KGC23 Keynote: The Future of Knowledge Graphs in a World of LLMs - Denny Vrandečić, Wikimedia

[@kevon217]

‘’it’s complicated”, love it

[@Salfie007]

Fast forward to 2:30 for better audio.

[@ArchitectTomorrow]

Yes thank goodness that got fixed!

[@mandymoo1188]

🎯 Key Takeaways for quick navigation: 00:00 *🎤 Introduction to Keynote* - Introduction to the keynote session by Denny Vrandečić from Wikimedia Foundation. 01:23 *🌐 Challenges and Changes in Knowledge Graphs and LLMs* - Knowledge graphs and LLMs are rapidly evolving, challenging existing paradigms. - Adoption of LLMs like GPT-3 has been unprecedented, impacting various sectors globally. - Researchers and practitioners are adapting to the implications of LLMs on knowledge graphs. 04:03 *⚠️ Narrow Focus and Disclaimers* - The talk specifically addresses the interaction between knowledge graphs and LLMs. - Disclaimers: The presentation does not include AI-generated content and avoids broader ethical and legal implications. 05:13 *🧠 Understanding Knowledge Graphs and Large Language Models* - Knowledge graphs represent relationships between entities, stored in graph databases like Wikidata. - Large language models (LLMs), exemplified by GPT-3, are neural networks trained on vast textual data. - LLMs, despite their capabilities, face challenges in computational efficiency compared to knowledge graphs. 10:23 *💡 Costs and Technical Challenges of LLMs* - LLMs incur high computational costs for both inference and training, posing financial and technical challenges. - Even with optimization efforts, LLMs remain computationally intensive compared to traditional knowledge graph lookup methods. - Industry leaders acknowledge the substantial computational overhead of LLMs. 11:33 *🔄 Evolving Landscape of LLMs* - The pace of change in the LLM landscape is rapid, with indications that the era of large language models might be waning. - Innovations like Meta's LLaMA model highlight the community's adaptability and creativity beyond GPT-3. - Technical limitations, including diminishing returns and cost concerns, influence the direction of LLM development. 12:53 *🌐 Challenges in Information Accuracy and Consistency* - Information accuracy and consistency pose challenges across platforms, exacerbated by reliance on sources like Wikipedia. - Discrepancies in information retrieval from platforms like Google, Bing, and LLMs reflect broader issues in data accuracy and verification. - Language-specific variations in information retrieval underscore the complexities of maintaining accurate knowledge bases. 17:01 *🤔 Limitations and Inefficiencies of LLMs* - LLMs exhibit limitations in handling specific queries, particularly those requiring mathematical operations or nuanced understanding. - The efficiency and reliability of knowledge retrieval through LLMs are questioned compared to structured knowledge bases like Wikidata. - Alternative approaches, such as augmented language models, offer potential solutions to mitigate LLM limitations. 20:09 *🧠 Understanding Knowledge Storage in Large Language Models* - Large language models (LLMs) store knowledge in their parameters. - Parameters in LLMs are essential for tasks like text-to-image generation. - Comparison between the parameter size of stable diffusion and GPT-3. 22:19 *📚 Role of Knowledge Graphs in Text Generation* - Questioning the necessity of vast parameter sizes in LLMs for text generation. - Introducing knowledge graphs as efficient knowledge extraction mechanisms. - Using knowledge graphs to store, curate, and extract valuable information. 23:15 *💡 Significance of Knowledge in a World of LLMs* - Emphasizing the value of knowledge in a world of infinite content generation. - Utilizing LLMs for knowledge extraction and symbolic representation. - Highlighting the importance of overfitting for truth in symbolic systems. 25:12 *🌐 Extending the Expressivity of Knowledge Graphs* - Discussing the limitations of knowledge graphs in terms of expressivity. - Introducing initiatives like Wikipedia Functions to enhance expressivity. - Proposing the introduction of a new special value, "it's complicated," in knowledge graphs. 26:50 *🚀 Enhancing the Future with Knowledge Graphs and LLMs* - LLMs have limitations including hallucinations, expense, and difficulty in auditing. - Knowledge graphs can address these limitations and provide ground truth for LLMs. - The future of knowledge graphs is promising, especially in conjunction with LLMs. 30:08 *💰 Cost Consideration in Knowledge Extraction* - Comparing the cost-effectiveness of using LLMs versus knowledge graphs for answering questions. - Considering whether cost consciousness will impact the hype around LLMs. - Money is a significant factor influencing the adoption and sustainability of LLMs. Made with HARPA AI

[@AEVMU]

Decentralized knowledge graphs deserve more attention.

[@StanleyDenman]

What in the world are you saying!

[@2DReanimation]

22:00: Yes, that's really where KG's would be optimal when integrated into an LLM -- trivia / facts that can't be reduced further than nodes and links in a KG. I mean if you think about the modelling horsepower that would remain in a 170B ANN after outsourcing fact learning and retrieval would be insane.

[@NataliiaLytvyn]

Really make sense!

[@paulina5247]

Such a great presentation! I learned a lot, thank you!

[@infraia]

Excellent presentation Denny!