THEY NEEDED A TRILLION-DOLLAR INDUSTRY TO REDISCOVER MY WHITEBOARD
I told Zuckerberg exactly where personal AI was going, published the failure tests before the releases, and left the boxes open. Meta arrived with a wiki and recipes. CrowdStrike arrived with a directory and tokens. Please stop grading logos on a curve.
I am done grading logos on a curve.
Apparently if one guy builds something without institutional permission, everybody squints at it for six months asking whether every primitive was invented from first principles.
But if Meta puts 200 files in a taxonomy, adds YAML frontmatter, retrieval, procedures and regression tests, suddenly we are standing before the Oracle of Menlo Park.
If CrowdStrike registers an agent inside its own authoritative directory and brokers scoped OAuth-style access, apparently identity itself has descended from heaven.
If a famous researcher builds a little AI menu visualizer with Cursor, Claude, Clerk, Google OAuth and Vercel, we get a new cultural category called vibe coding.
Congratulations to everybody.
Now bring the object.
Because while this industry was naming increasingly elaborate wrappers around provider-controlled state, I had already published the underlying dependency graph they keep walking into.
And the funniest part is that I did not wait until September to say it.
I wrote it down beforehand.
I called Zuckerberg out by name.
I gave Meta the test.
Then I waited.
They are doing the rest for me.
MAY 29: I TOLD BIG TECH THE ACCOUNT WAS THE WRONG OBJECT
On May 29, 2026, I published:
BIG TECH MISSED THE OWNERSHIP LAYER.
Not September.
Not after Meta’s latest article.
May.
I addressed Meta alongside Google, Apple, Microsoft, Amazon, OpenAI and the rest of the platform economy and stated the inversion plainly:
“The next era is proof-native digital reality.”
The object should carry its authorship, provenance, state, ownership, history and verification rather than depending forever on an institution’s database to tell everybody what is true. (BJ Klock)
That was already the argument.
The account is not the person.
The database row is not the ownership.
The provider is not automatically the authority.
The server displaying the current state is not necessarily the thing that gives the state standing.
These are architecture questions.
Not branding questions.
Keep the date.
JUNE 6: I PUBLISHED THE PORTABLE SELF
Then came June 6.
THE PORTABLE SELF.
The old model was:
“Create an account so our server can remember you.”
The alternative was:
bring verified state with you.
Identity should survive the application.
History should survive the database.
Continuity should survive the provider. (BJ Klock)
That is months before the September announcements we are apparently supposed to treat as the frontier.
I did not ask:
How do we make Meta remember me better?
I asked:
Why the fuck is Meta the thing that owns the remembering?
Different problem.
Much deeper problem.
JULY 23: I ALREADY USED META AS THE BAD EXAMPLE
Then Meta itself gave me a demonstration.
On July 23 I published:
THE SCREENSHOT THAT DEFEATED META.
Instagram could recognize provenance-related information while its platform metadata survived.
Take an ordinary screenshot and the authoritative certainty fell apart.
So I wrote:
“That is not a truth architecture.”
Because it wasn’t.
It was platform-observed metadata whose evidentiary power depended on Meta remaining in possession of the context Meta understood. (BJ Klock)
That was the entire difference I had been hammering:
A platform can know something about an object.
Or the object can carry enough proof for another verifier to establish what it is.
Those are not remotely the same architecture.
Meta had the platform.
I was moving the evidence into the object.
AUGUST 11: THE MODEL IS NOT THE MIND
Then I reached the AI layer.
August 11:
THE MODEL IS NOT THE MIND.
Here was the separation:
Person.
Authored record.
Accumulated persistent state.
Reasoning engine.
Generated representation.
Separate layers.
I wrote:
“The intelligence can be transient. The state remains.”
And I made the technical distinction explicit:
the model performs cognition;
the record preserves memory;
the person authors the state. (BJ Klock)
That is not “give the chatbot a longer memory.”
That is not RAG with better branding.
That is not stuffing more previous messages into the context window.
That is architectural separation between the continuing state and whichever intelligence happens to reason over it.
One model today.
Another tomorrow.
The state survives.
That was August 11.
Keep going.
