The Future of AI Agents: Long Context, Benchmarks & Real-World Deployment | ft. Subquadratic

What happens when you remove the token limit from AI agents? On this episode of The Data Layer, host Karla Heredia (Appen) sits down with Alex Whedon, co-founder and CTO of Subquadratic, and Sergio Bruccoleri, VP of Delivery at Appen, to unpack where agentic AI is really heading. They cover: Why transformer attention has quadratic compute costs - and why that's an industry-wide "failing grade" Subquadratic's sparse attention architecture and its path to linear compute scaling Why today's AI agents burn most of their steps just managing context instead of solving problems Long-horizon memory: building agents that remember across weeks, not just sessions What long context means for robotics and multimodal reasoning How Appen independently benchmarked Subquadratic's model up to 12 million tokens Why long-context benchmarks (like the RULER suite) are still immature - and what's missing The "bitter lesson" and why less human curation often beats more engineering Text-to-SQL, tabular data, and why enterprises are sitting on unstructured data goldmines What's next on Subquadratic's roadmap: enterprise design partners, agentic workflows, and vertical specialization Timestamps 00:00 Introductions - Subquadratic & Appen 01:33 The problem with quadratic scaling laws 03:35 How the Appen–Subquadratic partnership started 06:10 Where AI agents are heading in the next 12-18 months 07:21 Why agents waste steps managing context 09:51 Long context and the future of robotics 11:04 In-context learning: lessons from GPT-3 for robotics 14:11 How Appen evaluates models for agentic deployment 17:39 Is long context a feature or a foundation? 21:27 Where long context moves the needle for end users 30:08 Breaking down the benchmark: 98% retrieval at 12M tokens 35:44 How Appen keeps benchmarking independent 38:43 What's next for Subquadratic Subscribe to The Data Layer for more conversations on the infrastructure, data, and research shaping frontier AI. #AIAgents #LongContext #LLM #ArtificialIntelligence #Subquadratic #Appen #MachineLearning #AIBenchmarking #GenerativeAI #FoundationModels