In May 2001, Scientific American ran a feature by computer scientists Sir Tim Berners-Lee, James Hendler, and Ora Lassila. It opens with a story from a hypothetical future. A woman sitting in a doctor’s office uses her handheld browser to instruct a software agent to pull her mother’s treatment plan from the doctor’s own agent, crosscheck in-network providers against the insurance plan, reconcile two of her siblings’ calendars, and schedule an appointment. No one has to fill out any forms. Instead, the machines speak to each other by exchanging a type of data that carries its meaning along with it.
That piece was a manifesto for what Berners-Lee and his colleagues dubbed the Semantic Web: an extension of the World Wide Web “in which information is given well-defined meaning, better enabling computers and people to work in cooperation.” Over the next decade, the World Wide Web Consortium (W3C) produced technical standards to build the Semantic Web—most notably Extensible Markup Language (XML), the Resource Description Framework (RDF), and universal resource identifiers (URIs)—believing that they would “unleash a revolution of new possibilities.” By and large, they did not. The ontologies were too cumbersome, the markup was too tedious, and the incentives to annotate the internet never quite materialized. That edition of Scientific American and the dream of the Semantic Web were put on the shelf and left to collect dust.
Earlier this year, Berners-Lee’s manifesto was taken off the shelf and dusted off, when a startup called AutoAmerican filed comments at the FCC and the FTC on Mar 26th. They argued that phone company records tracking where users go and how they live should be made available, at the user’s direction, to intermediaries legally bound to serve them. In the AI era, this is the most valuable behavioral data. Less than a month later, Google published a specification called the Open Knowledge Format, a way of collecting an organization’s information into a portable and interoperable format directly readable by AI agents. One filing is about the knowledge a company holds, and the other is about the data a person generates. They are complementary answers to the same question Berners-Lee posed in 2001, now made possible by machines’ ability to overcome the challenges of ontology and annotation.
The Failure of the Semantic Web
In order to understand why the Semantic Web may be revitalized, one must first understand why it failed the first time. Berners-Lee, who was instrumental in the development of the Web, identified one of its major points of failure. As he wrote, the Web was “designed for humans to read, not for computer programs to manipulate meaningfully,” and thus “computers have no reliable way to process the semantics.” To correct this, the authors proposed adding three new pieces or layers to the existing Web stack.
XML is a writing system that lets anyone invent tags to mark up the parts of a page. Housed within XML is RDF, which supplies the meaning by stating facts as subject-verb-object triples. RDF then uses a URI to give unique web addresses to each term in those triples. If we take the hypothetical doctor’s office described above, whoever built the clinic would encode information so that a machine could understand that <clinic> <employs> <Dr. Smith> and <Dr. Smith> <accepts> <patient’s insurance>. On top of this the authors imagined software agents that find and chain together online services on their own, exchanging proofs and digital signatures to check that a claim came from a trusted source.
One immediate problem with this system is that developers could encode whatever meaning they’d like and so the taxonomy would necessarily be different between different databases. To translate meaning across databases, the authors proposed the creation of ontologies, “a document or file that formally defines the relations among terms.” The ontology would be able to translate meaning so that agents can understand, for example, that Texas is in the United States, that a doctor is a type of medical provider, and that a zip code and postal code are the same thing.
This solution is clever but laborious. Every piece of this system requires the person who held a piece of data to do unpaid work for a reader which did not exist yet. In the given example, despite the office manager at the doctor’s office never having taken a computer science course, she is expected to tag the page, write the triples, mint the URIs, and point them all at the right ontology. On top of this, she must keep all of this up to date so that someday an agent might arrive and book an appointment. Network effects magnified the problem; the system needed near-universal markup to be worth using and offered no immediate benefit for marking anything up before universality arrived.
The places where the markup did get done were the places where a large player saw immediate benefit and paid for it. The most notable example of success is that Google realized tagged pages were better for web search and so it began rewarding tagged pages with richer search results. This incentivized developers to tag their pages, but the taxonomy was discreet and the ontology was co-authored and maintained by Google. So W3C created the necessary standards but the Semantic Web mostly remained hypothetical.
