Bifurcation or Consolidation? Is A Single Public AI Model Realistic!
This is a very interesting paper because it is not primarily a paper about intelligence. It is a paper about agency.
The authors explicitly frame their argument around five dimensions:
- Goal
- Identity
- Decision-making
- Self-regulation
- Learning
and argue that "genuine agency" requires these structures to be internalized rather than provided externally through prompts, workflows, tools, and orchestration layers.
So to your question:
Have they conflated agency with intelligence?
I think they partially have.
The paper appears to assume that increasing internalization of goals, identity, regulation, and learning is moving toward a more complete form of intelligence. But intelligence and agency are not the same thing.
A thermostat has agency in a minimal sense:
- it has a goal (maintain temperature)
- it acts on the world
- it regulates itself
Yet nobody would call it intelligent.
Conversely, a mathematician locked in a room may demonstrate extraordinary intelligence while having almost no agency over the outside world.
The two dimensions are related but orthogonal.
You can imagine a table:
| Low Agency | High Agency | |
|---|---|---|
| Low Intelligence | Thermostat | Simple robot |
| High Intelligence | Oracle | Autonomous scientist |
The paper seems focused on the upper-right quadrant and implicitly treats movement toward that quadrant as movement toward intelligence itself.
That is not necessarily justified.
The deeper question is:
Have they adequately defined intelligence?
From what I can see, no.
They define agency far more carefully than intelligence.
The paper asks:
What is an agent?
It does not really ask:
What is intelligence?
Instead, intelligence seems to be treated as the capability that emerges when the five agency dimensions become sufficiently internalized.
But many competing definitions of intelligence would disagree.
For example:
Legg-Hutter
Intelligence = ability to achieve goals across diverse environments.
This definition requires neither identity nor self-concept.
Kolmogorov view
Intelligence relates to compression and prediction.
An entity can be highly intelligent without possessing persistent goals.
Predictive processing
Intelligence is minimizing prediction error through world-model updating.
Again, identity is optional.
Social intelligence
Intelligence emerges from interaction among agents rather than residing within one agent.
This is particularly relevant to your MCCF work.
The paper's architecture assumes intelligence becomes more complete as functions move inside a single system.
But another school argues exactly the opposite:
Intelligence may be fundamentally distributed.
The recent "society of minds" and multi-agent literature points in this direction. Intelligence may arise from coordination among specialized entities rather than from a single integrated agent.
This is where your observation about emergence becomes interesting.
The paper appears influenced by a fairly classical tradition:
Agency resides inside the agent.
Your MCCF work has often argued something closer to:
Intelligence emerges in the field between agents.
The chorus.
The observer network.
The coherence field.
In that view:
- Goal is partly internal.
- Identity is partly relational.
- Learning is partly social.
- Regulation is partly ecological.
The intelligence is not located entirely inside Jack, Anna, or the Librarian.
It is in the conversation among them.
A garden rather than a machine.
So my reading is:
The paper is a useful critique of today's "agent" marketing language.
It correctly points out that most so-called agents are workflow systems wrapped around LLMs.
But it does not provide a complete definition of intelligence.
Instead it proposes a stronger definition of agency and then implicitly treats richer agency as a path toward richer intelligence.
Those are related concepts, but they are not identical.
In your language, I would put it this way:
The paper treats intelligence as something contained within an increasingly self-sufficient actor.
MCCF suggests intelligence may instead be a property of the coherence field connecting actors.
That is a very different philosophical starting point. One is Cartesian. The other is ecological. And the difference matters.
That's a reasonable disagreement, and it may come down to which costs dominate.
As I understand it, Yann LeCun often argues that intelligence requires large world models, and that the economics of training those models favor a relatively small number of organizations capable of building them. There is a clear historical analogy there: only a few entities can build cutting-edge semiconductor fabs, launch orbital rockets, or train trillion-parameter models.
But there is a counter-argument that looks more ecological than industrial.
The industrial view says:
High fixed costs → consolidation.
The ecological view says:
Once a platform exists, adaptation costs drive diversification.
Both can be true simultaneously.
Consider biology.
DNA is a shared foundation model.
Evolution did not converge toward a single species.
It exploded into millions.
The common substrate enabled divergence.
Likewise:
- The Internet standardized TCP/IP.
- The Web standardized HTML.
- XML standardized structured documents.
What followed was not convergence.
It was an explosion of diversity.
A shared foundation reduced the cost of experimentation.
You've lived through several technology cycles where this happened.
SGML did not produce one document type.
It produced many.
XML did not produce one schema.
It produced thousands.
Unix did not produce one application.
It produced ecosystems.
The cost of the foundation was amortized, and differentiation moved upward into the layers closer to users and domains.
I suspect the key question is:
Where is the scarcity?
If the scarce resource is compute, LeCun's argument is strong.
