The Executive Guide to Large Language Models

Quick summary

A non-hype briefing for leadership on how LLMs are built, how they fail, and where to focus your AI investment.

At their core, Large Language Models (LLMs) are advanced prediction engines that generate text by calculating the probability of the next word in a sequence. While they are built on complex neural network architectures called Transformers, the industry's success with AI currently relies far less on the specific architecture and much more on massive data processing, rigorous evaluation, and massive computational scale.

To understand how these tools can drive efficiency across your company, it helps to understand the two main phases of creating them: Pre-training and Post-training.

1. Pre-training: Building the Foundational Knowledge

During pre-training, the AI is fed an astronomical amount of data — essentially the entire internet — to learn grammar, facts, syntax, and context.

• The Data Engine. Building an LLM requires downloading hundreds of billions of web pages (often over a petabyte of data). Raw internet data is messy, so companies heavily filter out toxic content, remove duplicates, and prioritize high-quality information.
• Strategic Data Weighting. Developers intentionally feed the model specific types of data to build desired skills. Training on computer code, for example, has been shown to significantly improve the model's general logical reasoning.
• Tokenization. Models do not read text like humans do; they read tokens (chunks of characters). This is why models can struggle with tasks like complex arithmetic or precise coding formats — an important constraint when designing business applications.

2. Post-Training (Alignment): Creating the AI Assistant

A pre-trained model only knows how to mimic the internet — ask it a question and it might just generate a list of similar questions. To make the AI useful for your employees, it must undergo alignment to become an instruction-following assistant.

• Supervised Fine-Tuning (SFT). Developers use highly curated, human-written examples to teach the model how to format helpful, accurate answers. This is where the model stops being a predictive text generator and starts behaving like a knowledgeable worker.
• Human Preference Optimization (RLHF). Human testers rate the AI's answers; the model learns to maximize responses humans find helpful and to minimize harmful, toxic, or hallucinated output. This is the brand-safety control surface for the enterprise.

3. Scaling Laws and the Cost of Compute

A fundamental rule driving the AI industry is scaling laws: increasing the amount of training data, the size of the model, and the computing power reliably and predictably improves performance.

• The Investment Reality. Training top-tier models requires massive data centers with tens of thousands of specialized GPUs running continuously for months, costing tens of millions of dollars.
• Inference Economics. Running these models is also highly resource-intensive, which is why optimizing how the AI uses hardware memory — and how much context it can hold — is a major focus for every serious player.

What this means for your organization

Fine-tuned, post-trained AI assistants can automate reasoning tasks, synthesize large documents, and generate code or structured text aligned to your specific business guidelines. The real leverage is not the model you license — it's the data governance, tool alignment, and operating model you build around it.

This presentation is based on the technical insights of Andrej Karpathy, a founding member of OpenAI and former Director of AI at Tesla. The core concepts are derived from his landmark lecture, "Intro to Large Language Models," which has become a foundational resource for understanding the engineering and evolution of modern AI. You can view the original presentation and explore his extensive work on his YouTube channel.