Generative Programming Paradigm
📌 The speaker introduces generative computing (or generative programming) as embedding software workflows within AI processes, contrasting it with traditional imperative programming (step-by-step instructions) and statistical machine learning (model creation via input/output examples).
🤖 Generative AI (GenAI) involves two stages: first, using inductive programming to create a generalist model, and second, programming that model using prompting (natural language instructions).
⚠️ Despite excitement, 95% of AI pilots in business fail due to issues arising from using long, natural language prompts for complex control flows.

Challenges with Prompt Engineering
📌 Natural language programming (prompting) is difficult to maintain, easily broken by new constraints, and presents security risks (e.g., vulnerability to password revelation).
💻 The core issue is forgetting decades of computer science and software engineering best practices when using LLMs as generalist computing devices.
🚫 Traditional engineering tools like debuggers, unit testing, abstraction boundaries, and design patterns (like divide and conquer) are neglected in favor of trial-and-error prompt engineering.

Introduction to Project Malaya
📌 Malaya is an open-source Python library designed to bring software engineering principles back to LLM interaction, aiming for predictability, maintainability, composability, and security.
⚙️ Malaya's key principle is moving control flows out of the prompt and into code, resulting in much shorter prompts (ideally 1-2 lines) that are easier to reuse and port across models.
🛠️ Malaya is a way to build agents, not an agent orchestration framework (like LangChain), allowing integration with existing systems.

The Instruct-Validate-Repair (IVR) Pattern
📌 Due to the stochastic and often incorrect nature of LLMs, a first-class concern must be checking and repairing outputs, similar to error-correcting codes in quantum computing.
🔄 The core design pattern in Malaya is Instruct-Validate-Repair (IVR), where an instruction and requirements are given, followed by automatic validation (defaulting to LLM as a judge or code checks) and subsequent repair strategies (like rejection sampling).
🔬 This pattern enforces abstraction boundaries (execution is separate from validation) and enables code reuse, structuring agent creation methodically rather than through guesswork.

Verbalized Algorithms and Efficiency
💡 Malaya allows applying classical computer science algorithms to natural language tasks via verbalized algorithms, where the LLM acts as a specialized oracle for pairwise comparisons or specific decisions.
📊 Using a verbalized merge sort approach (where the LLM only compares pairs) showed huge gains for smaller models, allowing a 1.7 billion parameter model to outperform a 32 billion parameter baseline model in sorting correlation.
🚀 The library incorporates technical efficiencies like Activated LoRA, which optimizes the use of efficient tuning modules by training them to work directly with base model representations, speeding up evaluations significantly when multiple functions are needed.

Key Points & Insights
➡️ Do not put control flow in the prompt; LLMs are poor at following complex language-encoded control structures, which is a primary cause of agent failure.
➡️ Structure LLM interactions using the Instruct-Validate-Repair (IVR) pattern to manage inherent LLM unpredictability with built-in error checking and repair strategies.
➡️ Utilize Malaya to reintroduce software engineering rigor by encapsulating control logic in Python code, ensuring prompts remain short, reusable, and modular.
➡️ Consider using LLMs as oracles within established algorithms (verbalized algorithms) to achieve superior performance and robustness, especially for smaller models.

📸 Video summarized with SummaryTube.com on Dec 15, 2025, 09:15 UTC