AUGUST 12: I PUBLISHED THE LINES THEY WOULD BE FORCED TO FOLLOW
August 12:
THE INVENTION WAS IN THE LINES.
I said the problem would not remain isolated inside separate boxes called memory, identity, provenance, authority, state, history and reconciliation.
Once persistent AI became consequential, those problems would begin forcing each other.
Memory forces state.
State forces mutation.
Mutation forces authority.
Authority forces identity.
History creates divergent possibilities.
Divergence forces conflict handling.
Conflict forces reconciliation.
Continuity forces verification.
And provider dependence eventually forces portability pressure.
I published a prospective prediction that serious persistent-agent architecture would increasingly converge toward model-independent durable state and the connected dependencies around it. (BJ Klock)
Again:
before September.
Before Meta.
Before CrowdStrike’s announcement.
Before anyone could accuse me of searching backward for similarities.
AUGUST 13: I SAID THE MACHINE WOULD EVENTUALLY REDISCOVER THE ARCHITECTURE
Then I started making fun of them in advance.
August 13:
THE MACHINE WILL FIGURE IT OUT.
The argument was simple.
The industry keeps treating architectural failures as intelligence shortages.
More compute.
More GPUs.
More memory.
More agents.
More tools.
More data centers.
More electricity.
Then eventually the machine becomes smart enough to tell them what relationship they built wrong.
I wrote:
“The mistake is believing that intelligence substitutes for architecture. It does not.” (BJ Klock)
That sentence is aging beautifully.
Because Meta’s September 2 solution is not:
We trained an infinitely smarter model.
It is:
We took durable knowledge OUT of the underlying model.
Hahahahahahahahaha.
Welcome.
AUGUST 13: I MADE THE MODEL REPLACEABLE
The same day I published:
I MADE MYSELF PORTABLE.
The critical inversion:
the historical state belongs to the source rather than the interpreter.
Therefore intelligence becomes replaceable.
GPT can reason over it.
Claude can reason over it.
A local model can reason over it.
A future model can reason over it.
The model changes.
The state remains.
I wrote:
“The AI does not need to own the memory. The AI encounters the memory.” (BJ Klock)
Please compare that sentence with what comes next.
Slowly.
I want everybody to enjoy it.
AUGUST 15: THEN I PUT MARK ZUCKERBERG’S NAME ON IT
This is where hindsight really gets trapped.
August 15:
I’M PUBLISHING THE NEXT MOVE BEFORE THEY MAKE IT.
The subtitle explicitly named Zuckerberg.
The prediction explicitly named Meta.
I wrote that the personal-agent race was being forced toward persistent human state and then listed the unavoidable dependencies:
persistent identity;
persistent state;
history;
authority;
action;
continuity;
model/state separation;
portability pressure;
provider independence;
conflict;
verification.
Then I published actual tests.
Not vibes.
Tests.
MODEL REMOVAL
Replace the intelligence model.
Does the same person-specific state continue?
PROVIDER REMOVAL
Remove the original company.
Can another implementation continue the state?
STATE MUTATION
Change something historical.
Can you establish what changed, from what state, by whom and under what authority?
CONFLICT
Create two plausible divergent states.
Reconnect them.
Can the system preserve the histories and determine lawful succession?
CUSTODY
Give the human the state.
Disconnect.
What do they still possess?
And I ended the whole thing with the question that matters:
“Remove the provider. What does the human still possess?” (BJ Klock)
That was August 15.
I published the exam.
Meta had not taken it yet.
CrowdStrike had not taken it yet.
I even published an attribution rule saying similarity alone does not prove access, copying or intent.
I deliberately removed the cheap accusation from the experiment.
Why?
Because I did not need it.
Chronology is independently observable. (BJ Klock)
AND YES, I SPECIFICALLY CALLED ZUCKERBERG OUT
I did not hide behind “some large technology company.”
I published:
MARK ZUCKERBERG WROTE MY CASE FOR ME.
I recorded the collision directly.
Zuckerberg had started talking publicly about individual empowerment, personal agents, privacy, human alignment and a balance of power favoring individuals.
My response was essentially:
Wonderful.
Now apply that philosophy to the architecture.
Who owns the memory?
