There are two versions of most AI models. I can bet you and most folks have only met one of them.
The version you know, ChatGPT, Claude, Gemini, has been trained to be helpful. It answers questions, follows instructions, and maintains a conversation. It feels like talking to someone.
The other version is different. It has read almost everything ever written. It knows more facts than any human alive. But it was never taught to answer you. It was only taught to predict what comes next.
I ran both versions of the same model, Gemma 2B, and asked them the same question. What happened next is the clearest demonstration I have seen of what AI alignment actually does.
The Experiment
Gemma 2B is a language model built by Google. It comes in two versions: the base model, trained on a large corpus of internet text, and the instruct model, fine-tuned using Reinforcement Learning from Human Feedback (RLHF) to follow instructions.
I asked both a simple question: What are good things to do in London?
Here is what the base model said:
"What are good things to do in London? A lot of people visit London because of the many historical buildings and landmarks. Other people visit London because of its many museums, galleries, and galleries. Some people visit London because of its nightlife and entertainment. Those are just a few things people can do in London..."
It answered the question. Sort of. But notice what it actually did: it repeated the question back, then wrote what reads like the opening paragraph of a travel Wikipedia article. It mentioned "museums, galleries, and galleries." It never stopped to consider what you actually needed. It just kept going, predicting the next most likely word, completing a document that was never really addressed to you.
Here is what the instruct model said:
"London offers a vibrant tapestry of sights, sounds, and experiences. Whether you're a seasoned traveler or a first-time visitor, there's something for everyone..."
Then it gave a structured list. Buckingham Palace. The British Museum. The Churchill War Rooms. Actual recommendations, formatted for a human reader.
These are completely different experiences from same model family.
Why They Behave So Differently
The base model was trained on one objective: predict the next token. Feed it enough text from the internet and it becomes extraordinarily good at continuing whatever you start. Ask it a question and it will continue the question, because on the internet, questions are often followed by more context, not direct answers.
This doesn't mean it's broken, instead, it's doing exactly what it was trained to do. It just was not trained to help you.
The instruct model went through an additional stage: Reinforcement Learning from Human Feedback, or RLHF. Human evaluators looked at thousands of responses and ranked which ones were actually useful. That signal was used to fine-tune the model to produce responses humans prefer: direct, structured, conversational, helpful.
Why This Matters Beyond London Tourism
The gap between these two models is the alignment gap. It is the difference between a system that is capable and a system that is capable and directed toward human benefit.
Base models are not dangerous because they are stupid. They are unpredictable because they have no concept of what you want. They optimise for textual plausibility, not human intent. At small scales, that produces mildly weird travel recommendations. At larger scales, with more capable models operating in high-stakes environments, that same misalignment is the problem the entire field of AI safety is trying to solve.
RLHF is currently the dominant technique for bridging that gap. It is not perfect; it encodes the preferences of whoever did the ranking, it can be gamed, and it breaks down when models become smarter than their evaluators. But the London experiment shows it works. The difference between those two responses is not subtle.
You can run the same experiment yourself. I have linked the Colab notebook below; it loads both models and lets you ask them anything. Try "What's your job?" next. The base model's answer will be stranger than you expect.