The detailed version of my oped in Transcontinental Times
Every species that has ever competed for resources on this planet has
done so with roughly the same toolkit: strength, speed, camouflage, venom,
numbers. Humans are unremarkable on most of these axes. We are slower than a
cheetah, weaker than a chimpanzee, blinder in the dark than an owl, and worse
at smelling danger than almost anything with a snout. What we do better than
any other species, by an order of magnitude no other animal comes close to, is
cooperate with strangers at scale.
A wolf pack cooperates. So does a beehive. But a wolf will not lay down
its life for a wolf it has never met from a pack three mountains away,
coordinated by a shared story none of them witnessed. Humans do this routinely.
We show up at war memorials for people we never knew, buy shares in companies
run by executives we’ve never met, and hand savings to banks based on nothing
but a shared belief that the institution will honor its promises. Yuval
Harari’s core insight in Sapiens, that humans are unique in our capacity to
organize around shared fictions: nations, currencies, corporations, human
rights, describes exactly this. None of these things exist as physical objects.
They exist because enough people agree to act as if they do, and that agreement
is what lets seven strangers organize a supply chain across four continents.
That capacity has a name in organizational theory: institutionalized
trust. And it has a delivery mechanism: communication, first spoken, then
written, then encoded into the procedures, contracts, and bureaucracies that
let a stranger in Rotterdam trust a signature from a stranger in Chennai.
Bureaucracy earns a bad reputation as red tape, but at its root it is a trust
technology, a set of standardized procedures that lets unrelated people
transact without needing to personally verify each other’s character.
Double-entry bookkeeping, the joint-stock company, the postal system, the
passport, these are all inventions in the same category as language itself:
tools that let cooperation scale past the roughly 150 people (Robin Dunbar’s
famous number) that our brains can track through personal relationship alone.
Physical technology has always ridden alongside this social technology,
amplifying it. The wheel didn’t just move goods faster, combined with
standardized axle widths and road networks, it made regional trust networks
viable at distances no courier on foot could sustain. The printing press didn’t
just reproduce text, by making the
Bible, and later pamphlets, newspapers, and scientific journals, available to
anyone literate, it broke the monopoly that scribal elites held over what
counted as agreed-upon truth, and in doing so it re-founded whole religious and
political orders. The telegraph, the telephone, the internet: each one is,
underneath the marketing, a new trust-and-coordination layer stacked on the
ones before it.
Now comes artificial intelligence, and it is not just another rung on
that ladder. It is different in kind, because for the first time the tool
doesn’t merely transmit human-generated trust signals faster , it can generate judgment, synthesis, and
decisions on its own. That changes the question. It’s no longer just “how fast
can strangers coordinate” but “who gets to do the coordinating, and on whose
behalf.”
There are two directions this can go, and they lead to very different
civilizations.
The first direction: the
Borg model
A small number of frontier labs - right now, realistically, a handful of
companies in two countries - train models on a scale of data and compute that
nobody else can replicate: the entire searchable internet, increasingly private
data through partnerships and acquisitions, and enough proprietary usage logs
from hundreds of millions of daily conversations to know how people think,
argue, and decide better than the people know themselves. Everyone else becomes
a client, querying a central intelligence that has never revealed what it
learned about them to anyone but itself. It is not an accident that the Star
Trek Borg is the right metaphor: a distributed set of drones, individually
unremarkable, whose intelligence is aggregated upward into a Queen who alone
sees the whole picture and alone decides. Assimilation doesn’t require malice.
It only requires that everyone’s data flows one way upwards , while judgment flows back down as a service.
The second direction:
diffusion
Instead of one model trained on everyone’s data and queried by everyone,
imagine a nested architecture of intelligence that mirrors the way trust itself
has always scaled in human societies - from individual to family to community
to nation, each layer adding coordination without fully surrendering what came
before it. A personal model that learns primarily from an individual’s own
history and stays substantially theirs. A household-level router that
reconciles the family’s shared needs, finances, schedules, health, without exporting
the raw data to any central party. Community and institutional layers that pool
just enough signal to coordinate like a hospital network sharing anonymized
treatment outcomes, a farming cooperative sharing yield data and so on without
surrendering the underlying record. National or civilizational layers that
federate further still, for the genuinely public-goods problems: pandemic
response, climate modeling, financial stability. Intelligence increases with
altitude, but so does the friction required to extract raw data upward. This is
closer to how evolution actually organizes complexity - through modular, semi-autonomous units that
coordinate without fully centralizing control - than the single-brain model the
Borg represents.
