AI Reality

Move faster without pretending the costs are invisible.

AI is useful. It is also physical: electricity, water, land, chips, policy, private data, vendor dependency, and communities asked to absorb infrastructure they did not design.

The answer is not to retreat from the tools. The answer is to get more deliberate: local where possible, remote where justified, private by default, and sovereign enough that your work still belongs to you.

Try local AI Read the position

The reality

This is not cloud magic. It is infrastructure.

Responsible AI starts by refusing the comforting story that prompts happen nowhere. The better question is simple: which work deserves a remote data center, and which work should happen on the machine in front of you?

Power

The bill is not abstract.

AI demand is now part of the electricity planning conversation. Even when the global share looks small, the local load can land hard on a specific grid, utility queue, or rate base.

Water

Cooling choices become community choices.

Data centers need to move heat. In some places that means water demand, wastewater considerations, and pressure on local utilities that were not built around massive compute loads.

Privacy

The prompt is part of the system.

AI work often starts with private context: customer records, financial details, source code, strategy, health-adjacent notes, legal drafts. Where that context goes matters.

Agency

Convenience can become dependency.

If every workflow requires a remote model, a remote account, and a vendor policy you do not control, your most important tools can change underneath you overnight.

Existential risk

The wildest number in AI is not a benchmark.

P(doom) is shorthand for the probability that AI produces an existentially catastrophic outcome. The unnerving part is not one estimate. It is the spread: people with serious technical, financial, and moral skin in the game place their numbers from effectively zero to almost certain.

Schrodinger would have had a field day. Until the box is opened, AI is both civilization-scale leverage and civilization-scale threat. The rational move is not theatrical certainty. It is engineering with containment, observability, reversibility, and local control wherever those controls are possible.

The answer

Local-first AI is not nostalgia. It is leverage.

A local model running on your own hardware will not replace every frontier model call. It should not. But it can take over a surprising amount of daily work: reading drafts, sorting notes, checking tone, classifying documents, summarizing private context, and helping a team think before anything leaves the building.

That changes the relationship. You are no longer asking whether every thought belongs in a remote system. You are deciding, case by case, when outside intelligence is worth the trade.

Open source proof

mere.run is this argument as software.

Sawfwair's open-source mere.run project is a Swift package, CLI, and optional macOS studio for local-first AI on Apple Silicon. It is the practical alley this page points toward: capable tools, local model storage, and private work that can happen on hardware you control.

Today's public quickstart is macOS-first and validated on Apple Silicon. That constraint is honest. Sovereign AI starts by being real somewhere, then widens from there.

Visit mere.run View source package Read docs

macOS quickstart

curl -L https://mere.run/releases/mere-run.dmg -o mere-run.dmg
open mere-run.dmg

swift run mere.run model capabilities
swift run mere.run text anonymize "private context"
swift run mere.run api serve --engine text-chat-gemma4

local text chat, code generation, embeddings, and PII anonymization
local image generation, OCR, vision inspection, grounding, segmentation, and tracking
local speech synthesis, transcription, music, and video generation
Hugging Face-backed model pulls into a shared local model store
a loopback OpenAI-compatible API for supported text engines

First-party demo

Local AI is not mysterious.

We can run a 500 MB language model from Google in the browser right now. First cache Gemma while online. Then run a prompt immediately, or turn off Wi-Fi first for the stricter proof.

WebGPU Checking this browser...

Device hint Hardware details stay in your browser.

Network Online, ready to cache or run

No account, no server round trip for the prompt, and no mystery wrapper: after the first cache step, the language model runs in this tab.

sawfwair local lab Gemma 4 E2B / ~500 MB first run

Runtime Ready. Nothing is fetched until you choose Cache Gemma.

Step 1

Cache Gemma in this browser

Do this while the network is on. The browser keeps the model files in its cache for this tab.

Step 2

Run a prompt locally

For the strongest proof, turn off Wi-Fi first. You can also keep it on: after caching, the prompt runs in this tab.

Online allowed

Prompt

Output

Step 1 fetches Gemma 4 E2B into the browser cache. Step 2 runs the prompt locally. For a stricter proof, disconnect first; for convenience, keep Wi-Fi on and run now.

How to stay ahead

Adopt the tools. Keep your footing.

Rule
Use remote frontier models when the task truly needs frontier capability.
Rule
Use smaller local models for routine drafting, classification, search, review, and private context work.
Rule
Keep sensitive documents on hardware you control unless there is a clear reason to send them out.
Rule
Design AI features with an off-ramp: exportable data, replaceable models, observable costs.
Rule
Treat existential-risk estimates as decision inputs, not slogans. Do not outsource your priors to the loudest certainty.
Rule
Treat energy, water, and privacy as architecture constraints, not public-relations footnotes.

What Sawfwair does about it

Five percent of every engagement plants trees.

Sawfwair sets aside 5% of every engagement for tree planting and restoration work at the Brookside Agroforestry Research Center in Hazelbrook, Prince Edward Island.

It does not make compute impact disappear. It is a standing choice to connect digital work back to soil, watersheds, habitat, and open field research close to home.

The sources behind the stance.

IEA Energy and AI report Global data center electricity demand, local grid concentration, and 2030 projections.Lawrence Berkeley Lab 2024 U.S. Data Center Energy Usage Report U.S. historical and projected data center electricity demand.EPA: Water and Data Centers Cooling, water reuse, utility coordination, and community water considerations.NIST AI Risk Management Framework Risk framing for trustworthy AI systems and generative AI governance.P(doom) Shorthand for estimated probability of existentially catastrophic AI outcomes.mere.run Sawfwair open-source local-first AI toolkit for Apple Silicon Macs.mere.run source package Swift package, CLI, and optional macOS studio for local inference.Gemma Gem Browser extension reference for Gemma 4 E2B/E4B running locally with WebGPU.