Chris Huber Reitz

Chief of AI & Strategy | Building Organizational Intelligence Design | AI Lecturer @ Columbia University | Trained 15,000+ on AI

[Chris Huber Reitz @chrisreitz]

“Follow the path of the unsafe, independent thinker. Expose your ideas to the dangers of controversy. Speak your mind and fear less the label of ’crack-pot’ than the stigma of conformity. And on issues that seem important to you, stand up and be counted at any cost”
THOMAS WATSON

[Chris Huber Reitz @chrisreitz]

Jack Dorsey's new Buzz chat app makes AI agents members of your team instead of integrations.

Same agent. Same model. Same skills.

In Slack agents are an add-on. In Buzz each agent has a name you can @-mention and a seat in the channel.

Nothing about the agent changed. It’s just showing up more like a human.

We have spent two years arguing about how much to trust an agent. Every trust framework I know assumes the agent arrives with a record and you grade it.

This one hands out the badge at the door and lets the record accumulate. Membership is granted by installation, not earned by demonstration.

That might be the right call. It’s how people assimilate into teams.

But notice what changed hands. The default is now yes, and the burden moved to whoever wants to say no.

https://t.co/UBWrow2E0u

[Chris Huber Reitz @chrisreitz]

Your AI strategy is a literature review.

Open it and read the evidence. The market sizing came from an analyst. The capability claims came from vendors. The urgency came from an article, or a peer at dinner. Every citation is solid. Almost none of it is an observation of your own company.

This week's Signals issue starts in a board meeting. An executive is asked whether his company is behind on AI. He has no number of his own, so he does the responsible thing and brings the best study he can find.

The study is a survey. It asked twelve hundred people in his exact position the exact question his board just asked him. It collected their impressions, because impressions were what there was to collect.

Nothing in the room is false. Nothing in the room is firsthand, either.

The full issue traces what that looks like at scale: why Clippy had perfect numbers, why collecting more of this evidence makes the picture worse instead of better, and the one move that flips the scarcity into leverage. One before-and-after reading from your own systems, dated, with a name attached, outweighs a study of twelve hundred companies. Your sample is real. Theirs is remembered.

All the evidence is secondhand. Yours can be the only firsthand account.

https://t.co/zxrPIkl4ms via @LinkedIn

[Chris Huber Reitz @chrisreitz]

A test suite is an anchor. A type checker is an anchor. A hash of a frozen file is an anchor. You cannot persuade any of them.

An LLM verifier is not one. Change the prompt and it changes its mind.

You can convert some arguing nodes into anchors. Commit to a number before the outcome and it can no longer be talked out of it.

Cheap to check, expensive to fake. That is the signature.

[Chris Huber Reitz @chrisreitz]

A widely shared piece on agent workflows says the verification node is the whole trick, then a few sections later concedes that topology alone does not buy truth.

Both are right. The gap between them is the interesting part.

Adding a checker does not create trust. It moves it. You now trust the checker. Add another and it moves again.

An anchor is not a position in the graph. It is a node that earned the right to end the argument, because proof accumulated on it.

If every node in your graph is an LLM, your graph has no anchors. It is a room of confident strangers agreeing.

The field has rich language for the shape of agent work and almost none for how a node earns standing.

Calibrated authority is the missing half. Grant a node authority in proportion to the evidence that has accrued to it. Where no evidence can accrue, the honest grant is zero.

So the useful question about your workflow is not which node verifies.

It is which node cannot argue back, and what it did to earn that.

[Anatoli Kopadze @AnatoliKopadze]

https://t.co/fyBGWQbsvn

[Chris Huber Reitz @chrisreitz]

This week I joined the board of trustees of Transylvania University, founded in 1780.

We ask young people for the most conviction exactly when they have the least visibility. There is no oracle for your future, only doors you cannot see through.

A university, done well, is an all-weather bet. It helps you commit to something real and remain able to re-choose as the terrain changes.

I’ve spent years writing about how institutions should meet AI. Governance is where that stops being an opinion and becomes a responsibility.

