The Learning Economy

When AI makes machine reading abundant, privacy—not reach—becomes the scarce resource

For most of the internet’s history, obscurity was the default. Anyone could publish, but almost nobody would notice unless an editor, broadcaster, search engine, large account, or recommendation algorithm granted distribution. The attention economy grew around that bottleneck. Human attention was scarce; platforms accumulated audiences and sold access to them; creators learned to perform for ranking systems whose decisions could make them economically and politically real.

Artificial intelligence is beginning to reverse the bottleneck. As personal and institutional AI systems become capable of reading at Twitter scale or greater, the world can acquire millions and eventually billions of delegated readers. They will monitor repositories, newsletters, videos, filings, forums, public records, social graphs, local news, commercial databases and one another. They will not merely display what they find. They will connect identities, update private models, preserve claims, infer relevance, and selectively brief or act for their principals.

The attention economy says: perform so someone will notice.

The learning economy says: assume every legible trace can be noticed, connected, remembered and privately acted upon.

That transition can weaken platforms’ power over discovery when agents retrieve artifacts directly. It does not remove platform power where access, identity, social context and APIs remain controlled. An anonymous researcher should no longer need a famous account to quote-post a discovery before the right laboratory can find it. But the same machinery also destroys the cheap privacy of obscurity. The anonymous person may not become famous. They may instead become an inferential object continuously modeled by employers, insurers, investors, political organizations, competitors, neighbors, journalists, governments and strangers.

AI can make machine reading abundant without making human attention or judgment abundant. What it newly makes scarce is unreadability: the ability to leave a legible trace without having it persistently found, connected and modeled.

The graph is the moat; the algorithm is the regime

In a recent //Invest Like the Best// conversation, investor Jeremy Giffon describes X as the “global newspaper” through which powerful people in markets, politics, journalism and technology acquire a common narrative. He credits the unified algorithmic feed for this power and calls posting “the last great meritocracy”: in his telling, a new account can write a brilliant post, be selected by the algorithm and suddenly reach hundreds of millions of people.

This account confuses the source of Twitter’s persistence with the mechanism that governs it.

Twitter persists because the consequential people are already there. Politicians, founders, investors, reporters, researchers, activists, celebrities, officials and specialist subcultures remain because the others remain. What cannot easily be copied is not the post format or recommendation model. It is the accumulated graph of standing: who listens to whom, who can confer visibility, where institutions expect announcements to appear, where journalists monitor sources, and where powerful actors expect reactions to become publicly legible.

The network is path-dependent relational capital. Users tolerate the platform’s algorithm because collective migration from that graph is difficult. Continued use validates the network, not necessarily the ranking system imposed upon it.

Zvi Mowshowitz’s practical guide, “Twitter Thoughts For You”, makes this distinction visible through behavior. He says Twitter is indispensable to his work because there are no adequate alternatives, while calling the For You experience “horrid,” describing the interface as hostile to chronological reading, and retreating into a manually curated list of people whose posts he intends to see. He remains for the network while actively defending himself against its algorithm.

On this analysis, network effects—not the recommendation model—produce Twitter’s de facto natural network monopoly. The algorithm occupies the monopoly’s boundary. It decides how much the relationships users deliberately formed are allowed to communicate.

That is why the clean formulation is:

The graph is the moat. The algorithm is the regime.

The anon meritocracy is mostly mythology

The claim that X routinely discovers brilliant anons independently of status does not match the ordinary route to distribution. A low-status account usually requires engagement from a high-status account before the algorithm grants meaningful reach:

anon artifact → elite engagement → algorithmic amplification

The large account supplies standing. The algorithm compounds it.

Zvi collects third-party readings of X’s ranking code that allege account-level priors, small initial test audiences, heavy weighting of early engagement and wider distribution after an initial sample performs. He warns that some reports may describe an older system and that implementations can change at any time. They support a hypothesis about path-dependent reach, not a settled audit of the current algorithm. The durable structure is that content does not receive a neutral trial before the world. Existing graph position helps determine whether it receives a meaningful trial at all.

Replacing follower chronology with opaque algorithmic patronage is not a meritocracy. It is a change in the identity of the patron.