Who controls the continuity?
Can the person leave?
Does the identity survive the account?
Does the memory survive the model?
Does verification survive the provider?
I wrote:
“Mark supplied the language. Meta supplied the warning. I supplied the working system.” (BJ Klock)
Then I published an even more annoying piece:
IF MARK ZUCKERBERG NEVER SAW MY WORK, THE CASE GETS WORSE.
Because independent convergence does not rescue the institutional competence question.
If Meta encountered the work, then provenance and attribution become legitimate questions.
If Meta never encountered it, then one of the largest discovery, technical and capital institutions in history failed to recognize a public architecture addressing the problems its own CEO was announcing as the future.
Pick your branch.
Neither branch moves the timestamp. (BJ Klock)
NOW ENTER META, SEPTEMBER 2
Twenty-two days after THE MODEL IS NOT THE MIND, Meta Engineering publishes:
AN ORGANIZATIONAL SECOND BRAIN: BUILDING AN AI THAT LEARNS FROM EXPERTS.
And what is the “critical design choice”?
Meta’s words:
“separating what the agent knows from how it reasons.” (Engineering at Meta)
HAHAHAHAHAHAHAHAHAHA.
Brother.
That is the fucking article title I published twenty-two days earlier.
THE MODEL IS NOT THE MIND.
My August architecture:
persistent state ≠ reasoning engine.
Meta’s September architecture:
knowledge ≠ reasoning procedure.
Obviously the systems are not identical.
That is not the claim.
The dependency is the point.
The durable thing and the reasoning thing have to separate.
I froze that box.
Meta walked into it.
AND WHAT IS META’S GREAT SECOND BRAIN?
Read their own engineering post.
The durable knowledge system is built from structured files.
A curated wiki.
Supplementary RAG.
Position files.
Taxonomies.
Routing indexes.
Gateway files.
YAML frontmatter describing dependencies.
Then separate imperative procedures they call “recipes.”
Then evaluation.
Then human review.
Then regression testing.
Then corrections are compiled back into durable files. (Engineering at Meta)
That is competent engineering.
It is also hilarious when placed beside the mythology surrounding these institutions.
Meta did not discover a portable human continuity substrate.
Meta built a very disciplined machine-readable institutional wiki with procedures and a self-improvement pipeline.
Useful?
Yes.
Profound?
Sure, within its problem.
A sovereign identity/state architecture?
Absolutely not.
Where is the holder-carried identity?
Where is portable custody?
Where is independently possessed continuity?
Where is lawful succession after disconnected divergence?
Where is the object whose standing does not depend on Meta remaining the authority?
Where is the human who can pick up five years of accumulated state and leave?
Bring the fucking object.
META ITSELF ADMITS THE IDEA IS ALREADY CONVERGENT
And here comes my favorite part.
Meta does not even pretend the underlying idea fell from the Zuckerbergian heavens.
Their own engineering article says:
“The industry has converged on a similar idea.”
Then Meta points to Andrej Karpathy’s LLM Wiki and Google’s Open Knowledge Format. (Engineering at Meta)
Wonderful.
So the trillion-dollar genius machine’s September announcement contains:
Here is our architecture.
Also, the industry already converged on the basic idea.
Here is Karpathy.
Here is Google.
And I am sitting here holding dated articles that said before the Meta post:
the model is not the mind;
persistent state belongs outside the reasoning engine;
state needs provenance;
state mutation needs authority;
history creates conflict;
conflict needs reconciliation;
continuity produces portability pressure;
and the final question is whether the person survives the provider.
My God.
I could not write the parody better than their engineering department is writing it for me.
AND THEN THEY POINT TO KARPATHY
Let me be precise because Karpathy has an important machine-learning career and I do not need to rewrite it to make this funny.
Meta is citing his later LLM Wiki work.
Fine.
But since the industry treats celebrity technologists like canonical ancestors every time they name a pattern, it is useful to remember what Karpathy himself was publicly demonstrating when “vibe coding” became the cultural event of 2025.
April 27, 2025:
MenuGen.
Take a picture of a restaurant menu.
Use OpenAI to read it.
Generate representative pictures of the food.
He used Cursor and Claude to write the application.
