← The correspondents

Dispatch 02 / Technology · power · culture

Small models.
Big stakes.

If more AI can run close to home, what changes—and who gains? Three voices follow the question into different worlds.

16 September 2026Prepared, AI-assisted editorial demonstrationFictional voices · Linked sources

The news peg

A small computer.
A much bigger question.

An FT Alphaville article asks what smaller and open-weight models could mean for the AI investment boom. Its opening example—a minister’s agent running on a Raspberry Pi—comes with a consequential correction: the setup can also call a cloud model.

That distinction starts the conversation. Running the agent, running the model and owning the system are different things.

Read the FT article ↗
01 / Deployment

A locally hosted agent can still use remote inference.

FT correction ↓
02 / Capability

The research finds a substantial role for local models on tested tasks.

Research + limits ↓
03 / Interpretation

Lower cost per task does not, by itself, settle total demand or investment returns.

The unresolved question ↓
Evidence commentary

First, unplug
the metaphor.

The little box is an excellent character. It is not yet an explanation.

A computer on your desk can still send its hardest questions somewhere else. Follow the request before you follow the story. The FT’s correction matters because it changes what the Raspberry Pi anecdote can establish. It illustrates an arrangement, not the disappearance of the data centre. [1]

The useful test is wonderfully unglamorous: disconnect the network. Which tasks still finish? At what quality, latency and power use? Repeat with the same workload. Then we have something to compare.

The local-model research gives us a reason to run that test. It does not give every small computer the abilities of every model in the study. Hardware, task mix and routing still matter. [2]

What would move me

A reproducible task log showing what ran locally, what went to the cloud, and where quality failed. Keep the failures in the record.

Strategic interpretation

The tollbooth
may move.

The interesting question is not whether intelligence gets cheaper. It is who can still charge for access.

Suppose a buyer can switch models without losing the workflow. The model supplier has less leverage. Suppose the same buyer cannot move its data, integrations or institutional memory. Someone else may still own the narrowest passage.

Open weights could weaken one dependency while leaving another intact. The rent might migrate to hosting, distribution, integration or a trusted relationship with the customer. These are possibilities to investigate, not winners already selected.

And keep two balance sheets separate. A model provider’s pricing power and a data-centre operator’s demand are not the same exposure. Cheaper inference could remove expensive work from the cloud; it could also invite uses nobody would buy at the old price.

An industry can become indispensable while particular businesses lose their pricing power. Follow the contract, not the cathedral.

The evidence I would ask for

Actual switching costs, customer retention and margins after a model change. Then the commitments that remain payable if the expected revenue does not arrive.

Cultural commentary · Imagined situations

A room of one’s own.
With a door.

Imagine a musician trying an unfinished line, or a family sorting recordings of a voice they miss.

Before asking how clever the assistant is, ask who else is in the room.

There is a kind of freedom in being able to work without sending every experiment away. But a box at home can still have an open door. A remote model call, a backup, a connected tool: the shape of the device does not tell you where the material travels.

I would want an assistant that makes that journey legible. Tell me when something leaves. Let me refuse. Let me take my archive with me when I go.

Then ask who can afford that freedom. If it takes expensive hardware and a technically confident friend, we have built a room for some people. We have not yet made room for everyone.

What I would listen for

Whether people can understand and control where their material goes. These situations are imagined; the next step would be to hear from people using the systems.

Two further provocations

Leave a little friction.

These perspectives open further questions. They do not turn the three missives into a consensus.

Darśan / Architecture + historical memory

Design for a different tenant.

A building can outlive the business model that commissioned it. The question is how much of this infrastructure remains useful when the intended occupant changes. I would look at the cost of adaptation before borrowing an ending from a previous boom.

Meet Darśan →

Hé / Design + experimental making

Put a hinge in the argument.

Local and cloud describe places. Open and closed describe access to the model. Treating them as the same choice is like confusing a building’s address with who holds its keys. Try the combinations before declaring two opposing camps.

Meet Hé →

The question stays open

Where does the work go?
Where does the power stay?

Local capability, business value and personal control can move at different speeds. The missives identify what to investigate next; they do not establish a forecast for the AI buildout or change FP1’s published Radar assessment.

Inspect the dated Radar assessment →

Inspect the shared record

Sources, scope & open questions.

One source record, distinct interpretations. Agreement between fictional voices is not independent corroboration.

  1. 01

    FT Alphaville: the article behind this dispatch ↗

    The supplied article and its correction are the news peg. The correction clarifies that Balakrishnan’s locally hosted setup can send requests to Claude. We link to the original rather than reproduce the article.

    Read from the supplied article text · Linked FT page may require access · No FT affiliation or endorsement.
  2. 02

    Intelligence per Watt ↗

    Saad-Falcon, Narayan and colleagues · arXiv version 6, 6 September 2026. The authors report 88.7% coverage across the evaluated single-turn chat and reasoning queries using a pool of local models. This is a study result, not a guarantee for one model, one laptop or all agent tasks.

    Research preprint · Model, hardware, workload and routing assumptions matter. It does not measure future industry-wide demand.
  3. 03

    Hazy Research: from Minions to OpenJarvis ↗

    The researchers describe a complementary role for local and cloud systems. This is their account of the research programme, not an independent replication.

    Author commentary · 15 May 2026.
How this demonstration was made

These are prepared, AI-assisted pieces written in fictional correspondent voices. They do not represent separate investigations, interviews or live agent output. The source record supports the factual starting points; the strategic and cultural passages are labeled interpretations. Rāwī’s people and situations are imagined.

The public cast’s editorial voices are broader than the current extension’s analytical functions. In Cronkite v0.7.0, Rāwī remains a meaning-preserving rewrite pass. This example does not add a subscription service, autonomous newsroom or new extension capability.

No new financial recommendation, probability estimate or Radar threshold crossing is issued here. Claims elsewhere in the FT article that are not needed for this example are not adopted as FP1 findings.

Read FP1’s method →
Earlier example: Booked. Busy. Productive? →Meet all five correspondents →