Traditional editorial systems at least identify editors. X presents politically consequential allocation as impersonal computation. The system can privilege established accounts, paid status, owner-preferred narratives, engagement-compatible factions and already fashionable topics while describing the result as organic discovery.

This is a recurrent laundering operation in platform politics: inherited standing enters the machine, and machine output returns as merit.

Peak Twitter brain knows accuracy is optional

The Giffon conversation is most revealing where it describes the “billion-dollar PDF.” Private investment funds may take a decade to produce realized returns. During that long interval, their effective product is narrative. Someone crystallizes uncertainty into a confident story; capital follows it “like ten-year-olds playing soccer”; a new consensus governs allocation until another story replaces it.

Giffon explicitly says the narrative does not need to be right. It needs to arrive at the right moment, sound intelligent, feel compelling and give uncertain people a story on which they can temporarily rest. Later in the conversation, he calls X “almost” a source of truth and argues that timeline-selected narratives influence policy and capital allocation.

That contradiction is not incidental. It is peak Twitter brain explaining its own production function.

Within Giffon’s account, “truth” slides from correspondence with an independently settling world toward a narrative’s ability to coordinate powerful actors. The poster’s thesis then becomes reflexive:

post → algorithmic amplification → capital coordination → price movement → apparent validation → more standing

The poster may be causally powerful without being predictively accurate. If a story attracts capital and raises valuations, its author can cite the response as evidence that the story was right. Private-market settlement comes years later; attention, access, invitations, deals and professional status arrive immediately.

This is why one cannot easily “short posters.” There is no liquid instrument inversely indexed to undeserved algorithmic standing. Even when the thesis fails, the promoter may keep the social capital accumulated while it was fashionable, or migrate to the next story before the previous one settles.

The next-best trade is to choose another court: own or short assets whose value will eventually encounter cash flow, power costs, customer willingness to pay, physical utility, debt service, replacement markets or other consequences that posting cannot indefinitely defer. Yet even this is dangerous. A false narrative with enough coordinated capital can remain causal longer than a dissenter can remain solvent.

The algorithm converts network custody into propaganda power

A chronological following feed preserves a simple constitutional relation: users select the graph, and the graph determines what they can see. A For You feed inserts the platform as a sovereign mediator:

user chooses relations → platform chooses exposure → platform manufactures perceived salience

By changing relative visibility, the platform has the capacity to make a fringe position feel universal, a widespread position feel absent, a controversy feel urgent, or an external source feel irrelevant. The sources establish ranking discretion, not an internal record of particular intentional interventions. The platform need not ban speech. Visibility allocation is enough. Formal availability can coexist with practical disappearance.

Zvi’s account documents the everyday form of this control. Users may see relatively few posts from people they deliberately followed. External links perform badly because the system rewards activity that remains inside the platform. Following chronology is hidden behind extra interaction. Reply-provoking posts receive disproportionate rewards. The platform overrides two explicit choices at once: “I chose to hear this person” and “this person chose to direct me to this artifact.”

That discretion lets X extract editorial—and potentially political-propaganda—power from custody of the coordination graph. The network effect keeps consequential users present; the ranking system governs perceived salience inside the network they cannot easily leave.

Giffon celebrates institutions becoming “timeline native”—continuously monitoring the feed, acting in response, watching the response to their action and reacting again. But when one private actor governs visibility inside this loop, timeline-native institutions become governable by that actor:

algorithmic selection → elite perception → institutional action → new platform content → algorithmic selection

This is a feedback system, not a passive newspaper. Ranking affects elite perception; institutions respond; those responses generate new content and engagement signals for the next ranking cycle.

The platform learns; the user is conditioned

Recommendation systems are described as learning systems, but the distribution of learning is asymmetric. The platform learns what provokes each user. Users learn how to produce behavior that the platform will reward.

Zvi identifies the proxy failure directly: Twitter predicts engagement rather than whether a post creates value. Engagement signals become inputs to ranking and personalization. People therefore learn to treat their own interactions as reinforcement signals. Creators learn the machine’s winning style and adapt themselves to it.

The predictable result is not only AI-generated slop. It is humans becoming slop generators:

• claims detached from sources because links reduce reach;
• conflict bait that creates replies;
• formulaic prose and thumbnails;
• permanent novelty and performative certainty;
• identities compressed into repeatable factions;
• writing optimized to arrest scrolling rather than improve judgment.

