Volume 36: A Maine Mill Town Wrote the Better Playbook

The data center fight is showing up in 300+ state bills, on Fox Business, and in towns most of us have never heard of. The headlines are loud and one-sided. Underneath them is a more complicated story about water, power, breakthroughs we have not seen yet, and one Maine mill town quietly writing a better playbook.

🧭 Founder's Corner: Why the loudest voices in the data center debate are misreading the room, and what the Town of Jay, Maine is showing the rest of us about how to build.

🧠 AI Education: Why fine-tuning a model on your company's data almost never works the way the vendor pitched it, and the cleaner mental model that does.

✅ 10-Minute Win: A three-question rubric for evaluating any new AI tool, with Claude doing the research and you keeping the final call.

Let's get into it.

1) OpenAI's Altman says AI unlikely to lead to 'jobs apocalypse'

Summary: Speaking in Sydney, OpenAI CEO Sam Altman said AI would not lead to a global "jobs apocalypse" and the technology had not claimed as many white-collar jobs as he had feared. 

Why it matters: The CEO who fueled the loudest version of the fear narrative is now walking it back. The 2026 picture is reshaped roles, not eliminated ones, and the professionals learning to work with these tools are still the ones winning.

2) Amazon starts selling its AI shopping technology to other retailers

Summary: Amazon is licensing the technology behind Alexa for Shopping to outside retailers through AWS, with Tapestry-owned Kate Spade signed on as a customer to launch a gifting assistant. 

Why it matters: The cloud-computing playbook is repeating for AI shopping. If you sell anything online, the question is shifting from "should we have an AI assistant" to "whose AI agent runs our storefront."

3) Weekly Rundown—Moffitt Cancer Center expands Reimagine Care's virtual oncology model

Summary: Moffitt Cancer Center is expanding its AI-enabled virtual oncology program with Reimagine Care after early results showed nearly 7,000 patient interactions with 97% independently resolved without escalation to providers and only 2.4% of interactions resulting in emergency department referrals.

Why it matters: Cancer care between visits has historically been a black box for both patients and clinicians. This is real-world evidence that AI plus a virtual care team can handle the in-between moments without sending people to the ER, and it sets a measurable bar for any specialty care AI rollout.

4) Your AI agent can now trade for you on Robinhood. And buy stuff with your credit card too

Summary: Robinhood launched Agentic Trading and an Agentic Credit Card, letting customers connect third-party AI assistants to carry out investing strategies or spending with minimal human involvement, with push notifications, spending limits, and optional manual approval. 

Why it matters: This is the first mainstream consumer product where an AI agent moves real money, not just researches options. Expect the same approve-as-you-go model to show up next in HR, procurement, and benefits platforms.

5) Beyond the EHR: Why one CIO believes AI will dwarf every prior health IT shift

Summary: Health system CIO Lundal told Healthcare IT News that artificial intelligence has the potential to fundamentally alter nearly every aspect of healthcare delivery and administration, and unlike EHR rollouts that followed an established roadmap, AI strategies can now shift in a matter of months. 

Why it matters: Many leaders are still planning AI in two-year horizons. This CIO's point is the strategy itself needs a faster cycle, and the organizations treating AI like a one-time deployment are already falling behind.

On Fox Business in May 2026, Kevin O'Leary suggested that two women from Utah might be working for the Chinese government. Their offense was organizing local opposition to his $100 billion Stratos data center in Hansel Valley. The project would cover 40,000 acres, more than twice Manhattan's footprint, and draw 9 gigawatts, more than the entire state uses in a year.

He called them "proxies for the Chinese government" and taunted, "come out, come out wherever you are." The reasoning, as Fortune reported it: anyone slowing American compute must be working for the only adversary who would benefit.

Rather than hide, Gabi Finlayson and Jackie Morgan laughed, mocked his flip-flops-with-a-suit look, and let the internet make him the punchline. Within days the CEO of O'Leary's own firm, Paul Palandjian, walked back the accusation, telling Business Insider the company accepted that the women were American political strategists.

The insult collapsed under its own weight. What I keep returning to is the gap behind it: the distance between a hot-mic accusation and the careful clarification the same firm issued days later. The pushback he tried to paint as foreign was not even partisan. Utah's own Republican governor, Spencer Cox, had already pressed the project for a water plan to protect the Great Salt Lake. When the loudest voice reaches for a spy story, something in the conversation has already gone wrong.

The Concerns Are Real, and So Is the Upside

Water is where the worry usually starts, and the numbers explain why. Some hyperscale facilities draw up to 5 million gallons a day, the Environmental and Energy Study Institute reports. The power grid is straining under the same demand. In the PJM region, the grid serving 65 million people across 13 states, supply costs jumped from $2.2 billion to $14.7 billion in a single year, and Brookings attributes close to two-thirds of that climb to data centers. Those costs land on the monthly bill. In Utah, residential electricity rates rose 15.2 percent in twelve months, the third-largest jump in the country, according to the Energy Information Administration.

