The Future Worth Building Is Human

Quick note on search intent: If you arrived here looking for the future worth formula used in financial analysis (also called future value), the future worth method in engineering economics, or a future worth calculator, see the short primer at the bottom of this article. If you are looking for the song "A Future Worth Dying For" by metalcore band Architects — including its lyrics or guitar tabs — that track appears on their 2012 album Daybreaker; dedicated tab and lyric sites will serve you better. The main article below addresses a third meaning: the philosophical and technical question of what kind of AI-enabled future is actually worth building for human society.

---

The future worth building is one in which humans don't just benefit from artificial intelligence — they actively shape it. That's the central argument of "The Future Worth Building Is Human," a long-form essay published on July 10, 2026, by Thinking Machines Lab — an AI research and product company that describes its mission as making AI systems "more widely understood, customizable and generally capable." The essay lands at a moment when the dominant tech narrative increasingly frames human involvement in AI as either a bottleneck to engineer away or a safety constraint to reluctantly tolerate. What makes the piece notable is that this critique of frontier-lab incentives comes from an organization building frontier models itself; it pushes back hard, grounding its case not in sentiment but in philosophy of knowledge, institutional economics, and pointed criticism of how today's frontier labs measure — and therefore optimize for — the wrong things.

(A note on names: an unrelated Manila-based data and AI consultancy also operates under the name "Thinking Machines." The essay discussed here is from Thinking Machines Lab, the frontier-model research company publishing at thinkingmachines.ai, and should not be confused with that consultancy.)

Why the Dominant Path Leads Somewhere Dangerous

Thinking Machines Lab opens with a diagnosis that will feel familiar to anyone watching the last two years of AI development: as the essay puts it, "Most AI in use today is trained in a handful of places and then frozen" — neither shaped by the people it serves nor able to learn from the collaborative work it participates in. The road most of the industry is racing down pushes toward centralization and autonomy, and it consistently frames human involvement as a trade-off — participation versus capability, ownership versus safe alignment.

The essay argues these are false trade-offs — not philosophical compromises to be accepted but engineering problems to be solved. That reframing is the load-bearing move in the entire piece. Once you accept that keeping humans meaningfully in the loop is a hard technical challenge worth taking on rather than a concession to caution, a completely different research and product agenda comes into view.

"The future is not a choice between human dominance and rapid obsolescence in the face of AI. Different roads lead to many different futures, and we get to choose which one to take."
— Thinking Machines Lab, "The Future Worth Building Is Human"

The Hayek Problem: Why Centralized AI Can't Know What You Know

The most intellectually substantial section of the essay draws on Friedrich Hayek's landmark 1945 paper The Use of Knowledge in Society to explain why centralizing intelligence at the AI layer is structurally prone to failure — not because AI isn't smart enough, but because of the nature of knowledge itself.

Hayek's argument, originally aimed at central economic planning, holds that productive knowledge is tacit, local, fleeting, and held privately by the people who do the actual work. As the essay characterizes it, "central planning fails not because of insufficient intelligence, but because of the nature of productive knowledge." No central planner can aggregate that knowledge fast enough or completely enough to make optimal decisions on behalf of everyone. The essay carries this argument directly into AI: a model trained on data generated in a handful of labs and optimized by a handful of alignment teams inherits a specific — and inevitably narrow — understanding of what is worth wanting and what is worth making.

Michael Polanyi's The Tacit Dimension (1966) supplies supporting evidence for the same principle: much of what makes a skilled professional, a functioning team, or an effective organization valuable simply cannot be written down or scraped from the internet. It lives in embodied practice and institutional habit, and — as the essay notes — is "constantly updated through feedback; it's not a static repository that can be written into a database." AI systems that cannot access that knowledge — and that aren't designed to elicit and integrate it — will systematically underperform in any domain where tacit expertise actually matters.