That is why I said yes to serving @Transy.

[Chris Huber Reitz @chrisreitz]

Check out the latest article in my newsletter: We expect the most conviction exactly where they have the least visibility. https://t.co/8yFggu8cyB via @LinkedIn

[Chris Huber Reitz @chrisreitz]

The authority layer is measurable, by the way.

I track how knowledge institutions draw this exact line in their public AI policies. Of the 55 policies I’ve analyzed, 46 draw the same line: machines handle what you can cheaply verify, humans keep what you cannot.

The Calibrated Authority Index: https://t.co/5t1fgpoFhN

[Chris Huber Reitz @chrisreitz]

Cerebras just published the architecture of its internal company brain.

Federated ingest. Hybrid retrieval. Answers that expire as they age. A who_knows tool that tells agents which human actually holds the expertise.

15,000 questions a day, asked by humans and agents alike.

Read it closely and one layer is missing.

Who decides a memory is still true? Which answers carry authority? How does an agent earn the right to act on what it retrieves? Permissions get one line. Trust gets zero.

That is not a knock on the engineering. It is the state of the frontier: a world-class lab ships the pipes in a quarter, and the layer that decides what the organization actually believes still has no owner.

Retrieval is how an answer arrives.
But Intelligence demands someone be accountable for it being right.

The substrate is becoming a commodity. The authority layer is the asset.

[Cerebras @cerebras]

https://t.co/FvKy1StDpQ

[Chris Huber Reitz @chrisreitz]

The companion piece, a ten-minute audit for sorting which tasks to hand an agent and which calls to keep: https://t.co/1Jbs2gLS0M

[Chris Huber Reitz @chrisreitz]

I build AI agents for a living. My first rule is restraint.

Not "what can it do?" but "what should it never do?"

Three rules I build into everything:

[Chris Huber Reitz @chrisreitz]

Everyone's racing to make AI do more.

The harder engineering is deciding what it does when it doesn't know, and making that boring and predictable.

Restraint isn't the absence of ambition. It's ambition that learned to say "I don't have that."

[Chris Huber Reitz @chrisreitz]

The full index, the methodology, and every institution's own words:

https://t.co/5nq1GC57Ij

[Chris Huber Reitz @chrisreitz]

In 1979, an @IBM training manual drew the line in one sentence: a computer can never be held accountable, therefore a computer must never make a management decision.

In 2025, IBM's own Think blog revisited the question and gave up. "There's no hard-and-fast line," it concluded. "It's a moving target."

Last week I published the AI policies of 51 of the world's most trusted institutions. Nature. The BBC. Oxford. The Associated Press.

44 of them drew the same line. Independently. Not one of them named it.

Machines get the work you can cheaply check. Humans keep the work you can't.

The line never moved. It was never about ethics. It's the cost of proof: where proof is cheap, the machine is welcome. Where proof is scarce, the human stays.

IBM had it right the first time. Accountability is the one thing a machine cannot hold.

47 years later, the data agrees with the training manual.

[Chris Huber Reitz @chrisreitz]

A lens for reading any AI vendor who walks in the door:
Are they making proof cheap, or asking you for faith?

Hold everyone to that bar.

[Chris Huber Reitz @chrisreitz]

We trust a machine exactly as far as we can cheaply check it. Past that, it's faith.

[Chris Huber Reitz @chrisreitz]

I'm publishing this as a living index that re-scores itself as policies change. It already caught ACM drop its AI writing-disclosure requirement last month.

The pattern every institution discovered and none named — I'm naming it: Calibrated Authority.

https://t.co/CqljShYdch

[Chris Huber Reitz @chrisreitz]

I coded the public AI policies of 51 of the world's most-trusted knowledge institutions — Nature, JAMA, the BBC, the NYT, Oxford, UNESCO.

I expected 51 different answers. I found one pattern. 🧵

[Chris Huber Reitz @chrisreitz]

Universities show the same reflex, quietly: almost none write a single AI rule — they hand the call to the individual instructor, the one human close enough to judge what a machine can't certify.