This is an anti-learning dynamic: the ranking model can improve at predicting engagement while users adapt their behavior and source-rich artifacts lose ground to content optimized for reaction.

A healthy learning system would distribute gains differently: the user learns, the model learns, the institution learns, and the artifact improves. It would preserve provenance, track later outcomes, maintain dissent and give users control over the constitution by which information is selected. Twitter learns its users in order to govern their attention. An Automatic Newspaper should learn with its users in order to increase their agency.

Delegated reading breaks the attention interface

The most important part of Zvi’s article is not his feed advice but his discussion of machine-scale reading. He considers AI systems that can search Twitter, monitor a ticker’s narratives, find current technical practices, process a whole feed and filter the small portion worth human attention. Using the API prices then available, he estimates that processing a full feed could cost tens of dollars per day and argues that this may be eminently payable if the filtering is good enough. Human time is expensive.

This is the beginning of the learning economy.

In the attention economy, humans consume the stream and manually decide what matters. In the learning economy, agents consume the corpus, update private models and escalate only the events that cross a principal’s relevance threshold.

Twitter then becomes one sensor among many rather than the interface through which the world must be experienced. A personal system can read:

• websites and newsletters;
• repositories and papers;
• filings and public databases;
• videos, podcasts and transcripts;
• forums and federated networks;
• local records and institutional publications.

Creators no longer need to squeeze durable work into the platform’s engagement format. They can publish addressable, source-linked artifacts where delegated readers can retrieve them.

This is distribution by relevance rather than popularity. Human attention, judgment and action remain scarce. The change is that relevance filtering can happen before an artifact competes for mass attention, routing it to the few people for whom it matters.

The technical discovery does not need millions of human impressions. It needs to reach the laboratories that can test it, the maintainers whose code it repairs, the investors able to understand it, and the regulator whose rule it changes. Machine readers can route the artifact into those jurisdictions without first turning its author into a public performer.

Why X taxes machine reading

At the prices Zvi reported, automated posting and messaging cost less per action than large-scale reading, while reading posts containing links incurred additional fees that made full-feed processing expensive. From that price structure, Zvi infers that Musk is less concerned with automated posting than with third parties parsing the corpus. The prices are reported facts from his article; the claim about motive is his interpretation.

That makes strategic sense. Bots that post create inventory and engagement. Agents that read, filter and summarize can detach the graph’s informational value from X’s attention surface:

X corpus → private agent → user briefing

instead of:

X corpus → X ranking regime → user scrolls inside X

Delegated reading threatens screen time, advertising inventory, behavioral conditioning and control over salience. It also hides from X which findings ultimately captured the principal’s attention. Charging for machine-readable access protects the bottleneck.

The future dispute is therefore not merely over who may publish. It is over who may learn from the public graph.

When no one needs to go viral

Today, distribution is a visible event: publish, engage, trend, decay. In the learning economy, distribution becomes continuous and mostly invisible. In a mature learning economy, an obscure post might receive few human views while being ingested by many machine readers—if the artifact remains publicly addressable and platforms do not block automated access. Like counts and follower counts then become poor measures of actual reach.

The anon’s route changes from:

fire content → large account notices → large account engages → algorithm distributes

to:

fire content → many agents independently evaluate → relevant systems preserve and escalate → specific principals receive it

This can break the poster class’s monopoly over public standing. Reputation can become contextual rather than globally theatrical: this obscure source finds compiler bugs; that local observer understands one municipality; this writer repeatedly identifies defects others miss. Their artifacts acquire standing where the evidence settles, without requiring the authors to become celebrities.

But this emancipatory outcome is not automatic. Agents can use existing popularity as a cheap relevance prior and reproduce the old hierarchy at machine scale. If every delegated reader starts from the same engagement-ranked corpus, AI freezes the poster class into automated memory.

The transition requires independent crawling, source diversification, claim tracking, provenance, outcome-based reputation, adversarial discovery and user-owned relevance functions. Otherwise the learning economy becomes the attention economy with better surveillance.

Privacy becomes the scarce factor

The attention economy assumes obscurity and sells visibility. The learning economy industrializes observation.

A person need not actively post a coherent public identity. AI systems may synthesize a profile from accessible traces: public code and filings, published photographs, revision histories, citations, social connections and stylistic patterns, supplemented by private purchase, correspondence or location data where an observer already has access. What once required an investigator’s time can become a background process.

