Predictions for AI in 2026
2025 has probably been the most revolutionary year for AI dating back to maybe 2022 when the first demo of GPT 3.5 in ChatGPT was launched. LLMs powered by reasoning, agents such as Claude Code, and image generation models like Nano Banana Pro have fundamentally changed how things within most fields are done and how they will be done going forward. As someone who has used AI every single day for the past year or so (in all different kind of tasks), I think this year has made something very clear.
AI is here to stay and change the world.
However, this is not a year in review post (if you are looking for one, at least in regards to LLMs, Karpathy’s is great) but rather one that tries to look ahead. Disclaimer: I am quite far removed from the frontier model development or even research circles, so my predictions are based purely off of vibes and things I’ve read. Additionally, this blog is mainly for fun, so if I am wrong (which I probably will be), don’t be too harsh on me.
(1) Meta will emerge as a leader in the AI race
I want to preface this take by saying that I had this view prior to the Manus acquisition that happened a couple of days ago. That move just further strengthens my confidence in Meta. The reason why I am so bullish on Meta despite them not having any of the top LLMs/agents at them moment is that I believe they are playing the long game correctly. Yann LeCun, who was the Chief AI Scientist at FAIR, has been notoriously stating that LLMs will not lead us to AGI/ASI and that scaling them further, while obviously proven to be useful, only steers us towards a dead end. I actually agree with this take (though unpopular), and a snippet from E.T Jayne that I had read earlier this year supports this.
A false premise built into a model which is never questioned, cannot be removed by any amount of new data - E.T Jayne, “Probability Theory: The Logic of Science”
The reason as to why I find this quote so meaningful and even predictive with relation to the modern state of AI is quite simple: it can both be true that LLMs will change/empower humans to do things never done before at speeds never seen before while also not reaching superintelligence. I’m not claiming that any efforts or resources poured into the scaling of LLMs is a waste, but I do think that at some point, we are going to have to look another way. While I could write a whole separate blog post about just this topic, I’m going to make it brief. I believe that superintelligence will be achieved via vision, not language.
Back to Meta, I believe that in 2026 they will be able to make the huge leap because they are in a position that Google was in back in 2017 with the development of the Transformer. They had all these pieces ranging from attention to residual connections that were gaining traction within the research world, and when put together achieved a revolutionary architecture. Similarly, I think Meta is in a spot where all they have to do is put it together. Think of all the models they have: SAM3 dropped this past month or so, and it has completely changed segmentation. Being able to segment images, video, and even audio with just a simple prompt is a huge improvement to the original SAM. They also have DINOv3 and V-JEPA 2, which I think are actually revolutionary learners in the vision intelligence world. None of these are LLMs, yet they are still re-defining what AI truly is.
The most amazing part about all this is that Meta open-sources these models. While they might not be as far ahead as the frontier labs in certain aspects, it would be naive to think that they won’t get better next year and find a breaking point. Additionally, Meta has the perfect hardware (Meta Ray-bans) to put these models to real-world use and training. Imagine a SAM3 that runs on the smart glasses, transforming what you see into an XR/AR world. This integration alone could put Meta in a position that not many labs could rival. Also, all of these predictions are based off what is publicly available. The recently formed MSL has not released any significant model or breakthrough yet, and I believe this will change in 2026.
I really hope I am right, because I think this investment into vision is the correct play.
(2) Open source will triumph
Do you remember how 2025 started? DeepSeek had just shocked the world towards the end of 2024. What made DeepSeek so popular and industry-shifting was the fact that they a) open-sourced their models and research and b) were innovative in all areas (reasoning, RL, etc.) I think that more and more, labs and people are realizing that open-source is not just a good thing for the industry, but a necessary thing. Agents such as Open Code that are competitors to Claude Code and models like KimiK2, Qwen, MiniMax M2.1, etc. have been revolutionizing what operating in public means. These are not just some tier below the SOTA models, but ones that are openly competing with them across benchmarks and functionality.
To me, this desire for open source traces back to what research is at its core. The foundation of research is placed in sharing your knowledge with the world, enabling it to know more and iterate upon that research to move ALL of humanity further along. And in a world where leaders like Dario Amodei claim that superintelligence could have the ability to solve problems humanity has been dealing with for hundreds of years such as world hunger or cancer, open source will be the main driving force for research. This is made even more clear with Anthropic’s acquisition of Bun (which is open sourced) and many others that have been happening recently. Labs that are focusing on open sourcing their work, and bringing everyone else along the ride have gained people’s trust compared to more closed labs, which is why I think we will see a bigger rise in such.
While I do have more predictions (and goals) for 2026, I think that they are too similar to other people’s (and thus would not be able to celebrate the correctness of such). I will end it off by just saying this:
There is so much to look forward to, because this is the worst that AI will ever be going forward.