Skip to content

Note · November 13, 2026

AI Beyond Chatbots

I can see a major issue with AI in public usage. People still talk about AI through their own narratives. For example, if someone uses only ChatGPT or only a Gemini Flash model, they just talk about what they experienced with that chatbot. But AI is beyond that.

Back in the days, we had automation tools. Automation tools helped us do things much faster than we could, because we often forget routine jobs. Later on, with AI, we have loops, and a loop can do a lot of work. Before, we had to write down all the decisions with if-else logic, like the expert systems we had. For explanation purposes, I call it the if-else system. It was not exactly that, but let's say it was.

After the transformer architecture and the GPTs, everything changed. Next token prediction became the biggest breakthrough in the industry. A lot of people do a lot of things with the large language model APIs from OpenAI, Anthropic, Google or other providers. Right now we can see many large language models and API services, but the foundation is still next token prediction.

By the way, what Anthropic has done to give better control, the tool-based approach, is better than anything else I have used so far. Others try to optimize the language model, but Anthropic optimizes the model and the tools at the same time. The ecosystem is very good, and everything connects well, because they also introduced the Model Context Protocol. That was the foundation that changed how we thought before.

So right now, for graphic design or anything else, I prefer a more programmable, step-by-step way, because programmable things let us change things deterministically. The large language model decides the tool, and I provide the toolbox. I am the toolbox manager, and the manager of all requirements and specs.

For example, if I make a graphic design post, I decide the taste. I am the judge, and I am the problem finder. Sometimes the large language model finds the problem itself, but I can spot it in a very clean way. Then I let the large language model reason about it. When we try to overfit it, we do not get a good output.

That is why I am not a big fan of putting everything into the large language model and just taking the answer. That was the thinking in the early days, but now everything has changed. I am watching each new model, and model by model I get dramatically different outputs, more optimized outputs. So I keep going, and I feel we can move much faster than before.

This note was voice typed, auto-corrected by LLMs, and published by a notes posting agent.

More notes