Let’s admit it upfront: it is extremely hard to predict how AI will evolve in the next, say, 10-30 years, simply because AI has been evolving so fast in the past 10 years. Things that were considered But I’ll try my best to discuss some future directions below.
We all know that the deep learning revolution is all about connectionism at the sacrifice of good old symbolism. Unfortunately, classical AI has mostly died out, partly due to prejudice and ignorance from deep learning researchers. For example, Hinton (“grandfather of deep learning”) famously dismissed those Chomsky-inspired “grammar-based” natural language processing, and other folks even claimed “syntax is not a thing”. As a result, current AI systems are very good at System 1 (fast, intuitive thinking, full of mistakes) but struggle at System 2 (slow and formal reasoning). If we just push deep learning itself, AI will always hallucinate, never be trustworthy, and can never reason reliably well.
A major future direction is improving the reliability of AI reasoning through formal verification, neuro-symbolic methods, and explicit intermediate steps. Models will increasingly be designed to check, justify, and revise their own outputs, reducing hallucinations and improving trustworthiness. This transition will make AI suitable for domains requiring strong guarantees, such as law, science, and engineering.
Current AI systems are heavily focused on text, while other modalities feel more like afterthoughts. Future AI systems will naturally integrate text, images, audio, video, and sensor data into a unified reasoning process. This shift will enable richer understanding of real-world contexts and allow models to respond in ways that reflect physical, visual, and linguistic grounding. As multimodal training scales, AI will move closer to genuine situational awareness.
Google Gemini 3 (released November 2025) is an example moving towards this direction.
A fundamental criticism of current AI LLMs, led by Turing Award winner Yann LeCun (who popularized CNNs in Unit 4) and ImageNet founder Fei-Fei Li, is that current AI is mostly about linguistic intelligence, but our world is much more than language! For example, many animals are highly intelligent but they don’t seem to have language. They argue that current LLMs are not even as smart as a cat when interacting with the real world (e.g., robots still struggle at very basic tasks), although in terms of text-based exams, each mainstream LLM seems to have earned 10+ PhDs and knows everything. They also argue that intelligence comes from interacting with the physical world where AI can learn physical intuitions. For example, a cat doesn’t know any physics, but if she saw a rock falling from the top of the hill, she knows the rock will keep falling and keep accelerating so she’d better run away. If you think about it, the cat can predict the world status in the next second, so she does have some physical knowledge (e.g., Newton’s laws). This is also known as “next frame prediction” or “next world prediction” similar to but much more involved than “next word prediction”.
Therefore, Yann LeCun and Fei-Fei Li and many other AI scientists have been trying to build “world models” instead of “language models”. Once this direction advances, robotics will have a breakthrough.
AI is evolving from passive responders into active agents capable of planning, memory, tool use, and multi-step action execution. These agentic systems will coordinate across longer time horizons, automating workflows and managing complex tasks with minimal human oversight. The research frontier lies in safe autonomy, reliable goal tracking, and structured decision-making.
A good example is AI coding. Nowadays, AI coding assistants are widely used in the software industry and becoming more and more independent. This trend explains why the Silicon Valley is hiring less new grads but more experienced engineers, because easy jobs can be done by AIs.
As exemplified by the 2024 Nobel Prize in Chemistry (AlphaFold for protein folding and deep learning methods for protein design), AI is becoming a core engine of scientific progress, from molecular design and protein modeling to mathematical conjecture and materials discovery. Future systems will assist researchers by generating hypotheses, running simulations, and interpreting complex datasets at scale. This synergy between human intuition and machine-driven exploration may accelerate breakthroughs across every scientific field.
Of course, we should note that AI can not do wet lab experiments yet and probably will not be able to do it in the near future (say, by ~2030), so AI’s scientific proposals still need to verified by human scientists in the wet lab in order to “close the loop”. However, AI can greatly speed up this process by proposing the most promosing designs and experiments, so that they are most likely to succeed in wet lab. AI can also benefit from the feedback from wet lab experiments, fine-tune its models, and propose new hypotheses.
As of 2025, we have already seen quite some AI-directed scientific research (with human wet lab), for example the “Virtual Lab of AI scientists” by Prof. James Zou of Stanford.
Finally, we speculate that AI will self-evolve, because AI is getting better and better at coding, and in the future it can change its own code (rather than just models). Once that happens, AI can evolve 24/7 on itself, much faster than it is now. However, this will also bring safety and alignment issues, so we need to be very careful about any development in that direction.