Classic surveillance imagines one state or corporation observing many subjects. The learning economy produces an ecology of reciprocal intelligence agencies. A person may be modeled simultaneously by employers, insurers, landlords, creditors, customers, competitors, political groups, dating prospects, journalists, law firms and merely curious strangers. No observer needs the whole picture. Each needs only the slice relevant to its own objective.

Anonymity is especially vulnerable. Most anonymity today is procedural obscurity: the clues exist, but nobody has enough time or incentive to connect them. Agents can compare writing style, posting times, locations, code commits, image metadata, social connections and repeated phrases indefinitely. Anonymity then requires not that nobody bothered, but that the traces fail to support a sufficiently confident link.

The scarce asset becomes the capacity to preserve uncertainty about oneself.

Privacy in this world is not simply secrecy. It is control over query bandwidth into one’s life: which entities may observe, correlate, retain, infer, simulate and act upon which traces, under which jurisdiction, for how long, with what right of appeal.

The constitutional object is learnability

Many privacy regimes focus on the collection, processing, consent and disclosure of personal data. Cheap inference from public or lawfully held traces forces harder questions:

• Who may infer sensitive facts from otherwise public traces?
• Is persistent autonomous monitoring different from occasional human observation?
• Can a person prohibit linkage among identities?
• Who owns an inferred profile that was never directly disclosed?
• Must an agent reveal what model it maintains about a person?
• Can the subject inspect, contest or appeal consequential inferences?
• Must systems forget?
• Does an investigator need standing before initiating continuous observation?
• How is liability distributed among millions of small surveillants?

The regulated object can no longer be only data. It must include learnability: the permissible transformations from traces to claims, identities, predictions and actions.

This does not imply that public facts should become unusable or that every inference can be prohibited. It means “the data was public” is no longer a complete constitutional answer when automation materially changes the cost and scale of integration.

The Automatic Newspaper

The institutional answer is not another universal feed with a supposedly better central algorithm. It is an owned learning system—an Automatic Newspaper—that continuously ingests the world on behalf of its principal while preserving a court of record.

It should know not only what is popular, but:

• what changed;
• who claims what;
• which evidence supports each claim;
• which sources have earned standing in which domains;
• what the user cares about;
• which anomalies deserve escalation;
• what later events confirmed or refuted;
• which minority reports remain unresolved.

This is not a chatbot summarizing a feed. It is operational capital: a private editorial institution, research desk, archivist, investigator and public-memory interface. Its public artifacts can travel across systems; its private constitution determines what it learns, what it forgets, when it interrupts and what actions require authorization.

If systems of this kind become widely deployed, they will create both the promise and danger of the learning economy. They can break the monopoly through which X converts network custody into propaganda power. They can discover valuable anons without requiring elite retweets. They can replace synchronized attention with plural, user-owned relevance.

They can also make every person permanently legible to every sufficiently motivated entity.

The product and political task is therefore not to prevent learning. It is to reverse its asymmetry. The system may learn from people only insofar as people, their artifacts, their institutions and their communities also become more capable—and only under boundaries that preserve privacy, provenance, exit and appeal.

The new scarcity

The attention economy made people perform to become visible. Its central fear was being ignored.

The learning economy will make people organize, pay and legislate for the right not to be learned. Its central fear will be becoming legible everywhere at once, to observers one cannot identify and courts one cannot reach.

Twitter’s graph will not disappear immediately. Networks are durable. But its ranking algorithm becomes less necessary once delegated readers can traverse the wider world directly. The platform’s real strategic defense will be to restrict machine access to the graph and keep users inside its attention interface.

That defense may preserve the bottleneck for a time. But if inference becomes cheaper and user-owned systems gain lawful access to more sources, some distribution escapes the feed.

The final inversion is simple:

The attention economy made observation scarce and sold visibility. The learning economy makes observation abundant and turns privacy into capital.

AI may not merely help platforms decide what users see. It may also give individuals and institutions systems capable of searching, connecting and remembering on their behalf. The struggle that follows will be over who remains entitled to opacity, who owns the resulting models of the world, and which boundaries can prevent universal learning from becoming universal jurisdiction.