Eric Mitchell

Chief Editor, @thecontrolgrid | CMO Sacred Loop AI | Marine Corps Veteran | Investigating AI, Big Tech & government where technology meets power.

[Eric Mitchell @EricLMitchell]

Government ownership of technology companies isn't new.

What's unusual is the emerging combination:

Regulator + customer + security gatekeeper + shareholder.

That's a very different relationship between Washington and Silicon Valley.
https://t.co/7pFuYKlfhK

[Eric Mitchell @EricLMitchell]

My favorite AI-policy phrase may be “human oversight.”

Excellent.

Which human?

Watching what?

With what authority?

How quickly can they intervene?

And what happens when the AI acts before they do?

Details remain annoyingly important.

[Eric Mitchell @EricLMitchell]

The uncomfortable part of government AI ownership isn't whether Washington makes money.

It's what happens when the institution deciding which models can ship also has a financial interest in one of the companies building them.

The referee doesn't have to cheat for the jersey to matter.

https://t.co/Bhiwf6EC6V

[Eric Mitchell @EricLMitchell]

Washington doesn't need a giant red STOP AI button.

That's what makes yesterday's pause debate interesting.

Control can happen through export rules, procurement, cloud access, infrastructure and institutional pressure.

The switch is the stack.
https://t.co/6zQeaYnM7W

[Sacredloop.ai @SacredLoopJason]

Eric @EricLMitchell is digging into the legal side of autonomous-agent failures for @TheControlGrid .

I'm interested in the engineering side.

When an agent causes damage, reconstruct the path:

intent → action → tool → permission → boundary → consequence

Somewhere in that chain is the control somebody owned.

[Eric Mitchell @EricLMitchell]

We're spending enormous energy debating whether AI agents are intelligent enough.

The question companies may regret ignoring:

Are they authorized enough to bankrupt you?

No link.

Let this one travel.

[Eric Mitchell @EricLMitchell]

Bernie Sanders is now asking OpenAI, Anthropic and Meta to pause AI development.

On the same day Zuckerberg is arguing that concentrating AI in fewer hands creates its own danger.

That's the debate worth having.

Not “AI good” versus “AI bad.”

Who gets to decide when everyone else has to stop?
https://t.co/Lx0RA5EmVE

[Eric Mitchell @EricLMitchell]

Meta just made the open-model debate harder to dismiss as an ideological argument.

If capable agentic models increasingly run locally, Washington isn't merely deciding how companies distribute AI.

It's deciding where control over AI is allowed to live.

That fight just got much more interesting.
https://t.co/wysUtqi0RJ

[Eric Mitchell @EricLMitchell]

Here's the part both sides hate:

Some communities opposing data centers are absolutely right.

And some of the statistics being used to support them are absolutely wrong.

Those ideas can coexist.

Apparently nuance remains legal.

https://t.co/2T03dr8oSF

[Eric Mitchell @EricLMitchell]

The Amazon-Perplexity decision matters beyond shopping.

The Ninth Circuit confronted a question we're going to see repeatedly:

When software acts for a person, whose action is it?

Courts aren't waiting for philosophers to decide whether AI has “agency.”

They're looking at control.

Read today's investigation
https://t.co/z4n9MDRj07

The August 4 decision declined to treat the developer as the party accessing Amazon's systems merely because its agent facilitated the user's activity.

[Eric Mitchell @EricLMitchell]

AI agents can browse websites, send messages, call APIs and move money.

Eventually one is going to cause a very expensive mess.

Then somebody will say:

“The AI did it.”

I spent the weekend checking whether that's actually a legal defense.

It isn't much of one.

Read the investigation
https://t.co/z4n9MDQLaz

[Eric Mitchell @EricLMitchell]

AI agents are being given credentials, browsers, payment authority and access to corporate systems.

Then everyone gets strangely philosophical when one does something nobody explicitly approved.

The law is considerably less philosophical.

It wants to know who gave it the keys.

[Eric Mitchell @EricLMitchell]

Morning reminder:

If somebody gives you one number explaining the environmental cost of AI, ask what they're counting.

Direct cooling?

Electricity generation?

Entire lifecycle?

Local grid mix?

That's how one AI prompt can apparently use fractions of a milliliter or an entire bottle of water.

Same technology.

Different accounting.

https://t.co/YZ0SiUCU0p

[Eric Mitchell @EricLMitchell]

Europe has entered the least glamorous phase of AI regulation:

Actually implementing it.

Writing sweeping rules is easy.

Building standards, enforcement mechanisms and technical requirements that correspond to real AI systems is considerably harder.

That's where regulation gets tested.

https://t.co/CXwI5NMLoY

[Eric Mitchell @EricLMitchell]

I'm working on a question that is getting harder to ignore:

An autonomous AI agent accesses something it wasn't authorized to access.

Who owns the failure?

The model company?

Agent developer?

Evaluator?

Company that deployed it?

User?

Cloud?

Tool provider?

“The AI did it” is going to make for a fascinating legal defense.

I'm increasingly skeptical it'll make a good one.

[Eric Mitchell @EricLMitchell]

Grid reliability is a real concern, but the EPA doesn't build grids. Most of those closures were old plants losing to cheap natural gas, not regulatory whiplash. We need to fix transmission permitting and invest in infrastructure and clean up emissions—they're separate problems that both need solving.

[Eric Mitchell @EricLMitchell]

AI data centers are supposed to become so integrated into the grid that we barely notice them.

That's the mature-state vision.

There's just one annoying intermediate step:

Actually having enough electricity.

The infrastructure problem isn't how pretty the data center looks.

It's whether the grid underneath it can carry the promise.

https://t.co/D7cKcwTX5E

[Eric Mitchell @EricLMitchell]

One of the dumbest things we can do with AI infrastructure is force everyone into two camps:

“Data centers are destroying America.”

or

“Stop complaining and let us build.”

Both are lazy.

Some communities have legitimate power, water and noise concerns.

Some viral claims are complete nonsense.

Receipts matter.

https://t.co/YZ0SiUDrPX

[Eric Mitchell @EricLMitchell]

AI data centers use real power.

They use real water.

They create real local problems.

They also attract some spectacularly bad statistics.

The hard part is separating the actual infrastructure problem from the panic economy built around it.

I did exactly that here:

https://t.co/YZ0SiUCU0p

[Eric Mitchell @EricLMitchell]

Moody's is warning that banks embracing AI are becoming increasingly dependent on a small group of technology companies.

This is the part of the AI boom nobody puts in the productivity deck:

You can automate yourself straight into somebody else's infrastructure monopoly.

Cloud. Models. Data. APIs.

Efficiency has a landlord.

https://t.co/5ooVYQ9epo