Agents of Intelligence

Can AI make software testing more efficient without introducing new risks? In this episode, we explore TestGen-LLM, Meta’s automated unit test improvement tool powered by Large Language Models (LLMs). Unlike conventional AI-driven code generation, TestGen-LLM follows an Assured LLM-Based Software Engineering approach, ensuring that every AI-generated test is verifiably useful, reliable, and regression-free. With successful deployments on Instagram and Facebook, this tool is revolutionizing the way engineers enhance test coverage. Tune in to discover how AI can assist—not replace—developers in writing better tests.

When it comes to less popular knowledge, how should we train AI? Should we fine-tune it or let it retrieve information on the fly? In this episode, we break down a groundbreaking study that compares these two approaches—Fine-Tuning (FT) vs. Retrieval-Augmented Generation (RAG)—to see which one better equips AI models for niche factual knowledge. We also explore a novel approach called Stimulus RAG, which boosts retrieval accuracy without expensive fine-tuning. Tune in to find out which method wins and what it means for AI customization!

Data retrieval is getting a major upgrade! In this episode, we dive into Structured-GraphRAG, a new framework that enhances AI-powered retrieval by integrating knowledge graphs (KGs) with large language models. Using a case study on soccer data, this approach drastically reduces hallucinations, improves accuracy, and speeds up response times by over 98%. Join us as we explore how Structured-GraphRAG is setting a new standard for AI-driven information retrieval.

How well do AI models really think? In this episode, we explore FRAMES, a groundbreaking evaluation benchmark designed to push Retrieval-Augmented Generation (RAG) systems to their limits. Unlike traditional benchmarks, FRAMES assesses factual retrieval, reasoning, and synthesis together, exposing key weaknesses in today’s most advanced AI models. Tune in to discover why even state-of-the-art systems struggle with multi-hop reasoning—and what it means for the future of AI reliability.

Bigger isn’t always better. In this episode, we break down Microsoft Research’s latest AI breakthrough—Phi-1, a 1.3 billion parameter coding model that outperforms much larger models by focusing on high-quality, textbook-style data. Discover how this approach challenges traditional scaling laws, slashes computational costs, and paves the way for more efficient AI development. Tune in as we explore the future of coding AI and why “textbooks are all you need."

AI isn't taking over jobs as fast as we think—it's all about the economics. In this episode, we dive into MIT’s latest research on computer vision automation and unpack why cost, scale, and deployment matter more than just technical feasibility. From small businesses to AI-as-a-service models, we explore what actually makes automation worth the investment and what that means for the future of work.

Generative AI is evolving beyond standalone models into fully functional agents capable of reasoning, planning, and interacting with the world through tools. In this episode, we explore the architecture of AI agents, how they differ from traditional models, and the role of tools like LangChain and Vertex AI in their development. Discover the future of AI autonomy and what it means for industries and everyday applications.

Discover how multi-agent AI frameworks are redefining the fight against hallucinations in large language models. This episode explores how layered agents, guided by the OVON framework, reduce speculative content through iterative refinement and structured data exchange. Learn about novel KPIs, empirical results, and the potential of agentic AI to enhance trust and transparency in generative AI systems.

Explore how MedAgentBench benchmarks large language models (LLMs) as medical agents, moving beyond chatbots to tackle real-world clinical tasks. This episode unpacks the dataset's 100 clinically derived tasks, its FHIR-compliant interactive environment, and insights into the current state of LLM performance. Learn how AI can reduce administrative burdens and improve healthcare delivery.

Explore how cutting-edge AI transforms remote robotic surgery. Using the Informer model, researchers tackle network-induced issues like jitter and packet loss, ensuring real-time precision for Patient Side Manipulators. Discover the role of predictive AI in overcoming latency challenges, enhancing accuracy, and reshaping surgical possibilities with the Tactile Internet.