And this is not a fringe reaction. Gallup found that 7 in 10 Americans do not want an AI facility built in their own community. The opposition O'Leary tried to cast as foreign is, in fact, most of America.

For all that, the same buildout could underwrite the breakthroughs we have been waiting on for a generation. That possibility is the part the headlines rarely hold long enough to examine.

What If We Measured Both Sides?

This debate runs on a lopsided ledger. The costs are concrete and immediate. Gallons, megawatts, acres, the tax revenue a town gives up, all of it landing on a meter or a balance sheet. The benefits are harder to pin down, because most have not arrived yet, and a benefit that has not arrived is easy to value at nothing. So we tally the cost in full and the upside at zero. What would it look like to weigh both columns, even when one of them is still a guess?

In May 2026, the upside stopped being hypothetical. In 1946, Paul Erdős posed a deceptively simple question. Scatter points on a flat plane, and how many pairs can sit exactly one unit apart? For decades, the square grid seemed to be the best anyone could do. Then a researcher fed the problem to an OpenAI reasoning model, a general-purpose system rather than one built for mathematics, and it found a better arrangement by importing tools from algebraic number theory that no one had thought to connect to the geometry. As Scientific American described it, "After 80 years of fruitless struggle by human mathematicians, a major geometry conjecture has at last been solved." Experts told OpenAI the proof would have earned a spot in a leading journal even if a human had written it. The thing that solved it was a chatbot.

Read that and it is tempting to think the machines had finally outsmarted us, on an ordinary Tuesday. So I went looking for the asterisk, and the people closest to the work supplied it. Human mathematicians had to clean up the model's output before the proof held. The model had proved a better arrangement existed without working out how much better, and a Princeton mathematician, Will Sawin, pinned that down. Sébastien Bubeck, who leads OpenAI's mathematical work, put it plainly. The model "did not invent something fundamentally new that nobody saw coming. It just executed like an amazing mathematician."

None of this proves AI will solve everything. What it proves is smaller and more interesting. This infrastructure, paired with serious people, can reach places neither would reach alone. What else might sit behind that door? Drug discovery, climate modeling, materials we cannot make yet. The honest answer is we do not know, and sitting with that uncertainty beats pretending we already know which way it breaks.

The Better Case No One Is Making

The companies building these data centers keep losing the argument for them. I wrote in February about the labs fumbling their message at the Super Bowl, and the same instinct is now playing out on concrete and steel rather than ad buys. Lost in the noise is the better case, the one about jobs, tax revenue, and the kind of infrastructure a town keeps long after the trucks leave. The bar is on the floor. Showing a community what it stands to gain, and giving it a real stake in the outcome, would do more than any press release.

Two thousand miles east of Hansel Valley, one town has been quietly making exactly that case. The Town of Jay sits in Franklin County, Maine. For generations the Androscoggin Mill carried the local economy, until a 2020 boiler explosion began its decline and Pixelle Specialty Solutions left in 2023, taking the rest of the jobs and 22 percent of the town's tax base with it. Earlier attempts to revive the site went nowhere. So the town spent two years working to put a data center where the mill once stood.

That patience is starting to pay off. According to Governor Mills' April 24 veto message, the $550 million project would bring more than 800 construction jobs, at least 100 high-paying permanent positions, and substantial property tax revenue. What sets it apart is how it would be built. Rather than break new ground, the developers plan to reuse the mill's existing buildings, water, and electrical infrastructure, which, as Mills noted, would avoid the very impacts on the environment and on ratepayers that drive the opposition elsewhere. The project is already under contract, has cleared several permits, and carries the backing of the town and the wider region.

What makes Jay matter is the governance. Mills did three things at once. She vetoed the statewide moratorium that would have killed the project, signed a separate bill stripping data centers of state tax breaks, and created a council to recommend rules for the rest. Her reasoning held both truths. A moratorium, she wrote, can be "appropriate given the impacts of massive data centers in other states," but a blanket ban would have killed a project her own community wanted.

I do not want to oversell this. One hundred permanent jobs is fewer than the mill once supported, and critics have said so fairly. But the lesson was never in the job count. It is in the process. Two years of local work, infrastructure reused instead of land broken, governance written in rather than bolted on after a fight. That is the rare outline people on different sides could actually sign.

Bigger Than One Town

Jay is one town. The same fight is unfolding in statehouses across the country. MultiState counted over 300 data center bills in the first six weeks of 2026, spread across more than 30 states, with outright moratoriums proposed in 11. So far not one has cleared its originating chamber. The country is deciding how to build this one legislature at a time.

