Right Now: And 6 AI Predictions For The Next Few Years

Forbes

03-04-2025

AI (Artificial Intelligence) concept. Deep learning. GUI (Graphical User Interface).
First of all, let me say that it's extremely valuable to run across articles that don't just make vague and general predictions, but actually talk about the internal workings of LLM models and how they are going to progress throughout the next few years.
Even non-tech people know, intuitively, that AI is getting a lot better quickly, but people with an interest in the technology itself want to know how this is happening.
I'm going to present six things that I've noticed as concrete predictions about model development from about 2022 to 2027. A lot of this I got directly from a LessWrong post ostensibly authored by someone named L Rudolph L that really drills down into what we're likely to see in upcoming quarters, based on what we've already seen the past few years. But other sources talk about these trends, too, and I've noticed them coming up again and again as leaders plan for what's ahead.
Early machine learning systems were often almost entirely supervised. There was the idea that with data labeling and other forms of direct supervision, you get a better result and more targeted responses from a model. But then the models got better.
DeepSeek's announcement, for one, leads companies away from this kind of traditional supervision toward something called supervised fine-tuning. Combining this with reinforcement learning, systems can become more autonomous and able to act without detailed human prompting. That's a big part of what's rocking the tech world right now.
Another thing that is presented in the above piece is the idea that the curation of model results will produce even more targeting, which the author colloquially calls a 'magic sauce' for AI.
Again, this substitutes a new process for traditional analysis of weighted inputs, prompting and supervision.
I came up with a very basic three-stage process detailing how robust AI works based on reading the above article and others.
It looks like this:
What does that mean?
It's the prediction that models will be able to take any kind of input - text, media, image, video, and analyze it with cognitive reasoning to figure out what to output in any of those above forms. That's a few steps beyond what we have now, and it's going to be part of the massive revolution that happens in 2026 and beyond.
Another general observation is that some companies will focus on enterprise, while others will focus on consumer products. L Rudolph L suggests, for example, that OpenAI will have the edge in consumer applications, because ChatGPT was the first model of its kind to become a household name. I heartily agree with that, and it's likely that other contenders will exclusively focus on enterprise and leave consumer-facing products alone.
Here's one that's also extremely interesting to me based on some of our brightest experts predicting the kinds of human-computer interactions that look and seem authentic to the human user.
L Rudolph L says it this way, which I thought was a pretty good way to introduce what's likely to happen next year:
'In particular, OpenAI has finally almost caught up to Claude's personality level. It is also way more impressive to normal people because it can also—if prompted to do so—generate real-time video and audio of a talking face. OpenAI doesn't explicitly encourage this, but winks at this, since it knows this will get some users addicted (especially as they now have a more nuanced policy for sexually explicit model outputs than the previous blanket ban).'
Notwithstanding that parenthetic disclaimer, it's natural for humans to want to have human experiences with AI. So we'll see where all of this goes.
While we've heard a lot from big names about the value of open source, some predict that the biggest models will continue to be internal. This crops up in L Rudolph L's article too.
'(Anthropic) comes up with an upgraded version of computer use that is far faster and more seamless,' the author predicts, also suggesting that OpenAI will end up using Claude. 'Again, the largest model is kept internal. Its training data curation and post-training finetuning was focused on programming, ML research, MLOps, and maths. Anthropic employees started adopting it internally in mid 2025, giving researchers and engineers what's essentially a team of AI interns to manage. They then spent 6 months giving the models tailored feedback, which they massively boosted with methods for dataset augmentation, and filtered for correctness with scalable oversight techniques like debate, before feeding it back into the model as finetuning data.'
That's my list of takeaways from doing a little light reading on AI predictions as they stand right now.
It's easy to say that 'we'll have AGI' or 'we won't have AGI' but this really puts the entire process under a magnifying glass, and shows us how these advances are likely to unfurl, again, seemingly at lightning speed. All of this is so new – you can think back to the Covid years as a vast medieval era of pre-AI life, And then look to the future to see, in the parlance of Monty Python, something completely different.