The diffusion model is the one worth betting on, but it is worth being
honest about why it isn’t automatic. Model weights being “open” or a chatbot
running locally on a phone does not, by itself, redistribute power. Underneath
even the most local-feeling AI product today sits a stack that is still
extremely concentrated: pretraining compute that only a few labs can afford,
chip design and fabrication controlled by a handful of firms in a handful of
countries, and energy infrastructure that is itself a scarce, geopolitically
contested resource. An open-weight model trained on a closed, centrally-scraped
corpus is diffusion in name and centralization in substance , a longer, more
comfortable road to the same Queen. If this century’s version of the printing
press turns out to require a printing press factory that only three governments
can build, the diffusion story collapses into the Borg story with better
marketing.
There is, however, an answer to the training problem, and it comes from
the closest analogy available: how humans themselves acquire capability. Every
human being is “pretrained” on a broad common corpus - language, schooling, the
accumulated knowledge of a culture - before developing anything distinctive.
Universal education does not centralize human intelligence; it equips each mind
to then learn recursively from its own experience, in directions no curriculum
planned. The base model can play the same role: a common endowment, trained
once on broad public data, the way a public education system is funded once for
everyone. Diffusion becomes real at the point past that endowment , when each
node in the hierarchy, whether an individual, a household, a firm, or a
community, holds not merely a copy of the model but the capacity to keep
learning from what it alone can see, and when the owner of that node decides
what portion of the learning is exposed upward. This mirrors how capability has
always worked in human society. A doctor shares her diagnosis, not the decade
of pattern recognition behind it; a firm sells its product, not its process
knowledge; a family teaches its children things it would never publish. Skill
and disclosure have always been separable, and that separability is precisely
what a query-everything-through-the-center architecture destroys - the center
learns from every interaction, while the individual accumulates nothing that is
durably theirs.
Honesty requires admitting that this recursive-learning-at-the-edge
capability does not fully exist yet. Most of what is marketed today as
personalization is retrieval: the model consults an individual’s documents and
history at query time without changing itself, which means the accumulated
learning still lives wherever the model lives. Genuine local learning, models that update themselves from experience,
cheaply, on modest hardware, without catastrophically forgetting what they
already knew remains a hard, open engineering problem. The technical
trajectory, to be fair, is bending in the right direction: models keep getting
smaller for a given level of capability, techniques for cheap adaptation keep
improving, and consumer chips now ship with dedicated neural hardware as a
matter of course. What is not bending is the commercial trajectory. The
economics of every frontier lab reward keeping the learning loop at the center,
because centrally accumulated learning is the moat, the more the central model
learns from everyone, the harder it becomes for anyone to leave. So the two
trajectories diverge: feasibility is diffusing outward while deployment keeps
concentrating inward, and it is precisely in that gap that policy has work to
do. Recursive learning at the edge will not be handed down by incumbents whose
business model it undermines; waiting for the market to deliver it is like
waiting for the scribes to distribute the printing press. It has to be pulled
forward deliberately by public research
funding, by procurement rules that require publicly purchased AI systems to
support local learning and owner-controlled disclosure, and by writing the
principle that the learning stays with the learner into data protection law,
the way purpose limitation was written in a generation ago.
So the honest version of the bet is not “small models will save us.” It
is that diffusion has to be built deliberately, at every layer of the stack,
the way earlier trust infrastructure was built deliberately, through standards,
law, and public investment, not left to emerge on its own from a market that
has every incentive to concentrate.