When this reaches a ballot, the question is rarely AI or no AI, but how to build it. A community can push for governance written into the buildout, keeping some leverage over water and power, or back a moratorium that freezes everything while federal rules take years to arrive. By the time Washington acts, the window may have closed.

And the stakes reach well past any single town. The United States and China are effectively the only two countries building AI infrastructure at full scale, and none of us can read China's actual hand. We are being asked to weigh in on something with national consequences while missing half the picture. In November, some of this lands on ballots as local and statewide measures, and the rest rides on the lawmakers voters send to write the rules.

Most of us will never have a data center proposed in our own town. But the story will keep finding us anyway, in the news, on a ballot, or in what the people we elect decide to do about it. The work is not to pick a side from the headlines, but to understand it well enough to see what they leave out. Some of these buildouts will be great for local communities. Some will not be. There are no absolutes when we are all living through a transformative time in human history. No headline will sort that out for us, and the loudest voices only make it harder to see. That leaves the work to us. Information is power, especially in conversations this ambiguous. The clearer we see it, the better the choices we make, in November and long after.

The Assumption

Here is the belief most professionals hold about fine-tuning: take a generic AI model, train it on your company's documents, and now it knows your business. The hospital's protocols. The bank's compliance rules. The law firm's case history. The assumption is reasonable. "Training" is how humans transfer knowledge to other humans, and the word carries the same intuition when applied to AI. Every vendor pitch promising to fine-tune the model on your data reinforces it.

Where It Breaks Down

Consider what happens when a hospital actually tries this. A clinical operations team partners with a vendor to fine-tune a large language model on thousands of pages of internal protocols, formulary documents, and clinical guidelines. The delivered model sounds completely different from the base version: right vocabulary, hospital voice, internal forms cited by name. Then a clinician asks a real question about a real protocol, and the model invents an answer that is not in any of the source documents. The more new facts the vendor pushed in, the more confidently the model fabricated ones that did not exist.

What Is Actually Happening

Fine-tuning does not put knowledge into a model. It shapes how the model behaves.

A pre-trained model has learned most of what it knows before any team touches it. Pre-training is where the model first encounters language, the world, and how concepts relate. It runs at a scale fine-tuning does not approach. By the time a team gets access to a model like Claude or ChatGPT or Gemini, the heavy lifting is done.

Fine-tuning is a smaller, targeted process that comes after pre-training. It teaches the model patterns: how to format an answer, what tone to use, when to refuse a request. Pre-training is everything a pharmacist learned about medicine. Fine-tuning is the day-one orientation at a new pharmacy: the intake script, the escalation path, the patient counseling style.

A 2024 Google research paper presented at EMNLP, the leading natural language processing conference, studied what happens when fine-tuning tries to teach a model genuinely new facts. The finding was clear. Models learn new facts very slowly during fine-tuning. And as those new facts get absorbed, the model's tendency to hallucinate increases in direct proportion. The paper's conclusion: factual knowledge mostly comes from pre-training, and fine-tuning teaches the model to use that knowledge more efficiently.

OpenAI's own documentation now reflects this reality. The use cases listed in their official fine-tuning methods table are classification, translation, tone, style, and instruction-following corrections. Not knowledge. The same documentation also notes that OpenAI is winding down its self-service fine-tuning platform. That is not an accident. The industry is converging on a clearer picture of what fine-tuning actually does.

The Revised Mental Model

Here is the cleaner mental model: fine-tuning changes the shape of the answer. Retrieval changes the substance.

Wrong format, tone, or style is a fine-tuning question. Information the model does not have is a retrieval question, often solved by RAG, which we covered in Vol 19 through 22. A one-time adjustment for a single task is a prompting question.

The reframe changes three things at work. The question to ask a vendor shifts from "will you fine-tune it on our data" to "how does this system access our information at the moment of the response, and how is that information kept current." The success criterion for internal AI pilots shifts away from "the model knows our policies" toward "the model can retrieve our policies on demand." The timeline shifts too: fine-tuning projects take weeks of data preparation, while retrieval and prompting changes can be tested in an afternoon.

What to Watch For

• Vendor demos that promise "fine-tuning on your data" as the answer to a knowledge problem. The right question back is how the system retrieves information at the moment of the response, not how it was trained.
• Internal AI projects where the success criterion is "the model knows our internal documents." That is a retrieval problem wearing the wrong jersey.
• Outputs that sound right but cannot be traced to a specific source document. Confidence is not a signal of accuracy. It is often a signal that fine-tuning was used where retrieval would have been correct.
• Industry signals like OpenAI winding down its self-service fine-tuning platform. The market is converging on prompting and retrieval as the primary levers for most professional use cases.
• Any AI pitch that conflates "training" with "knowledge." The two are related, but they are not the same thing, and the difference now matters at every vendor meeting.