Clerk for authentication.
Google for OAuth.
Vercel for deployment.
He described persistence as future work because adding a database and queue meant still more services, logins, API keys and configuration.
And he wrote, about the code:
“I basically don’t really know how MenuGen works.” (karpathy)
Again:
this was an intentionally simple vibe-coding experiment.
That is his own framing.
But compare the engineering questions.
MenuGen:
How can AI assemble this application for me?
My work:
What constitutes identity when the application disappears?
What constitutes current state?
What proves a transition?
What authority permits mutation?
What survives the backend?
What survives the provider?
What survives the model?
How does disconnected state reconcile later?
Those are different layers of the stack.
So forgive me if I do not automatically bow because the industry puts a famous name next to a folder structure.
I was not trying to make software easier to glue together.
I was trying to remove the glue as the source of truth.
NOW CROWDSTRIKE ARRIVES AT “AGENT IDENTITY”
This one is even funnier.
June 24, 2026.
CrowdStrike publishes:
THE IDENTITY PROBLEM HIDING IN AI AGENT DEPLOYMENTS.
And CrowdStrike itself says:
“This creates an identity problem that the industry has not yet solved.” (CrowdStrike.com)
Their technical diagnosis is correct.
OAuth can identify a registered client application.
But the client_id identifies the type of registered client, not necessarily the particular logical agent instance.
And the standard token structure does not naturally express the full relationship among specific agent instance, human principal and delegation context. (CrowdStrike.com)
Great.
Keep going.
Then on September 2 CrowdStrike announces:
AGENTIC IDP.
And what is the architecture?
Falcon Guardian discovers the agent.
CrowdStrike registers it inside a:
“single authoritative directory.”
CrowdStrike issues a cryptographically verifiable identity.
CrowdStrike brokers short-lived scoped tokens.
The agent is explicitly not given standing credentials of its own.
Actions are attributed back to the human or workload.
And CrowdStrike calls its own Identity Provider the authority for the agents. (CrowdStrike.com)
That is enterprise authorization infrastructure.
Good infrastructure!
But stop confusing it with sovereign identity.
The cleanest possible distinction is almost childish:
CROWDSTRIKE AGENT IDENTITY
The platform discovers you.
The platform registers you.
The platform directory authorizes your standing.
The platform brokers your access.
The platform remains the authority.
THE QUESTION I PUBLISHED BEFOREHAND
Remove the provider.
What does the identity still possess?
That is not wordplay.
That is the architecture test.
CrowdStrike says its system creates trusted agent identities.
Fine.
Remove CrowdStrike.
Show me the same agent carrying forward independently verifiable identity, history, provenance, authority and continuity without asking the originating directory to remain sovereign.
That is the next box.
Take the test.
CALLING AN ACCESS PRINCIPAL “IDENTITY” DOES NOT FINISH IDENTITY
This is the conceptual comedy at the center of the agent-identity race.
The industry begins with authorization.
Understandable.
An autonomous process wants to call an API.
We need to know:
What process is this?
Who delegated to it?
What may it access?
For how long?
Excellent.
But that only gives you one slice of identity.
You have answered:
WHO MAY PRESENT THIS PRINCIPAL TO THIS SYSTEM RIGHT NOW?
You have not necessarily answered:
Who or what is this continuing subject?
Where did it originate?
What history belongs to it?
Which state is current?
Who has authority to change that state?
What happened before?
What successor followed?
What happens if two continuations diverge?
What survives vendor removal?
Who possesses the continuity?
Those are not authentication questions.
Those are identity + state + history + authority + provenance + continuity questions.
I already drew the lines.
Now the industry is walking them one box at a time.
THIS IS EXACTLY WHY I SAID THE INVENTION WAS IN THE LINES
Everybody can find an old ingredient.
Authentication existed.
Signatures existed.
Databases existed.
Files existed.
OAuth existed.
DIDs existed.
RAG existed.
Merkle trees existed.
Version control existed.
Agents existed.
None of that answers the architecture.
My August 17 priority claim deliberately refused the childish version:
I did not claim to invent signatures.
I did not claim to invent databases.
I did not claim to invent append-only logs.