Where Autonomous AI Actually Works — and Why

The essay is careful not to be reflexively anti-automation. It identifies specific conditions under which AI operating without human oversight genuinely delivers value:

• Static, fully expressible goals — The objective can be written down completely and doesn't shift based on context or stakeholder values. In the essay's words, "the goal given to AI is static and expressible: to win a chess match, to prove a theorem."
• No hidden knowledge — All relevant information is either in the training data or in the input; there's nothing tacit a human participant could add. As the essay puts it, "the rules of chess and math are universal; the board is visible to all."
• Self-contained domains — Chess (where "the strongest engines are trained purely on self-play") and formal mathematics are the cited examples: closed systems where the rules are fixed and the winning condition is unambiguous.

Outside those boundaries — which describes almost every real organizational decision, healthcare interaction, policy choice, or creative project — intelligence alone isn't enough. The problem isn't that frontier models lack capability; it's that capability without access to local, tacit, human knowledge produces confident answers to the wrong question.

The Toyota Lesson: Mastery Requires Teaching the Machine

One of the essay's most striking empirical examples is Toyota's 2014 decision to bring expert craftsmen back onto production lines that robots had largely taken over — an episode the essay sources to Bloomberg reporting by Craig Trudell, Yuki Hagiwara and Ma Jie, Humans Replacing Robots Herald Toyota's Vision of Future (2014). The initiative was led by Mitsuru Kawai, and his explanation was precise:

"To be the master of the machine, you have to have the knowledge and the skills to teach the machine."
— Mitsuru Kawai, Toyota (as quoted in Bloomberg, 2014)

The point isn't romantic nostalgia for manual labor. It's that producing knowledge and applying intelligence are complements, not substitutes. Organizations that fully automate and eliminate the human knowledge-generation loop don't merely lose a warm feeling of craftsmanship — they lose the feedback loop that lets them improve the machines in the first place. Eventually, no one in the organization retains the expertise to diagnose why the machines are wrong, because the competence required to notice the error has been allowed to atrophy. This is a concrete, well-documented institutional failure mode, and the essay deploys it as a direct warning for AI strategy: the Toyota lesson generalizes from factory floors to any domain where humans hand over a complex process before they fully understand it themselves.

Two Unsolved Technical Problems Holding Back Human-AI Collaboration

Thinking Machines Lab identifies two specific engineering bottlenecks that prevent the participatory future they describe from being built today. Both are framed as tractable — difficult but solvable — which gives the essay its constructive rather than merely critical tone.

The Communication Channel Problem

The current dominant interface for interacting with large language models is, as the essay bluntly puts it, a communication channel amounting to "a small text box and a long wait." That bandwidth is far too narrow to carry the rich, contextual, often self-contradicting stream of intent that reflects how humans actually think through problems. People collaborate best when they can interrupt, correct, gesture, change their minds mid-sentence, and watch a collaborator react in real time — none of which the prompt-response paradigm supports well.

Thinking Machines Lab's proposed direction is a long-term bet on interaction models that handle live, multimodal collaboration natively inside the model itself — not in scaffolding bolted around a text-completion engine. The distinction has real practical weight. Scaffolding can approximate interactivity but imposes latency and architectural rigidity; a model that is natively interactive allows interactivity to scale with intelligence, so a more capable model becomes a more genuinely responsive collaborator rather than just a faster autocomplete tool.

The Evaluation Problem

The second bottleneck is more subtle but arguably more consequential: the AI industry is measuring the wrong thing. The essay takes direct aim at benchmarks such as the task-completion time horizon metric developed by researchers at METR — attributed in the source essay to Thomas Kwa, Ben West and colleagues in a 2025 paper, Task-Completion Time Horizons of Frontier AI Models — which tracks the time horizon of software tasks models can execute autonomously without human input.

The critique isn't that this benchmark is poorly designed for its stated purpose. It's that it only measures what AI can do alone, not what people and machines accomplish together. Optimizing for that metric creates systematic pressure toward autonomy over collaboration. Because the metric is what gets reported in leaderboards and lab announcements, it shapes investment, talent priorities, and product strategy across the entire industry. Measuring human-AI collaborative performance, the essay notes, "is more complex, and can't be done by a lab on its own" — it requires external, ecologically valid evaluation frameworks the field hasn't yet developed.

