We are in the third year of the generative AI boom. While some are predicting a stall for the industry as a whole, I personally expect 2025 to be a year of interesting, meaningful and useful advances. Generative AI is here to stay.
The advances to come may not always look as exciting as those of the past two years. The hype will probably cool down a bit.
My perspective as I write this: I’ve been writing for a living for 30 years. Generative AI directly affects my profession and therefore my livelihood. I see it as both a threat and a tool.
This dichotomy is always in the back of my mind as I read, think and write about it.
More specialized models
So far, the main contenders in the AI text generation market are general-purpose offerings: ChatGPT, Claude and Gemini can all help with many parts of writing, but they are not specialised in any way.
This may change.
The reason I believe this: I see what is happening in the “open source” AI world (see below). A short time after a new powerful and promising open AI model is released, you’ll see special fine-tuning of it. Often it is optimised for coding, it obeys instructions better, or all restrictions and safeguards are removed.
Commercial vendors might take note.
Claude 3.5 Sonnet, for example, is often praised for its coding capabilities. What if Anthropic saw this and realised an opportunity: Train a version of Claude specifically for this task that is even better! You don’t need a next-generation AI model to do that.
The same goes for writing: I prefer Claude 3.5 Sonnet to any of the other options. In my experience, it’s not even close.
What if Anthropic were to train a version of Claude that understands even better how professionally written texts differ from those written by a layperson? What if they created a Claude who understood the nuances of style and tone even better?
Even without any other improvements, this would be a big step up in quality.
Another example of specialization is reasoning models: They are specifically designed to think through a task before producing a result. This is no longer general-purpose AI. It is expensive, but it could prove very helpful in certain cases.
More specialized services
Putting GPT-3.5 into a chat interface was a stroke of genius. OpenAI famously didn’t realise it themselves. They just saw the number of users suddenly and violently explode. Within two months, the quiet little research lab had become a major company.
But this chat interface is not the best way to use these AIs in all cases. I find it cumbersome for writing, and I hope that people smarter and more resourceful than me are working on new ideas in this area.
The main problem with the chat interface: It mixes the drafting of a text with the discussion of that draft. This is not the way we work in real life: These two parts of the work are separate.
A better metaphor is something like Google Docs. You might have a section where you write and publish the brief for your proposed text. You can add guidelines, source documents, links.
Then the AI takes that and writes its first draft. It presents it to you in a separate environment.
In the next step, you can use a Google Docs-style commenting feature to mark up passages and explain what needs to be changed. The AI rewrites the draft live in front of you.
In other words: I wish there was a platform that worked exactly as if I were communicating with an external writer. But the writer is an AI.
I see similar potential in AI image generators. My main problem with them is that I don’t always know what I want to see in an image to illustrate an article. Sometimes I have a vague idea. But often I don’t.
I think ChatGPT’s interface works well in this respect, because I can explain what I’m looking for. Again, similar to what I just said about text generation: I want AI platforms, services and tools to mimic how we interact with humans. Because if I think back to when I worked in a big publishing house and we had a dedicated person for images and illustrations, I’d tell that person what my article was about. They would already know what style we generally wanted because they knew the target audience and what an issue of our magazine looked like. That’s what I want with image generators: To help me come up with good ideas before or during the image generation process.
While ChatGPT’s chat interface is good for this, its image generator Dall-E is unfortunately sub-par. Others like Ideogram or Midjourney are way ahead in terms of quality. But both put the onus on me to come up with an idea for an image, with a style, tone and more.
I think there will be people out there who see the same possibilities: Put these advanced AIs into user interfaces that are easy to understand and guide inexperienced people through the process.
Sometimes I feel that we are still in the “command line” phase of AI.
“Open Source” gains even more momentum
I know that many people do not consider open source AI to be truly open source. But that’s a discussion for another day.
With that out of the way: Having advanced AI models of many flavours, sizes and capabilities at your fingertips is exciting.
For one thing, these AIs can run directly on a device you already use, even your smartphone. If you need more power, the latest desktop and laptop PCs offer more and more AI power. And if that is still not enough and you work in a larger organisation, you can run your personal AI assistants on your own server.
That way, you are independent of any one vendor. And more importantly: Privacy is no longer an issue because your data and interactions never leave the environment you control. This is a powerful argument, especially for businesses.
On top of that: The open source scene is very active. As mentioned above: When a new interesting model comes out, it doesn’t take long for improved and specialised versions to appear. Companies, researchers and highly motivated individuals can also fine-tune an AI to their liking.
While a vendor like OpenAI offers a variety of turnkey solutions, it will have a hard time keeping up with the variety of open source AI.
I’m not saying that open source will win in the end. There’s a value to convenience and many companies will have no interest in hosting their own AI models.
But the progress I’ve seen with open AI models in the last 12 months has been breathtaking and surprising. I don’t see it slowing down any time soon.
Why I’m skeptical about agents
AI agents are currently the holy grail. The term refers to AI that can perform complex, multi-step tasks without constant human intervention.
At the moment, current AI models always need a human to tell them what to do, ideally step by step. Otherwise they get confused or lost.
AI agents will be much more advanced than that: You would give them a goal, maybe add some tips, tools and documents, and then they go off on their own. When they are done, they report back with the results.
AI agents are potentially a huge industry because they could replace even more people in their jobs.
Today, I see AI tools mostly as helpful assistants. They can make me more productive. But there’s still a lot I have to do myself. For example: When I use an AI to develop an outline for an article, I provide the idea and, in most cases, the source material.
An AI agent would do all that too. I would tell it about my project, the target audience and my goals. It would come up with ideas, research them, prioritise them, look for source material, come up with an outline, write the article.
In theory, many of the building blocks for these agents seem to be in place. They “just” need to be connected in a reliable way.
But as it turns out: That’s the hard part.
OpenAI likes to fan the flames of hype around this topic, especially its boss Sam Altman. In a way, it’s part of his job. OpenAI has just raised a lot of money. They have to keep their investors happy and people interested. So I take their predictions with a pinch of salt. They are self-serving.
I’m not saying they lie.
But I’m not convinced they’ve really cracked the code on AI agents.
I’ll believe it when I see it.
What about AGI, ASI …?
Even more pie-in-the-sky than agents are Artificial General Intelligence (AGI) and Artificial Super Intelligence (ASI). Many people in the industry see these as the end goal of their research.
It would literally be the end goal: An AGI is often defined as an AI that can perform any task that a human can perform at the same level as a human. Combined with robotics, an AGI would be a complete replacement for pretty much any human in any job or position – theoretically.
At that point, an AGI could also become an AI researcher and begin to improve itself. This process would most likely accelerate exponentially. In other words: It would start out slow and then suddenly become very fast.
The result: ASI – AI systems that outperform any human at any task. They would continue to improve themselves to a point that is unimaginable to us humans because, by definition, it is beyond our understanding.
That sounds like science fiction. And I don’t expect it to happen in the next 12 months. But it doesn’t seem as far-fetched as it did a few years ago.
Skeptics say we can’t do it with the means we have today. Others are much more confident and see it happening in the next five to ten years.
More predictions
Because everyone likes predictions (including me!), I added a tag Predictions to this site. Enjoy!