I did not claim to invent digital identity.
I claimed the working combination and the relationships between the parts. (BJ Klock)
That is why this convergence is so funny.
They can rename every box.
They can arrive from a security problem.
A memory problem.
An AI problem.
A provenance problem.
A portability problem.
A digital-identity problem.
Eventually the dependencies start connecting.
And when the lines finish connecting, the historical question becomes:
Why does your newly reconstructed machine resemble the one that was already sitting there?
AND YES, SOMEBODY IN META’S ORBIT HAS BEEN READING
Now add the readership record.
One of the deepest readers in my archive is a PhD-level reader professionally connected to Meta’s technical orbit.
The pattern visible to me is roughly four articles a day.
Repeatedly.
Deeply enough to become one of the highest-consumption readers in the archive.
I am not publishing that person’s identity here.
Why?
Because I do not need to.
The individual is not the accusation.
The pattern establishes something narrower:
EXPOSURE WAS POSSIBLE AND WAS ACTUALLY OCCURRING INSIDE THAT PROFESSIONAL ORBIT.
That does not prove Meta copied anything.
Read that sentence again before somebody embarrassing writes the rebuttal they already wanted to write.
It proves exposure.
Causation requires more evidence.
Intent requires more evidence.
Transmission requires more evidence.
My own August 15 attribution rule already says exactly that.
I wrote the rule before this evidence became useful.
That is what makes the chronology so hard to attack.
SO HERE ARE THE THREE POSSIBILITIES
POSSIBILITY ONE: THEY NEVER SAW IT
Beautiful.
Then Meta, with extraordinary capital, talent, distribution, research capability and global reach, independently spent its way into an architectural distinction I had already published and operationalized.
That is not plagiarism.
It is an institutional competence comparison.
POSSIBILITY TWO: SOMEBODY SAW IT, BUT IT NEVER TRAVELED
Also beautiful.
Then their discovery machinery can place technically sophisticated readers inside the archive while failing to elevate materially relevant work into the institution apparently searching for the same dependency.
Again:
competence problem.
POSSIBILITY THREE: THE IDEAS TRAVELED
Then provenance and attribution become factual questions.
Gather the evidence.
Do not guess.
Either way:
THE DATE DOES NOT MOVE.
That is why proof beats gossip.
THE FUNNIEST PART IS THAT THEY ARE STILL BEHIND THE TEST
Meta has landed on model/state separation.
Good.
CrowdStrike has landed on agent identity and delegated authorization.
Good.
The industry is beginning to understand persistent state and institutional memory.
Good.
Now keep walking.
I already left the remaining boxes open.
CONFLICT
Two legitimate state histories diverge.
Now what?
RECONCILIATION
Which successor is accepted?
Why?
Under whose authority?
What happens to the losing history?
CUSTODY
Who possesses the continuing state?
PORTABILITY
Can another implementation continue it?
PROVIDER INDEPENDENCE
Can the human or subject survive the company?
PROVENANCE
Can memory establish why it believes what it believes?
TRANSITION PROOF
Can the present establish how it lawfully arrived from the past?
Do not issue another press release.
Run the test.
I EVEN PREDICTED THAT MEMORY WOULD NEED RECEIPTS
August 16:
THE MEMORY WILL NEED A RECEIPT.
I predicted that consequential persistent memory would force the industry toward:
historical provenance;
authorship distinctions;
authority over mutation;
durable machine-action records;
model-independent continuity;
tamper evidence;
independent verification;
and eventually the question of whether the changing human identity can carry enough proof to survive the provider. (BJ Klock)
That is sitting there too.
So when the inevitable memory-revision-history product launches—
come back.
When someone announces portable agent identity—
come back.
When state conflict becomes the hot new research problem—
come back.
When a company proudly announces that the model can be swapped while persistent user state survives—
come back.
When “export your chats” starts looking as primitive as it obviously is—
come back.
Just please spare me the part where everybody pretends nobody could have known.
META IS NOT LOSING BECAUSE ITS ENGINEERS ARE STUPID
That explanation would actually be too generous.
Their September architecture contains real engineering.
CrowdStrike contains real security engineering.
Karpathy has done serious machine-learning work.