How This Connects

This is the first of a three-part series on fine-tuning. Vol 3 introduced data and labels as the building blocks of how any model learns. Vol 10 walked through how a large language model is born from the training process. This series picks up where those volumes left off and walks into the room where the actual decisions get made today. Next week, Vol 37 covers RLHF, the most consequential fine-tuning method ever developed and the reason every commercial model you use responds the way it does. Vol 38 closes the series with a practical decision framework for choosing between fine-tuning, RAG, and prompting for any problem at work.

Part 1 of 3 in the Fine-Tuning series.

The 3-Question Tool Test

Another AI tool launches on LinkedIn. Three more drop on YouTube the same day. A coworker forwards a newsletter with five more. Most of them are not transformative. You either chase everything and finish the week with eight half-configured accounts, or ignore everything and miss the two tools that would have changed your work.

The 3-Question Tool Test breaks the cycle. Three questions you answer before signing up for anything: Does it solve a pain I actually have? Does it work inside my flow? Does it earn the time? Honest answers give you a clean decision: adopt, save for later, or skip.

The twist: you use Claude (or any LLM) to run the analysis while you keep the judgment. Claude researches the tool, does the time math, and stress-tests its own answer. You bring your real pain, your real workflow, and the final call.

The Workflow

1. Pick a tool you have been curious about (1 minute)

Open the tab, post, or newsletter that triggered the curiosity. Have the tool's name and website ready. Knowing only the marketing copy is fine. Claude fixes that in the next step.

2. Get a plain-English briefing from Claude (2 minutes)

Open Claude at claude.ai. Replace [TOOL NAME] with your actual tool and paste:

Copy/Paste Prompt: "In under 200 words, tell me: (1) what [TOOL NAME] does in one sentence, (2) the top three use cases real users adopt it for, (3) the pricing tiers, and (4) the top two complaints from real users in reviews or forums. Write plainly, no marketing language."

Two minutes of grounded research replaces twenty minutes of skimming the tool's homepage.

3. Run the 3-question evaluation (4 minutes)

Stay in the same Claude conversation so it has the briefing as context. Fill in the bracketed sections with real specifics and paste:

Copy/Paste Prompt: "I am evaluating [TOOL NAME] using a 3-question rubric. Walk through each question with me honestly. Q1: Does it solve a pain I actually have? My specific pain right now is: [DESCRIBE IN 1 TO 2 SENTENCES]. Does this tool directly remove this friction, or is it adjacent? Q2: Does it work inside my flow? My current tools and daily workflow look like: [LIST 3 TO 5 TOOLS AND HOW YOU MOVE BETWEEN THEM]. Where would this tool fit, and does it integrate or create a copy-paste tax? Q3: Does it earn the time? Estimate realistic time saved per week if I used it actively, conservatively. Then estimate the learning curve hours to get there. Give me a 90-day ROI."

The specificity in the brackets is the work. Generic answers produce generic evaluations.

4. Pressure-test the evaluation (1 minute)

LLMs agree too easily. Paste:

Copy/Paste Prompt: "Now play devil's advocate. Based on everything I said above, give me the single strongest reason I should NOT adopt [TOOL NAME]. Do not soften it."

If Claude's strongest counterargument feels weak, that itself is useful information.

5. Make the call and log it (2 minutes)

Three outcomes.

Adopt: sign up now and schedule a 30-day review on your calendar.

Save for later: note the tool in a running list and name the trigger that would change your answer.

Skip: write the tool name and the question that failed it. Future you will be tempted by the same tool in three months, and the log saves the re-evaluation.

The Payoff

You walk away with a decision on one tool, a transferable framework for every future tool, and a small log that compounds. Next time a colleague forwards a hot new tool, you have a structured workflow instead of an open spiral. The rubric is yours; the engine is Claude.

The AI Concept You Just Used

AI-assisted decision triage. Most people use AI to write or summarize. Fewer use it as a thinking partner for decisions they would otherwise make on instinct. The pattern is portable: define your rubric, give the LLM specific context to apply it, ask it to push back on its own answer, then make the call yourself. The same shape works for evaluating vendors, conferences, and book recommendations.

Transparency & Notes

• The rubric is generalizable. It applies to productivity software, browser extensions, and most paid subscriptions, not only AI tools.
• Claude's research in Step 2 is a starting point, not a citation. If the decision is significant, click through to real reviews and pricing pages.
• The devil's advocate step is the most often skipped and the most valuable. Skipping it makes the rubric a rubber stamp.
• The framework is conservative by design. The cost of missing one useful tool is lower than the cost of accumulating twenty abandoned ones.