A Brief History of Large Language Models

I've been explaining AI to colleagues and clients in the financial industry lately. Most aren't technical and want to understand what AI means for their work. This is my attempt to tell that story - not the long history of deep learning and AI winters, but the part where AI became relevant for everyday work.

In just three years, large language models went from impressive novelty to useful intelligent companions that can reason and use tools. Here's how it happened.

2022 - ChatGPT

OpenAI launches ChatGPT. At first, it's a small circle of tech enthusiasts playing around with it, making jokes, testing its limits. But everyone is impressed. Something fundamental has changed. For the first time, you can ask any question - it doesn't matter which one - and you get an answer that's coherent. The AI seems to understand you and tries to give you the best possible answer. That feeling, that sense of being understood by a machine, was completely new.

2023 - RAG

The models get smarter. They can now see pictures and listen to audio. But there's still a limitation: they only know what's in their training data. They don't learn anything new while talking to you. A model trained on 2023 data still thinks Biden is president. It doesn't know about your customer database, your internal processes, or last quarter's sales figures.

Experts come up with a solution: if you give the model a document about current events or your company before asking questions, it learns from it on the spot. They call this Retrieval Augmented Generation. This is when businesses start to see the potential - you can give an AI your company handbook, product specifications, or customer data, and it can answer questions about them. For a year it's all anyone talks about.

2024 - Reasoning

Then researchers discover something interesting. If you give a model more time to think, the answer gets better. You can ask it to think through something in a few lines, but also in hundreds of pages. The longer it thinks, the better the answer. Just like humans, really.

For businesses, this means AI can tackle more complex problems. Not just answering simple questions about policies, but analyzing market trends, working through strategic decisions, or debugging complex code. The quality improves with thinking time.

2025 - AI Agents

Now these capabilities come together. An AI agent gets time to think and has access to tools during that thinking process. It can search the internet, check emails, query databases, update CRM systems, generate reports. The parallel with human history is striking: tools also caused an explosion in what we could do.

The models aren't just getting smarter through better training anymore. They're improving through longer thinking and using tools. For organizations, this means AI can handle end-to-end workflows. Instead of just answering a customer question, it can look up their account, check inventory, process an order, and have a confirmation ready.

2026 and Beyond - Learning on the Job

What comes next? Some experts predict the next breakthrough will be about learning on the job. Currently, models always start fresh. They don't know anything about you, your team, or your company beyond what you tell them in each conversation. RAG lets them read documents, but that's still just context - temporary knowledge they use and forget.

The next step might be different: models that learn from their interactions with you. Not just remembering what you said, but changing how they think based on your conversations. A model that learns the dependencies between your systems, understands which data sources are authoritative, recognizes patterns in how exceptions are handled, maps the complex relationships between departments and processes. It becomes better and more useful the longer it works with your company.

It's the difference between a consultant who flies in for a meeting and a colleague who's been working with you for years. For businesses, this could mean AI that doesn't just execute processes, but understands the complexity and interconnections that make large organizations work. The shift from providing context to genuine learning could be as significant as the jump from ChatGPT to AI agents.

What This Means

The shift from 2022 to 2025 isn't just about better models. It's about how we've come to better use their capabilities. ChatGPT was reactive - it waited for your question and responded. Today's AI agents are proactive - they can break down complex tasks, gather information, use tools, and work through problems step by step.

Intelligence isn't just about having more knowledge. It's about having time to think and the right tools to work with. We've known this about humans for centuries. Now we're discovering it applies to this form of artificial intelligence too.