(Thoughts brewed in Jun 2026, can soon be outdated. The views here are personal and biased, they should not be treated as scientifically rigorous arguments.)
As of Jun 2026, you've probably seem some "scary" news and venture investment announcements about developments in self-improving AI. Some people (maybe including you reading this post) might have started wondering "what's the point for me to enter/continue any learning & training programme", especially for some math/stats/AI research directions where the ground truth signals seem to come from pure logical & computational deductions.
Of course I am not against using & improving AI in proper ways - in fact we should think about pro-actively and, again, properly doing it. But I'll just bring up some interesting points, mostly by analogies, about why I think humans still want to do learning & research themselves. This is also not saying we shouldn't use AI during learning/research - none of the thoughts below follow an "AI vs human" in a defensive way. But we need to find the best way to live and work with AI of any form, based on biological, psychological and social factors in human life and society.
- We do human body training even when many of us don't do physical labour work. Our human brain also needs training.
- Many AI4Science research up-to-date suggests using AI models to "warm start" solution searching in scientific discovery. Human's "instinct/gut feeling", gained from learning & work experiences, is analogous to these "model shortcuts".
- Foundation model pre-training learns "generalist knowledge" to support "skills" for many tasks. Human learning of fundamentals (e.g., calculus & linear algebra as foundations for many math/stats/AI advanced topics) is analogous to "pre-training".
- If frontier models need "pre-/mid-/post-training", why not so for human brains?
- Pre-/mid-/post-training differs mainly on the type of data & the amount of compute needed. Same for human researcher pre-/mid-/post-training.
- In many cases post-training works amazingly only when pre- and mid-training are done well. Same for humans.
- All frontier AI labs loop the pre-/mid-/post-training cycles (so no one-off efforts). Why not so for human brains?
- The training, i.e., "weight updating" mechanism, of our human brain is different from that of AI models (more specifically deep neural networks).
- Frontier AI model training empirically follows scaling laws, not only in terms of model size, but also in terms of data & compute. So does human brain training, but with different scaling laws & data filtering mechanisms.
- Current AI model training paradigm largely converges to gradient descent with certain choices of optimisers and hyper-parameter configurations. This choice is validated by many practices. Similarly, a number of human STEM education systems work well and have been validated extensively. E.g., many frontier AI labs have great researchers/engineers receiving foundational training from those education systems.
- The (still largely unknown) biological & psychological mechanism for human brain learning probably won't change that much in the next 20 years. What will change in the future is more about the percentage of usage and the ways of using human & AI brains together.
- Validation metrics change constantly, so does the education system, but certain principles emerge and last long, probably due to "hardware" constraints.
- The ultimate limit of "scaling" is "how to compute better" for both AI and human.
- Computational complexities for AI algorithms and "human algorithms" on different tasks are different. The progress of AI at this stage indicates needs for humans to re-evaluate their own computational complexities.
- Computational costs and constraints come in different forms for AI and human. Human learning helps us build "human algorithms" that work efficiently under our own constraints.
- Human/AI compute, when scaffolded strategically and effectively, leads to "human/AI creativity" - again they may not be the same.
- Not all directions of "(self-)evolusion" will work for AI compute - human and AI are still finding plausible directions for this. Same for evolving human compute -- we, together with AI, should find solutions for this as well.
- Human brain needs entertainment. Best human researchers find proper ways to entertain themselves.
- "Passion" entertains human brain for long term. Human learning & experiences help us identify our own passion (we are not at the stage of AI directly injecting dopamine into our brains yet).
- Shannon's information theory may not work for human brain entertainment: the "mutual information" between external utility and internal satisfaction may not be symmetric.
- Human researchers find satisfactions in both innovation (during research) and impact (after research). One stops being a researcher if for them the former drops to very low percentage.
- Best human researchers tell their own research stories very well. Researchers enjoy personal interactions, good interactions need solid human skills to back-up.
- Existing frictions are big for human society to fully embrace AI. Addressing these challenges should be human-led work.
- Digitalisation is a pre-requisite for embracing AI. Very few countries & regions have this ready. Digitalisation itself might even cause bigger job losses than AI.
- Billions of people in the globe are suffering for reasons that digitalisation & AI alone cannot address.
- Psychological & societal biases (and the diverse form of them) prefer different directions to scale human & AI computes.
- We as humans want to overcome the difficulties for the society to embrace AI in a better way, but we probably don't want the solution to be fully AI-generated. We as humans need to do this work (and train ourselves better to do this work).