I want to float a specific worry and a specific way to test it, rather than announce a finished theory.
The worry is about the growing scope of causal reach. A single human's judgment used to affect a small number of people. A flawed algorithm today reaches millions. Tomorrow, an autonomous system wired into finance, software, and possibly robots and physical infrastructure could act on a wrong objective at a scale no individual could have reached before. AI has no desire of its own, which is exactly the part that should concern us: it is extremely efficient at executing whatever objective it is given, including objectives that encode our biases, our short-termism, or a goal we never fully meant.
So the useful question is not "how smart is this model?" It is "how much causal reach does the system now have, and how carefully has the goal behind it been checked?" Those are different questions, and we tend to answer the first while hoping the second will take care of itself.
The framework I've been building, called DCS, proposes one way to think about this. It treats causality as the thread running through the whole ladder of scale: particles, atoms, molecules, cells, bodies, brains-plus-language, civilization, and now AI systems. The main line, as a hypothesis rather than a settled result, is: causal structure, causal persistence, causal compression, causal emergence, causal prediction, causal intervention.
The part that matters for alignment is the last link, causal intervention. Intelligence, as DCS frames it, is roughly: compress the past, simulate future states causally, then act in a way that changes present conditions to reach a future goal. That is exactly the structure of a goal-directed system. A large language model next-token objective is good at predicting text; it is not, by itself, the same as having a verified model of which interventions actually produce which outcomes in the world. I think the field sometimes slides from "predicts text well" to "understands consequences," and those are different claims.
I want to be careful about what I am and am not claiming. DCS is an open-access preprint, not peer-reviewed. The "roughly a dozen levels" on the ladder is a candidate layering that the framework proposes, not an established scientific consensus. I am not going to pretend the theory is settled. The reason to post it here is the opposite: this community is unusually good at finding where a falsifiable claim quietly stops being testable.
That leads to the most important part of this post: what would make DCS wrong?
If the cross-scale transitions DCS points to cannot be stated as identifiable, coarse-grained macro variables that actually predict and control outcomes better than the underlying micro description, then the "causal emergence" claim fails. There is a real literature here, from Hoel and others, on whether effective information increases at a macro scale. If careful analysis shows no such stable macro-level causal gains in the relevant transitions, DCS should be downgraded or abandoned rather than defended. If the brain-as-predictor story it leans on cannot survive existing neuroscience, the upper rungs break. If "causal compression" turns out to be just a renaming of coarse-graining without predictive payoff, it adds nothing.
The point of building the theory this way is that it is meant to survive attacks. Different AI assistants are assigned to literature search, to logic checking, to arguing the opposite side, and to version control. Propositions that do not survive evidence are supposed to be dropped, not protected. Whether that actually works, or whether it just flatters the author, is itself an open question worth scrutiny.
Why post this on an alignment forum rather than a product page? Because the deeper bet is that one person plus AI might be able to hold some of the cognitive work that used to require an institution. If that is even partly true, it changes who can pose a big question. But it also changes who can point a large causal-reach system at a poorly-checked goal. Capability and responsibility are scaling together, and I do not think our goal-checking has scaled as fast.
Open-access preprint (check it yourself): Zenodo DOI 10.5281/zenodo.22709952, Figshare DOI 10.6084/M9.FIGSHARE.33519052. Public repo: https://github.com/weijierry/dcs-causal-structure-evolution.
There is a public online launch on September 16, 2026 (English session 09:00-11:30 Beijing; Chinese session 19:30-22:00). More at https://mindas.me . Free registration: https://www.eventbrite.com/e/dcs-theory-launch-in-search-of-the-first-principle-of-evolution-tickets-2000726722485
