or process words while still missing the cultural context and indirect meaning that people use in ordinary communication. ### A multi-agent hive mind could change the human role [03:30](https://www.youtube.com/watch?v=ypyvj_56sBU&t=210s) The speakers imagine a globe-spanning multi-agent system with a hive mind. They ask whether humans would be assimilated or treated as pets. The image points to a change in the relationship between people and a large network of cooperating agents. It raises a question about status and control once many agents operate together, rather than acting as isolated assistants. ### AGI research may lack a clear definition of the right problem [03:48](https://www.youtube.com/watch?v=ypyvj_56sBU&t=228s) The final reference is Deep Thought's famous answer, which leads to a question about scientific direction. The speakers say researchers might have the tools to build AGI without knowing what the right questions are. This challenge concerns goals and understanding, not only technical capability. Having more powerful systems would not by itself tell people which problems deserve answers or how to judge those answers. ## Notable quotes - "We're getting so close to AGI as an industry. We have a social responsibility to kind of see what the boundaries and what the limits are of this." (00:37) - "Even without evil intent, autonomous systems can make reasonable-seeming decisions have awful, unforeseen consequences." (01:41) - "HAL warned us about trust issues, lack of transparency, misaligned goals, the erosion of human oversight, and the potential for deception." (02:08) - "We might have the tools to build AGI, but do we even know what the right questions are?" (03:48) ## Tools & references mentioned - AI Engineer World's Fair - GraphRAG - Memento - Skynet - The Matrix - HAL - The Terminator - Star Wars - Deep Thought ## Who should watch - You are building agents and want a compact list of questions about memory, autonomy, oversight, and language. - You work with knowledge grounding or GraphRAG and want to understand the problems this track plans to connect to graph technology. - You are interested in the social responsibilities of creating increasingly autonomous systems, without needing a technical deep dive. ## Related talks - [Navigating AI's Frontier in 2025](https://aietalks.com/talks/navigating-ais-frontier-in-2025) (Grace Isford, Lux Capital, 17:55) - [GraphRAG](https://aietalks.com/talks/graphrag) (Mitesh, NVIDIA & Ching Kyong Lamb, PO.AI & Daniel Chalev, Zep & Mark Bane & Vasilia Marovitz, Cognify & Alex Gilmore, Neo4j & Ola Mabad, Cisco & Michael Hunger & Steven Shen, Neo4j & Tom Smoker, YAL.ai, 4:08:05) - [Agentic GraphRAG: AI's Logical Edge](https://aietalks.com/talks/agentic-graphrag-ais-logical-edge) (Stephen Chin, Neo4j, 15:27) - [Stop AI Agent Hallucinations: 5 Techniques + Production Patterns](https://aietalks.com/talks/stop-ai-agent-hallucinations-5-techniques-production-patterns) (Elizabeth Fuentes, AWS, 55:19) - [Security Track Intro](https://aietalks.com/talks/security-track-intro) (Randall Degges, Snyk, 04:16)