Vision and Values

Cost of the Average

Signal and Strategy

There is an economic shift happening. Generating code is easier, faster but not necessarily cheaper or more effective.

Can AI lift you above the average? No. If everyone has access, it is by definition not a differentiator. You need a strategic advantage.

Hall says the work is now in the signal layer. That is where human insight, vision, and judgment become a strategic advantage: knowing a problem closely, imagining a better outcome, and choosing the tradeoffs worth making to get there.

Gippity and Claudius do not share your values or your vision. Their priorities are shaped by the guys that claim "AI will destroy humanity".

Researchers tested strategic advice from leading LLMs. Across thousands of simulations, recommendations kept converging on fashionable strategies even as the business context changed. They call it “strategy trendslop”. trendslop isn't "wrong" persay, but if everyone has the same strategy, what makes it work better for you? What's your strategic advantage?

A strategy starts with a direction, and choices under uncertainty. Who are we serving? What will we do for them? What are we willing to sacrifice?

To answer these you need vision and values. Vision to see the effect of potential outcomes and values to evaluate and measure those effects.

Fluency Feels 'Right'

AI knows everything about your customer. It's easy to have multiple threads running market research, product analysis, technical planning, and it's easy to spend hours doing this. In fact, this is often easier than talking to your customer.

“the end game for all … RLHF models is optimizing for engagement”

— Diogo Almeida, ChatGPT and InstructGPT coauthor, What's next after RLHF?, 7:38–7:56.

His point: these assistants are designed to make you "feel good". They are useful and at the same time optimized to maximize engagement.

Almeida argues that optimizing for human approval can work against dependable automation. A useful agent sometimes needs to stop, expose uncertainty, or tell me my plan is bad. To me, and I'm not a Gippity cocreator, but seems like the same logic applies to other problem spaces besides automation.

Furthermore, we bring our own baggage along when crafting solutions with AI.

Our feeble human psychologies are fraught with fallacy and bias. Gippity and Claudius aren't designed to help guard against that. In fact, incentive structures suggest otherwise.

Take for example the personal validation effect. A CHI 2026 study showed that participants will be inclined toward the "favorable" AI response.

There is a psychological shift. AI isn't decietful it doesn't have intentions. It is optimized for engagement and to make you feel good, which according to the research affects our objectivity when evaluating answers.

A response can feel meaningful, but feeling meaningful is not the same as making a meaningful difference.

What Language Carries—and What It Leaves Out

What happens when an AI set's your aim? Besides subverting your values and its sycofantic misdirections, what else it might it lack?

Consider Jon. “Jon earns $72,000” describes his salary. “Jon feels underpaid” tells us the salary does not fit what he thinks the work should provide. It does not tell us whether he wants more income, recognition, time, or autonomy.

For this article, human value means how an outcome is experienced as beneficial, harmful, meaningful, or worth protecting. A value statement expresses part of that relationship in language. An operative priority is a ranking enacted in behavior or policy, whether or not anyone stated or authorized it.

A model can recite principles and behave according to learned or instructed priorities. That does not establish that it experiences those values or has authority to impose them. Human values concern what outcomes mean in people's lives. Operative priorities concern what a system rewards, protects, or sacrifices.

We do not need to upload Jon's entire inner life before helping him. We do need to ask what “better” means before building him the wrong thing. Share enough to form a responsible hypothesis, make the tradeoffs visible, and notice when reality disagrees.

Jon's “I want something better” branches toward legitimate futures:

Once those meanings are on the table, we can ask the harder question: which future should govern the work?

Vision Gives Direction. Authority Makes It Governing.

What you aim at determines what you see.

A vision is a picture of a future worth working toward, shared with others and open to revision. It gives us direction before we have every goal and plan worked out. “To the moon” is enthusiastic. It still leaves us asking who benefits, what it costs, and who agreed to the trip.

An inspiring vision does not automatically give its author the right to impose it. Four questions keep that authority attached to the people it affects.

Suppose our company offers a book-publishing service. We might help Jon turn his teaching materials into a book, reach more educators, and earn additional income without taking on the full work of publishing. That is a promising offer, not an authorized goal. We still need to learn whether Jon wants it, whether family time is a governing constraint, whether we can deliver it, and whether the benefit justifies the cost.

