My view on artificial intelligence in 2025
It is now about a year since I published My view on artificial intelligence, and it is time for a new assessment.
Through this year ChatGPT, and eventually Claude.ai too, have been my faithful companions. I have built Feedbach from nothing, completely on my own, on top of everything else that fills up life. These tools have made me more efficient, especially when I have faced problems where I lacked expertise. For example, I needed to make some database changes, and I estimated it would take about five hours of “yak shaving” to get there. Just as I was about to start, I thought I might try ChatGPT. Five minutes later the code was running, and I spent another five minutes testing it. Did I get a productivity gain? Yes and no. I got the job done, but I spent the other four hours and fifty minutes on something completely different from coding. Which is nice too!
How AI assistants have developed
The importance of good prompts
At the start of the year it was important to pay attention to how you worded your prompts to get the best possible result. If you asked for help with code, then after fixing all the compile errors you had to go through the code thoroughly and check for both obvious and more hidden mistakes. Flipped < and > operators could easily happen, and true often became false.
New Kid on the Block: Claude.ai
In the summer I started trying out the cool new kid in town, Claude.ai from Anthropic. Claude was clearly better than ChatGPT 4o at code results, and otherwise about as good as ChatGPT. A few small differences still mean that one can be better than the other in some situations. For example, Claude.ai cannot generate images, and it has no access to the internet either. If what you are wondering about needs input from recent times, ChatGPT is therefore better suited. In return, Claude.ai has a much larger context window than ChatGPT, so if we need AI assistance for large datasets, we can get better results with Claude.
The rise of the ChatGPT model o1
Just before I got around to changing my routines, the ChatGPT model o1 was launched in preview, along with o1-mini, which was specialised for coding. o1-preview impressed me a lot, first and foremost because it could “reason”.
Producing a text and then sending it back to the AI assistant for improvement was a common technique last spring for improving the results. Using the AI assistants iteratively, if you like. This was built in with o1-preview. Built-in quality control, in other words. At the start of the year people talked about the difference between ChatGPT 3.5 and ChatGPT 4.0 being that 3.5 could pass most exams, while 4.0 was on a level with the top 10% of students.
o1-preview, because it quality-checks, is at PhD level, yes, in all subjects!
After using o1-preview for the last few months, I have a trust in the results that almost matches the trust I can have in a human!
Context is King
My experience with the newest models is that the way you phrase things has become less important. In general, that means the AI assistants have become much better at understanding the intent behind what you ask for, even if you have not phrased it entirely correctly. Whether you are polite or not matters less now than it did a year ago, in my opinion (do check out this thread on LinkedIn where I suggest that you should still be polite).
Even though the AI assistants are better at understanding intent, you still cannot expect information that has to be made available to appear by itself. That is why working on context matters so much. If you ask a question without any context, you get a completely generic answer that you could probably find just as easily with a search. The value only appears when you add information. The context window, the amount of information you can send to an AI assistant, typically corresponds to a book of 300 to 500 pages today. That is the content you can feed the AI assistant, and that is used to give a good answer. Think about that next time you ask for help with a professional text and get disappointed because it does not understand you well enough!
If you put a little effort into creating a context for your problem, you will get much better answers. It will probably be the difference between having a very annoying assistant who gives answers you can never trust, and having this PhD-level assistant who understands your intentions and gives precise answers on the first try.
Where do we go from here?
If you compare this year's AI development with the year before, I think it is clear that things are moving faster. We are getting used to seeing AI built into the solutions we use every day. Sometimes it feels useful, and many times it feels forced. That is probably how it will be for a while. Many people talk about solving old problems more efficiently, but I think the changes we are facing will make the old problems less relevant.
AI agents
What happens when we give the AI assistants decision-making authority?
“I need new shoes. Buy something I like.” ...and the day after, they are in your mailbox!
An AI agent is autonomous and can act on its own, and make (certain!) decisions on behalf of the user. They will get to know each and every one of us well, and eventually be able to work out which shoes we like, or maybe even come up with suggestions we cannot think of ourselves!
An AI agent will be able to put together a dinner plan and order the ingredients, while taking into account all the allergies in the family, as well as your ambitions to sleep better and exercise more! If you are worried about pesticides, it will happily choose organic groceries, or products from producers who use less. It can take environmental requirements into account and choose locally produced goods, and it can assess the whole value chain of a product and analyse the environmental impact. These are things most people cannot be bothered to look into.
New problems
So what does it mean if AI agents are going to run various errands for us? They probably do not care much about complementary colours on websites or banner ads in online newspapers. They probably care more that the information is available as structured data, so they can make the best possible decisions.
How do you make the online shop “visible”, and make sure it gets considered, when the users are no longer humans but agents acting on their behalf? Which selection criteria will be decisive?
These are examples of new problems we need to start thinking about going forward.
What else?
It is hard to imagine getting much better language models than the best ones we have access to today. We can of course expect them to be faster, and hopefully less energy-hungry, but the results are already of very high quality in my opinion.
I think advanced language models will become more accessible, and that more alternatives will appear where security and privacy take the lead.
At the same time we have seen big progress in media production over the last few months. Many tools for producing, evaluating and editing audio, images and video have been launched. It is natural to expect these solutions to scale up. Maybe the video rental shops will come back, but instead of borrowing a film, you can build your own tailor-made one, generated in the server farm in the basement!
Many people are probably wondering about AGI, Artificial General Intelligence, where an AI adapts to its surroundings. Today we have AI models that solve the tasks they are trained on really well, but with AGI the model would also be able to solve tasks in areas it has not been trained on. I think we are far from achieving that, but the first step is probably to combine different AI models that can solve tasks within their own specialist areas, and in that way appear adaptable. There is still a big difference between that and real AGI.
Just for the record, I do not think we will get self-driving cars in 2025 either, but we will still be one year closer.