A History of NLP and Wisecube’s AI Journey

In this episode of Gradient Descent, we explore two competing approaches to fine-tuning LLMs: Reinforcement Learning with Human Feedback (RLHF) and Direct Preference Optimization (DPO). Dive into the mechanics of RLHF, its computational challenges, and how DPO simplifies the process by eliminating the need for a separate reward model. We also discuss supervised fine-tuning, emerging methods like Identity Preference Optimization (IPO) and Kahneman-Tversky Optimization (KTO), and their real-world applications in models like Llama 3 and Mistral. Learn practical LLM optimization strategies, including task modularization to boost performance without extensive fine-tuning.

Explore the evolution of prompt engineering in this episode of Gradient Descent. Manual prompt tuning — slow, brittle, and hard to scale — is giving way to DSPy, a framework that turns LLM prompting into a structured, programmable, and optimizable process. Learn how DSPy’s modular approach — with Signatures, Modules, and Optimizers — enables LLMs to tackle complex tasks like multi-hop reasoning and math problem solving, achieving accuracy comparable to much larger models. We also dive into real-world examples, optimization strategies, and why the future of prompting looks a lot more like programming.

Let’s dive into Agentic AI, guided by the "Cognitive Architectures for Language Agents" (CoALA) paper. What defines an agentic system? How does it plan, leverage memory, and execute tasks? We explore semantic, episodic, and procedural memory, discuss decision-making loops, and examine how agents integrate with external APIs (think LangGraph). Learn how AI tackles complex automation — from code generation to playing Minecraft — and why designing robust action spaces is key to scaling systems. We also touch on challenges like memory updates and the ethics of agentic AI. Get actionable insight… 🔗 Links to the CoALA paper, LangGraph, and more in the description. 🔔 Subscribe to stay updated with Gradient Descent!

In the second episode of Gradient Descent, Vishnu Vettrivel (CTO of Wisecube) and Alex Thomas (Principal Data Scientist) explore the innovative yet controversial idea of using LLMs to judge and evaluate other AI systems. They discuss the hidden human role in AI training, limitations of traditional benchmarks, automated evaluation strengths and weaknesses, and best practices for building reliable AI judgment systems. Timestamps: 00:00 – Introduction & Context 01:00 – The Role of Humans in AI 03:58 – Why Is Evaluating LLMs So Difficult? 09:00 – Pros and Cons of LLM-as-a-Judge 14:30 – How to Make LLM-as-a-Judge More Reliable? 19:30 – Trust and Reliability Issues 25:00 – The Future of LLM-as-a-Judge 30:00 – Final Thoughts and Takeaways

In this first episode of Gradient Descent, hosts Vishnu Vettrivel (CTO of Wisecube AI) and Alex Thomas (Principal Data Scientist) discuss the rapid evolution of AI, the breakthroughs in LLMs, and the role of Natural Language Processing in shaping the future of artificial intelligence. They also share their experiences in AI development and explain why this podcast differs from other AI discussions.