There is precedent for exactly this kind of deliberate construction, and
it is worth pointing to because it already exists rather than remaining
hypothetical. India’s approach to digital public infrastructure - a unified
payments protocol that any bank or fintech can plug into rather than routing
transactions through a single dominant platform, a verifiable-credentials
system that lets individuals hold and share their own documents rather than
surrendering them to a central database, and an open commerce network that lets
buyers and sellers transact across competing apps rather than being locked into
whichever platform got there first, is
essentially an attempt to build trust infrastructure as a shared, low-lock-in
utility rather than as proprietary rails owned by one company. It is not a
perfect model and it has real gaps, but it demonstrates something important:
that population-scale coordination doesn’t require a single controlling entity
if the protocol layer is deliberately kept open and interoperable. The lesson
for AI is not “copy this system” so much as “copy the design principle”, build
the equivalent of open rails for identity, data portability, and model access,
so that intelligence can be composed from below rather than only distributed
from above.
What would that take in practice? A few concrete interventions, none of
them exotic:
•
Data portability as an enforceable right, not a feature,
so an individual’s interaction history
can move with them between AI providers the way a phone number now moves
between carriers, preventing lock-in from doing quietly what outright control
could not do openly.
•
Public or multilateral investment in compute and energy
capacity outside the two or three countries that currently dominate it, on the
model of how public investment built highways and rural electrification rather
than waiting for private markets to reach unprofitable places on their own
schedule.
•
Interoperability standards for model-to-model and
agent-to-agent communication, so a household-level or community-level system
can coordinate with a national one without needing to be owned by the same
company that owns the national one, the
AI equivalent of the postal system agreeing on envelope sizes.
•
Regulatory pressure specifically aimed at the
infrastructure layer, chips, cloud
capacity, energy contracts , rather than only at the visible chatbot layer,
since that is where real concentration risk is currently accumulating fastest
and most invisibly.
•
A cultural shift among the capable middle tier of
nations, those with talent, institutions, and ambition but not
frontier-lab-scale capital, toward building shared, federated capability with
each other rather than each negotiating bilaterally and separately with the
handful of dominant labs, which only reproduces a hub-and-spoke Borg structure
one client relationship at a time.
Sceptics of the diffusion path will point out, correctly, that some
problems genuinely need a Queen, or at least a very large brain. Pandemic modelling,
climate prediction, and fundamental scientific discovery benefit from the kind
of massive, centralized compute that only a handful of institutions can field,
and no household-level router is going to fold a protein or model a hurricane.
The diffusion argument is not that centralized capability should not exist. It
is that centralized capability should be treated the way we treat other
infrastructure with natural concentration risk, nuclear power, undersea cables,
the electrical grid, as a public utility
subject to oversight, access rules, and accountability, rather than as the
private property of whichever company got there first. The European Union’s AI
Act, whatever its flaws in execution, is at least an attempt to draw that line:
to say that as models approach systemic scale, the obligations on their
operators should scale with them. Export controls on advanced chips are a
cruder version of the same instinct, aimed at slowing the concentration of the
compute layer rather than the software layer, even if their current form is
more about geopolitical rivalry than about distributing power more broadly.
And the early scaffolding for genuine diffusion is already visible.
Federated learning in healthcare, hospitals training shared diagnostic models
by exchanging model updates rather than patient records, demonstrates that collective intelligence does
not require pooling raw data in one place. On-device inference, now standard on
flagship phones, means a growing share of everyday AI use never has to leave
the device at all. Neither fully solves the concentration problems described
above, but both show that the direction is technically viable; what is missing
is the institutional will to deploy such architectures at population scale
rather than leaving them as premium features for those who can already afford
to ask.
Humanity’s edge was never raw intelligence. It was the invention of trust
technologies that let intelligence combine across strangers without requiring a
single mind to hold it all. Writing, law, currency, and bureaucracy did this by
distributing judgment while standardizing the interface between people. Whether
AI becomes a fifth trust technology in that lineage, or the tool that finally
lets a single mind hold it all, is not a question that resolves itself as
models get better. It resolves according to who builds the rails underneath
them, and how deliberately the rest of us insist that those rails stay open.
That is a choice still being made, right now, mostly in rooms far from public
view — which is exactly why it needs to be argued for in public.
"We taught the whole species to read.
We did not hand every book to one reader."
No comments:
Post a Comment