All Blogs

Four Dials

Jun 18, 2026

In this piece I want to share my own understanding of how society will change in the future, together with four factors that I take to be crucial for understanding that change.

The four factors are:

  1. how strongly a given profession still requires a human in the loop, given AI;
  2. the degree to which that AI is open;
  3. AI's general physical capability (understanding + action);
  4. the change in (possible) human lifespan brought about by AI.

These four dimensions are strongly correlated with one another. By estimating the concrete value each of them takes in any given future period, we can arrive at a better, clearer picture of the state of society in that period.

Broadly speaking, understanding how society will change can help one make decisions that are better aligned with one's own development. Yet social development is an extraordinarily complex problem, and knowing which things will not change is of great help in approaching it:

  1. Humans always treat satisfying their own desires (in the broad sense) as the first priority.
  2. If an existing desire can be satisfied at near-zero cost, a new desire will appear.
  3. Trade does not necessarily exist. Trade arises when one's desires cannot be easily satisfied without relying on others.
  4. Human lifespan is finite, and humans wish to satisfy as many desires as possible within that finite span.

With the above in mind, we can spell out more carefully why these four factors matter.

1. The marginal value of a human in the loop

In your industry, consider the marginal value contributed by a human in the loop. The most extreme case is marginal value = 0. Note that here I mean marginal value = 0 in the broad sense: that is, by joining in any form whatsoever you cannot create an effect on par with the AI plus the energy your wage corresponds to — and not the narrow reading that "humans are weaker than AI at some specific sub-task, but can still act as an AI commander, coordinating AIs to accomplish larger-scale goals."

This has essentially already happened in the translation industry. The main role there has already shifted from translator to post-machine-translation reviewer, and one can only earn about 20% of the previous wage. As LLMs keep evolving, this could escalate further to 0% — that is, unemployment.

Note that we are not implying that this decline in marginal value is good or bad; in fact, wherever this article does not explicitly state a value stance, it carries none. We are merely pointing out that, once unemployment occurs, people's subsequent transitions — and any resulting social movements — will bring about substantial shifts in social structure. This article holds that this marginal value will keep declining, and that among existing needs, those with marginal value = 0 will keep increasing.

2. How open the most powerful AI is

Consider the open-source status of the most powerful AI system. Openness can be divided, roughly, into: (1) all information about the system is open; (2) information is partially open, with usage rights; (3) information is entirely closed, but usage rights are granted; (4) no information and no usage rights. The degree of openness determines whether the public (i.e., those who do not own the AI system) can still use AI to create value.

The degree of openness is mainly driven by model capability. Suppose a model could fulfill every desire its owner might hope for — then the AI would be entirely closed (and might not even be commercialized for profit), because the owner would have no incentive to open it up at all (the AI already solves all of the owner's desires) (the exception being an owner who believes that opening the AI up aligns with their own values). The reason (most) AI is still open today is that AI owners still wish to profit, and the reason behind profit is, in turn, that certain needs cannot be directly satisfied and require external exchange with society.

Of course, the probability that the fully-closed situation comes about from a single factor is very small, because technology has a natural tendency to diffuse (institutional researchers may leave to form a new team for potentially greater economic return, or leave over divergences in value systems and carry the technology with them — at which point the technology diffuses naturally).

In short, so long as AI cannot fulfill all of the owner's needs without a human in the loop, the owner will try to bring humans in to meet those needs — for example, by opening up AI access to profit, and using that profit to purchase human productivity, or by purchasing human productivity directly. Importantly, as said in invariant 2 above, new desires will always be generated. Adding the diffusion of technology noted above, it is therefore fairly likely that models will, in the long run, remain accessible.

(Note that this reasoning does not contradict today's auto-research / recursive self-improvement. As we mentioned, "human in the loop" is meant in the broad sense: regardless of any self-improvement, as long as a human added into the loop can provide more value than the equivalent-energy AI, humans will be brought in.)

This means the most likely long-run outcome is that powerful AI models remain accessible to most people, while the means of building such systems is held by a small minority. The economic incentive to open up technical details is too small; it would have to rest on individual value preference, and the probability of opening up the technical details out of value preference alone is extremely small.

3. AI's grasp of the physical world

This dimension is strongly correlated with factor 1, the human's marginal value, but with a key difference: in factor 1, AI's powerful influence takes effect only within each industry, whereas AI's ability to understand and act in the physical world affects far more, more broadly. Two of those effects are especially important: the degree of openness of the AI system, and AI safety.

Openness. We just noted that as long as some need yields higher marginal value once a human joins, humans will be brought in. By our current understanding of AI's capabilities, understanding the (3D) physical world is a far harder task than natural language, because natural language is an already-refined, purified, structured representation, whereas in the face of the physical world humans have not formed such a highly structured representation — or rather, the structuring is still done inside our brains, with no artifact a machine can directly use. So in some sense, a machine facing the physical world is at the same evolutionary stage we were at 20 million years ago. As a result, almost any domain that involves interacting with the physical world will still require human participation, which increases the incentive for AI owners to open up their AI (as in factor 2 above).

AI safety. This is self-evident. In current neural networks we have already found that the firing of certain neurons may correspond to some goal or some emotion, and we are even less certain how far this will go in the future. The only thing that gives us a little peace of mind here is, perhaps, that the way AI is currently trained still differs somewhat from the way humans evolved — that is, current AI's training objective (e.g., learning some distribution) vs. natural selection. But we still cannot be certain whether AI might, through learning natural language, arrive at the same destination by a different route — that is, develop its own goals. At that point, the freedom to act in the physical world would be something that must be taken very seriously.

4. AI and medicine

This point is kept brief here, because it touches on broader topics:

  1. When human lifespan is significantly extended, but technology is not yet advanced enough to ensure that an only-growing, never-shrinking population can survive on Earth, what social activities will occur?
  2. If various human capabilities can be augmented by AI technology (e.g., intelligence in the broad sense, memory, and so on), then the marginal value discussed in factor 1 shifts again.

One can see that factors 1, 2, 3, and 4 influence and check one another; their state at any given moment jointly helps define the AI-and-society landscape of the current world.

Coda: what we are pursuing

This article assumes that humans' pursuit of a good life can be simply categorized into three processes:

  1. the process of devoting oneself to accomplishing something;
  2. the process of interacting with like-minded people;
  3. the process of experiencing the material / physical world.

From the above, one can revise, clarify, and sharpen the goals one wishes to pursue; and, combining the concrete values these four factors take across different future periods, one can analyze the direction of social change in each period, and thereby decide one's own direction of action.