What Does the Future Look Like?
Setting aside the re-orgs that happen every 3 months in my former company, massive changes happen in 2025. The release of DeepSeek-R1 in early 2025 now feels like a historical event in the last century. The great success of reasoning models with Chain-of-Thought (CoT) brought Reinforcement Learning (RL) back into the mainstream vision of AI. It also drove the development of AI4Coding and AI Agents, making large-scale deployment of LLMs a reality, with significantly boosting of productivity.
Hiring a large team used to be crucial, but now, the first popping question becomes: “Do we need more people?” Indeed, I can just spin up several Codex/Claude Code instances, give them detailed instructions, and they will work 24/7 non-stop. They build whatever I want, including this interactive blogpost in several hours! Their efficiency far exceeds any human being, always follow, no complaint. Working with AI, my biggest worry is whether I have given them enough workload, not wasting the remaining token quota for the day. This is why everyone is experimenting with long working hours for AI Agents. Human attention is forever the most expensive; humans need rest, vacations, and time to zone out or sleep. It is best to minimize human intervention, let the AI find the answer itself, and come back to check on it after a few hours, possible days of work.
For that 20 bucks I paid every month, I must squeeze every drop of value out of it.
This gives me goosebumps: just because of this measly 20 bucks, I have become a filthy and cunning capitalist. And the smartest and wealthiest minds in the world definitely will think about it too.
So, it is clear that, something will happen.
During the crunch time for Llama 4, I often received messages from my East Coast team members at midnight Pacific time. My friends in London were always on, staying up until 4 or 5 AM was routine. But as large models get stronger, the outcome of our hard work is to see that they reach and then surpass our own standard of work.
This is, in a way, the helplessness of being trapped in a Prisoner’s Dilemma.
The “Fermi Level” of Human Society
If the future is AI-centric, do we still need humans?
Consider a simple model of investment and return for a human. Traditionally, the more work experience accumulated (i.e., investment), the stronger the capability, and the greater the return. This leads to a monotonically increasing curve. This is why big tech companies have ladders: the job level generally goes up with years of service and experience.
Now it is different. Ladders have become meaningless, and past experience is irrelevant. Human value has shifted from being evaluated by “the quantity and quality of labor produced by the individual” to “whether one can improve AI’s capabilities.”
The equation now becomes: Human + AI > AI output.
This turns the investment-return curve from a monotonically increasing one into a soft-thresholding curve, that is, initially zero, then increasing only after a certain threshold. At the start, junior individuals are way behind AI, and the supply of AI will only get cheaper. Therefore, over a long period of time, individuals have no economic value. Only when they grow beyond the point that they can assist AI to become stronger, the economic values start to take off.
Moreover, after that, the boost talented individuals give to AI will be far higher than that of average people. Average people might spend time patching up one or two specific outputs of the AI. Top talents instead may propose systematic and universal solutions (e.g., fine-tune a model, or improves agentic frameworks). Combined with various resources at hand (GPUs and data, etc.), they can make the AI further stronger. As AI is widely deployed, this effect will be amplified geometrically. The “One Against a Thousand” trope from novels will soon become reality.
Under such a polarized investment-return model, if we treat “Human + all accessible AI” as a single agent, its capability distribution looks very much like the distribution of electron energy levels in materials:
- Below the level: Agents at or below this waterline are everywhere, begging for work to prove they are still useful.
- Above the level: Agents above this line are exponentially rare; acquiring and leveraging them is incredibly expensive, and they are often fully booked.
This waterline is the height of the “AI flood”; it is the “Fermi Level” of human society. Jobs and occupations below the Fermi Level might be disrupted overnight. Like a flood or an earthquake, it has been peaceful and quiet for centuries, and the next day the entire industry is wiped out.
As time changes, this waterline will move up. The velocity is proportional to the amount of data available that shows stronger behaviors than current model. If there are no massive breakthroughs in the training process of large models, then similar to autonomous driving, the better the performance becomes, the less useful data there is, and progress will slow down. The top tier of people will be able to maintain their moat for a long time. However, if there is a breakthrough in training, for example, finding new methods for synthetic data or new training algorithms that are way more efficient and achieve human-level, then all bets are off.
Of course, the above judgment assumes an infinite supply of GPUs and energy. Various resource shortages do happen, including shortages of energy, chip and memory manufacture. Whether Earth can satisfy the insane demand for AI from humans is still an unknown. Digging deep into this aspect could probably lead to a full-fledged study, which is way beyond the scope of this blogpost.
Independent and Active Thinking in the Era of “Omni Genies”
So, what happens next?
The future may no longer be like the traditional stories, in which our main character has a Wish, and endures to find the one and only Aladdin’s lamp, or collect the seven Dragon Balls, in order to make the wish come true. Instead, this will be an era of “Omni Genies”. Every AI agent is like a genie; they have superb capabilities and are eager to grant others’ wishes to prove their own usefulenss.
In this environment, what is truly scarce is not the ability to grant wishes, but the “Wish” itself, and the persistence to turn that into reality.
However, in this era of abundant AI capability, huge convenience often comes with huge traps. Large models provide extremely cheap thinking results. In a market where information exchange is not yet sufficient, these results can even be used directly to deliver work and gain economic value (e.g., those copywritings that reek of “AI flavor” at a glance). This readily available convenience will trap many into a lack of motivation to think. Over time, they lose the ability to create originally, and their minds are hijacked and assimilated by generative content and recommendation systems. This is the new era’s definition of “lazy people”: not physical laziness, but mental leisure: no time to think and no ability to conceive unique things.
Eventually, one becomes an empty shell, losing even the ability to make a wish.
So how do we maintain independent thinking? How do we avoid being assimilated by AI?
Tactically, we need to learn to constantly scrutinize AI’s answers, pick at its faults, and find new problems it cannot solve. Future new value may come from several aspects: (1) New data discovery. (2) New and deep understanding of problems that people haven’t thought about, and (3) New paths to solve the problems, including feasible innovative solutions and their results. This is good for long-term development of a person. In contrast, selling AI response may give short-term economic values, but as AI becomes well-recognized commodity, such opportunities will vanish. If one holds a position that only requires working on tasks assigned by supervisors/managers, without too much out-of-box thinking, then the position may be easily replaced by today’s AI flood.
Take AI Coding for example. After using it a lot, I feel that while it can quickly churn out a runnable code repository to meet requirements, as the repo gets larger, the pile of technical debt gets higher, local fixing of the codebase it provides becomes less satisfactory. Humans still excel at the overall framework design and planning. How to harness the coding agents to achieve one’s long-term goals faster will become part of the economic values.
In the long run, whether voluntary or not, everyone will face a transition from an “employee” role to a “founder” role. Working for others is no longer valuable; people will need to find their own “Sense of Purpose.” If there is a firm goal in mind, and a willingness to take all measures (including LLMs as core tools) to achieve it, then active thinking is a natural result. The grander the goal, the more active the thinking, and the greater the potential unleashed.
Therefore, if you heard that your child aims to hold a concert on Titan, the largest moon of Saturn, or wants to explore near the horizon of a black hole, never shy away from such seemingly absurd ambitions. Because that grand wish might be the fundamental source of lifelong motivation and active thinking, and the key to standing firmly above the “Fermi Level.”