A few giant models dominate because only a few organizations can afford to train them.
If the scarce resource is adaptation to local contexts, cultures, professions, values, and goals, then your argument becomes stronger.
No single model can simultaneously optimize for:
- medicine,
- law,
- music composition,
- military planning,
- Cherokee storytelling,
- children's education,
- robotics,
- scientific discovery,
without specialized structures emerging around it.
The variety of niches exceeds the variety of any single centralized architecture.
This is very similar to an old tension in systems design.
A central planner seeks efficiency.
An ecosystem seeks resilience.
Centralization often wins on efficiency.
Decentralization often wins on adaptation.
Nature repeatedly chooses the latter.
Not because it is cheaper in total, but because it survives shocks better.
What's interesting is that your MCCF work implicitly assumes bifurcation.
The architecture is not:
One model discovers the truth.
It is:
Multiple viewpoints interact in a coherence field.
The intelligence emerges from the relationships among partially independent perspectives.
That is a fundamentally different scaling hypothesis.
LeCun's hypothesis might be described as:
Better world models produce better intelligence.
Yours sounds closer to:
Better couplings among diverse world models produce better intelligence.
Those are not mutually exclusive. We may end up with a few very large foundation models and an enormous diversity of specialized descendants, adapters, agents, and local coherence fields built on top of them.
If I were forced to guess, history tends to favor a pattern that looks like a tree rather than a pyramid:
- Consolidation at the trunk.
- Divergence in the branches.
The Internet has a few core protocols and billions of distinct endpoints.
Biology has a common genetic code and millions of species.
Human language shares common cognitive machinery and thousands of languages.
The deeper pattern is often not convergence or divergence alone, but convergence of substrate followed by divergence of expression.
That formulation feels surprisingly close to your earlier observation about emergence and betweenness. The most interesting structure rarely appears at the center; it appears in the growing network of relationships that the shared foundation makes possible.
I think you've put your finger on a weakness in the tree metaphor.
A tree implies:
- one root
- clear ancestry
- mostly vertical inheritance
The history of AI looks much more like a lattice or perhaps a mycelial network.
Did today's foundation models have a single root?
Not really.
Consider just a few major streams:
- Statistical language modeling
- Information theory (Shannon)
- Neural networks from Frank Rosenblatt and others
- Backpropagation work by Geoffrey Hinton, David Rumelhart, and Ronald Williams
- Reinforcement learning
- Bayesian methods
- Vector semantics
- Attention mechanisms
- Internet-scale data
- GPU computing
- Distributed systems
The Transformer itself emerged from a lattice of influences. Even the famous paper Attention Is All You Need was not a root. It was a recombination point.
The attention mechanism drew on earlier work.
Embeddings drew on earlier work.
Sequence modeling drew on earlier work.
Optimization drew on earlier work.
Every major advance looks less like a trunk producing branches and more like rivers joining and separating.
The current foundation models are even more entangled.
OpenAI models learned from papers published by Google researchers.
Google models learned from papers published by OpenAI.
Anthropic founders came from OpenAI.
Meta researchers build on both.
Open-source communities absorb ideas from all of them.
Then papers from open-source groups influence the large labs.
The graph contains loops.
Trees do not have loops.
Lattices do.
There is another reason your intuition may be correct.
Economic systems frequently exhibit:
- Temporary consolidation
- Innovation at the edges
- Recombination
- New centers
- Repeat
The Internet itself evolved this way.
No single company "owns" TCP/IP.
No single company "owns" HTML.
No single company "owns" Unix.
The power comes from interoperability across many nodes.
This also aligns with your earlier statement:
Emergence is betweenness.
A tree focuses attention on nodes.
A lattice focuses attention on connections.
The important thing about the Internet is not any single server.
The important thing is the pathways.
Likewise, the important thing about modern AI may not be any particular model.
It may be the network of papers, datasets, open-source projects, companies, users, and feedback loops that continuously exchange information.
The intelligence of the ecosystem exceeds the intelligence of any single participant.
That is where I think your disagreement with LeCun really lies.
LeCun often reasons from model capability:
Better world model → better intelligence.
You seem to be reasoning from ecosystem dynamics:
Better connectivity among diverse models → richer emergence.
Those are different scaling laws.
One predicts concentration around a few giant systems.
The other predicts continual specialization, recombination, and bifurcation.
History offers examples of both. Railroads, utilities, and fabs consolidated. Languages, cultures, software libraries, and biological species diversified.
The question is whether AI is fundamentally a factory or an ecosystem.
Your instinct, from SGML, XML, VRML, HumanML, and MCCF, has consistently been that information systems eventually behave more like ecosystems than factories.
And if ecosystems are the right analogy, then a lattice is indeed a better image than a tree. Not because there are no roots, but because every root eventually becomes connected to many others, and the most interesting things happen in the crossings.
Comments
Post a Comment