A goal names a condition worth creating or preserving. Once the root goal is supplied, it becomes valuable to explore opportunities, compare solutions, and test whether an intervention produces the intended consequence. Strategy coordinates those commitments over time.

Even better reasoning cannot choose the destination for us. Judea Pearl's ladder of causation distinguishes association, intervention, and counterfactual questions. AI can help at every rung when the evidence and causal model are available. The ladder answers increasingly powerful causal questions; it does not add a fourth rung called “what should we value?”

A model may infer a goal from behavior or propose one from public patterns. Fluency, prediction, and causal competence can support a decision. They do not create standing to decide whose interests matter.

This is a normative argument, not a laboratory result. Evidence can tell us what happened. It cannot decide, without a governing judgment, whose experience should count or which sacrifice is legitimate.

A System Can Be Coherent and Still Be Wrong

Now for the irritating part: we can state our priorities clearly and still get them wrong. Every deployed AI system sits inside a hierarchy shaped by training data, post-training, system instructions, organizational policy, tools, permissions, and evaluation. People and institutions choose how the surrounding product orders and enforces those influences.

Someone chooses what counts as a good response, whose feedback matters, and how disagreements get resolved. “People preferred it” tells us something. It does not settle whether the result is true, valuable, authorized, or something we are willing to answer for.

A system can please evaluators, pass its tests, or obey its instructions while optimizing the wrong target. Our publishing service could celebrate books, pages, and revenue while Jon earns little after fees and loses more evenings to work. The dashboard would report success while the value proposition failed.

AI may identify that contradiction, but it cannot overrule the governing system unless people have given it permission to challenge, escalate, or stop.

We need a way back. Jon must be able to say the offer is making his life worse, and someone must have the authority to change it. Observe the consequences, include the people paying the cost, protect disagreement, and make sure an escalation can actually change the goal, metric, or boundary. Otherwise we have built a very efficient way to keep being wrong.

Human governance is no guarantee of wisdom. People choose bad goals, protect status, and rationalize harm. The reason to preserve human responsibility is not that people are always better than models. Organizations act through social authority, and those affected need somewhere to direct consent, challenge, blame, repair, and revision.

Shared Intent Lets Teams Think for Themselves

None of this means the boss needs to approve every move. That would turn human judgment into another queue. A strategy that only one person can interpret is not much use to the rest of the organization.

Suppose product learns that Jon accepts a longer timeline if it protects his evenings, while operations finds white-glove support unsustainable. Given only “grow book revenue,” one team may overpromise while the other removes the help that made the offer valuable.

Shared intent gives them a basis for different but compatible choices. The vision is not simply “publish more books.” It is to help educators earn from existing expertise without creating a second full-time job. Teams need the desired future, affected people, constraints, decision rights, visible commitments, evidence that should force reconsideration, and protected dissent.

Write that down, and teams have something they can use and challenge. Why does this action make sense? Whose needs have we missed? What outcome would show that the offer failed? A longer plan cannot answer those questions by itself.

Inside those boundaries, AI can research, compare explanations, generate options, and make bounded instrumental decisions. Removing a person from each action does not remove human responsibility for the objective, permissions, evaluation, escalation, or consequences.

Teams can choose different routes when they understand the destination, their boundaries, and how to change course. Shared intent gives them room to think. It should not become a polite name for obedience.

Conclusion — Retain Authority Over the Ends

Six hours, all my credits, and a discarded worktree did not prove that AI cannot help with strategy. What I had left undefined was what deserved the time, what was good enough, and what should be left alone. I handed over those tradeoffs with the task and called it “optimization.”

Unstated values do not disappear. A system imports priorities from its training, instructions, tools, evaluation, and request. Stated priorities can still be wrong. Coherence does not close the values question.

The work is to name the future we want and why it matters, make clear who can commit us to it, and give teams enough shared intent to act. Then listen when the evidence—or the people living with the consequences—says we got it wrong.

I still want the machinery. I would also like to keep my credits next time. But the advantage has to come from understanding a problem, choosing a future worth pursuing, and staying responsible for what happens when we pursue it. “Optimize this plan” cannot do that work for us.

This article sets the AI Factory's direction.

Understanding and Bottlenecks asks how to organize work when generation outruns our capacity to evaluate and understand it.