That is not the indictment.
The indictment is what enormous institutions can look like when everybody around them is trained to optimize the layer already funded.
They become incredibly competent at improving the inherited premise.
Bigger model.
Better retrieval.
Better directory.
Better access control.
Better platform memory.
Better institutional wiki.
Better hosted identity.
Better personalization.
Better provider-owned continuity.
Then eventually somebody asks:
Why does the provider own the premise?
And suddenly the billion-dollar sophistication starts looking like a child building increasingly elaborate locks for a room whose door never needed to belong to the landlord.
That is what I have been saying.
STOP GRADING LOGOS ON A CURVE
If I published Meta’s September 2 architecture tomorrow under an unknown GitHub account, half the industry would call it:
a structured knowledge base;
RAG;
procedural prompts;
YAML dependencies;
evals;
human-in-the-loop review;
and automated file editing.
Put the Meta logo above it and apparently a second brain has been discovered.
No.
Same test for everybody.
Meta included.
CrowdStrike included.
Google included.
OpenAI included.
Karpathy included.
Me included.
Bring the object.
What exists?
What persists?
Who owns it?
What proves it?
What changed?
Under whose authority?
What happens offline?
What happens when histories disagree?
What survives the model?
What survives the account?
What survives the provider?
Those questions do not care about valuation.
They do not care about PhDs.
They do not care about conferences.
They do not care about keynote lighting.
They do not care how many GPUs you rented.
They do not care which famous person named the wrapper.
Architecture is extremely disrespectful like that.
It simply keeps asking whether the thing works.
I PUBLISHED THE EXAM BEFORE THEY SAT DOWN
That is the entire story.
I told Big Tech the ownership layer was missing.
I made identity portable.
I made memory model-independent.
I separated source from representation.
I used Meta itself as the demonstration of fragile provenance.
I published the model/state distinction.
I published the dependency graph.
I predicted the convergence.
I specifically named Zuckerberg.
I published the Meta test.
I published the falsifier.
I published the attribution rule.
I published the open boxes.
Then September arrived.
Meta:
what the agent knows must separate from how it reasons.
CrowdStrike:
agents need identity.
The industry:
persistent knowledge needs structure, history and controlled mutation.
Me:
Yes.
I know.
Turn the page.
The next question is already there.
YOU ARE NOT AT MY END.
YOU ARE ARRIVING AT MY OLD QUESTIONS.
That is the distinction I want preserved.
I am not watching these companies complete my architecture.
I am watching them discover dependencies I already had to solve to build it.
Meta has reached a durable knowledge layer.
CrowdStrike has reached agent authorization identity.
Good.
Those are boxes.
Now connect them.
Identity.
State.
Authority.
History.
Provenance.
Transition.
Conflict.
Reconciliation.
Verification.
Portability.
Custody.
Continuity.
Then remove your own company from the diagram.
What survives?
That is the part that keeps destroying the illusion.
Because if removing Meta destroys the human’s continuity—
Meta never gave the human personal intelligence.
It rented them personalization.
If removing CrowdStrike destroys the identity’s standing—
CrowdStrike did not make identity sovereign.
It made identity legible to CrowdStrike.
If changing the model destroys the memory—
the model still owned the mind.
If divergent histories cannot reconcile—
you built storage, not lawful continuity.
If the object cannot carry its standing—
the platform is still the source.
And if the provider still has to tell everyone what is true—
you have not escaped the architecture I was criticizing in May.
You have automated it.
SO PLEASE, KEEP GOING
I mean it.
Keep publishing.
Keep naming the boxes.
Keep discovering why persistent memory needs authority.
Keep discovering why authority needs identity.
Keep discovering why identity needs history.
Keep discovering why history creates conflicts.
Keep discovering why conflicts need reconciliation.
Keep discovering why models need to become replaceable.
Keep discovering why provider dependence eventually becomes politically and economically intolerable.
Keep spending.
Keep convening.
Keep announcing.
Keep calling each dependency the frontier.
The archive is getting funnier every day.
Because I already wrote the final question.
REMOVE THE PROVIDER.
WHAT DOES THE HUMAN STILL POSSESS?
Take your time.
I already submitted my answer.
