Recap: Using AI for 2 years
Technology is mostly an evolutionary field, but things like the World Wide Web, Google Search, and the iPhone were a revolutionary shift in the way we use computers. Using ChatGPT for the first time in late 2022 was also such an eye opener. I wrote some blogs about it early on:
Well, that was the end of 2022, and like most people, AI became part of my daily digital habits. The first thing I noticed after two years is that most people have adopted it quickly, and while using it isn’t really a secret, it’s not really talked about. Almost as if using it at work is a bit of a cheat and something to be ashamed of.
Recently, I sat down with my team at work to share how each of us uses AI in our daily work, and how to use it efficiently and responsibly. I think this is important discussion because AI is far from perfect. For example, when using it as a programmer, contrary to many beliefs, it is more productive to be an expert than a novice. That’s because you often have to cut through AI’s bullshit answers.
Chatbots are often confidently wrong, and you need knowledge, and at least a strategy, to apply its answers. Blindly copy-pasting answers, as some students do, is a recipe for disaster. You need to read the answer carefully, and evaluate and test them to use them efficiently. Also, as prompt engineer, you need to be specific and make the questions small so you don’t mess up big time. This is true for coding, but can be used as a general guideline.
This also has to do with the nature of the technology. The chatbots that most people identifies with AI evolved from automated translation. When translating, the AI algorithm predicts the words based on the input language. This idea was then transferred to questions/answers in the same language.
The main problem with this technology is still reliability. When you use a calculator, you are almost certain that the answer is correct. This is because the calculator uses mathematical rules. AI doesn’t reason in the sense of deduction or by testing the answer against real-world knowledge. A hand in an AI-generated image often has 6 fingers, or a chatbot can’t count the number of letters in a word.
Still, the number of hallucinations has decreased significantly with each generation of models. Also, many of the early problems such as slow responses, technical errors, and availability have largely disappeared. Trusting AI, because you can blindly depend on its reliability, hasn’t been solved yet.
There are a few things you can do to get more reliable answers. The first is to use multiple bots. For many people ChatGPT is synonymous with AI, but now there are many options like Gemini (Google), Claude (Anthropic), Grok, DeepSeek, and PerplexityAI. You can benefit from this competition not only by choosing the cheapest and best model up-to-date, but also by comparing answers.
The second strategy is to use specialized tools. ChatGPT is very generic and works fine for most answers. This is convenient, but if you want a higher quality then it’s still better to use specialized tools. Models that are optimized for a specific field, for example deepl.com for translations, Claude Code for programming and so on.
The third option to improve reliability is to use reasoning, like OpenAI’s O models and DeepSeek do. Reasoning in the context of chatbots is just a fancy term for a model that breaks down your query into smaller pieces and works its way to the answer one step at a time, rather than trying to solve a complex problem in one go. The next step is agents that embed this AI reasoning into your personal data/code.
I expect that in the next few years we will see the most progress in this area of agentic reasoning. For larger goals, I would still recommend (as I did in 2022) that you break down complex tasks into smaller ones yourself. This is something most programmers are already familiar with, but this way of working applies to anyone working with AI today.
In short, AI has become slowly practicable, but you still need to guide it so it doesn’t get off rails, and it won’t cost your more time then you put into.