AutoAmerican and Mobile Metadata
So what’s different now? First, we start with the data, because Berners-Lee’s mentions of “personal data” is vague. As it turns out, most of the personal data that the Semantic Web needs already exists on the nearly 5 billion smartphones throughout the world.
Despite the lingering myth, your phone isn’t recording everything you say. On the telecom side, what it does record is known as Customer Proprietary Network Information or, more broadly, mobile metadata. This includes records of what numbers you called or texted, when and where you were when you called, what network infrastructure carried it, how your phone is configured, and all sorts of other snippets of information. Strung across months, that record is a map of your life precise enough that the pattern alone describes you. The record of this “personal data” sits inside the carriers’ networks and it moves with you.
Since 1996, what carriers can do with that information and user’s rights to privacy regarding that information have been governed by §222 of the Telecommunications Act. Crucially, §222 specifies that carriers “shall disclose customer proprietary network information, upon affirmative written request by the customer, to any person designated by the customer.” In other words, mobile metadata is, for all intents and purposes, yours. It is just held by the carrier which has a legal responsibility to do with it only what you ask them to.
This is where AutoAmerican comes into the picture. Essentially, AutoAmerican aims to be an intermediary between carriers and AI companies that would easily allow users to delegate access to their mobile metadata to AI agents. The structure of this arrangement was laid out by AutoAmerican’s co-founder and notable tech lawyer Richard Whitt in his book Reweaving the Web. Whitt and his co-founder, Sunil Daluvoy, want the carrier to release your metadata only when you direct it to, and only to a party that has taken on a fiduciary duty to act in your best interest. Their claim is that the company holding the keys to your behavioral record should be legally bound the same way your banker or lawyer is. It should be loyal to you, transparent about every transfer, and switched off the moment you say so. Since the wireless carriers are already legally constrained in how they use your data, AutoAmerican wants AI companies and any intermediary handling that data to be similarly constrained by fiduciary duty.
What makes this privacy startup a crucial step towards a more Semantic Web is where AutoAmerican sits in the stack. Berners-Lee’s design asked the holder of every piece of data to annotate it by hand for a reader that had not yet arrived. Because of this, it failed. But mobile metadata is already structured, with the carrier recording it in machine-readable form as a condition of running the network. The markup the 2001 design begged someone to do has already been done on five billion phones.
The Open Knowledge Format
If mobile metadata is one half of the equation for making an agentic future work, then the knowledge companies hold is the other half. Some corporate data such as corporate addresses and phone numbers is public while most of the important information such as a doctor’s schedule is private and there has never been a good way to structure this information such that an AI agent could easily parse and act upon it. Enter LLM-wiki and Google’s Open Knowledge Format.
OKF is, at its core, just a directory of plain text files where each of these files describes something a company knows. Using the same links as any web page, files representing database tables, metrics, APIs, or the steps a scheduler would use to make an appointment, link to one another and turn the folder into a map of interconnected company knowledge. Agents read and understand the whole bundle because OKF builds on “LLM wiki,” Andrej Karpathy’s existing system, adding a small set of conventions that let one team’s agent read another team’s wiki without translation.
When compared to Berner-Lee’s Semantic Web system, OKF is a huge improvement. The Semantic Web would have required an office manager with no computer science knowledge to do a lot of technical work. OKF only requires her to add a document that is probably already on a hard drive to a folder. RDF’s reader was a program that could infer nothing on its own, which meant the writer had to carry the meaning in formal logic; for instance, through an ontology file stating a zip and postal code are different things.That OKF is then read by a language model that already has enough context to know what it means which resolves the Semantic Web’s biggest challenge.
Putting It All Together
Very soon, the woman in the doctor’s office will have an agent who is able to book the appointment, facilitated by two kinds of information meeting and exchanging meaning. Unlike under Berners-Lee’s design, information holders are not forced to imprint formal logic onto each piece of information. The shape of the semantic web has transformed into one where the company and the user hand over a plain document and an already-structured record, and language models easily do the rest. Because no one has to annotate anything, the meaning Berners-Lee wanted the data to carry lives in the reader, and agents are able to act on it. It is clear now that 25 years of failed markup was the easy problem to solve, once we got advanced machine learning. Finally, the Semantic Web is actually buildable.