Dimension	Current Dominant Approach	Human-Centered Approach
Primary goal	Maximize autonomous task completion	Maximize human-AI collaborative output
Knowledge model	Centralized, aggregated at training time	Distributed, elicited from users continuously
Human expertise role	Input at prompt; otherwise bypassed	Active co-producer of model knowledge and values
Evaluation metric	METR time-horizon benchmarks (solo performance)	Joint human-AI task performance (not yet standardized)
Alignment method	RLHF / lab-directed value training	Model weights shaped by individual organizations and communities
Interface model	Text prompt + asynchronous response	Live, multimodal, natively interactive
Incentive structure	Lab benefits by absorbing user distinctiveness	Lab benefits when users leverage their own advantages

The Intelligence Curse and the Power Concentration Problem

The essay's most politically charged argument concerns what happens when AI can do most valuable work without human input. It cites a 2025 paper by Luke Drago and Rudolf Laine titled The Intelligence Curse:

"The social contract between corporations, governments, and citizens relies on individuals' productive capabilities on which the government's sovereignty and corporations' profits ultimately depend. Power that needs nothing from people loses the incentive to care for their needs and values, caring instead for its own preservation."
— Luke Drago and Rudolf Laine, The Intelligence Curse (2025)

This is a structural argument, not a conspiratorial one. The democratic accountability mechanisms that constrain how companies and governments behave depend, at a fundamental level, on those institutions needing something from ordinary people — their labor, their taxes, their votes, their spending. An economy where AI supplies most of that value unilaterally removes the material basis for the social contract. The essay doesn't predict this outcome as inevitable; it identifies it as the direction the current path tends toward, which is reason enough to want a different one.

The same structural logic extends to value alignment itself. Human values, like human knowledge, resist consolidation: they are plural, contextual, and often mutually contradictory in productive ways. Today, the values embedded in frontier AI models are decided by a small number of people in a small number of places. The essay references a document it attributes to Pope Leo XIV, Magnifica Humanitas — cited in the source essay as a 2026 text — for the point that "a more moral AI is not enough if that morality is determined by a few." Whatever one makes of that particular source, the structural argument holds independently: a single locus of value alignment, however thoughtfully managed, becomes a locus of power, and power concentrations invite capture by whoever controls them.

The proposed alternative is what the essay calls alignment as an ecosystem feature rather than a property of any single model: a diversity of AI systems raised in different contexts, shaped by different communities, disagreeing and competing and learning from one another. The analogy is to free speech and free markets — not as panaceas, but as mechanisms that surface competing ideas rather than forcing everyone to accept an averaged preference. This connects directly to the incentive argument: a lab that profits when customers leverage their own unique advantages has very different alignment incentives from one that profits by making customers more like each other. It's also why the essay draws a sharp distinction between renting AI from a vendor and owning and tailoring AI tuned to an organization's specific knowledge base — a distinction with direct implications for enterprise AI strategy, and one that echoes broader debates about who ultimately controls the direction of the most powerful AI systems.

Why Prompts Aren't Enough: The Case for Weight-Level Customization

One technically important argument in the essay is routinely glossed over in enterprise AI discussions: the difference between influencing a model through prompts versus shaping it at the level of model weights. Prompts, the essay argues, change surface behavior while leaving "deeper habits" intact. You can instruct a centralized model to behave in a culturally specific way through a system prompt, but the underlying assumptions about what constitutes a good answer, a reasonable decision, or a relevant consideration remain encoded from training — invisible, unchallengeable, and constant regardless of what the prompt says.

The essay references Gwern Branwen's Guardian Angels: LLM Personalization for Productivity and Security — cited as a 2026 essay in the source — to make a related point: allowing core model behavior to change significantly with prompts sacrifices safety, making a malleable centralized model "vulnerable to repeated attacks." The implication is that genuine value pluralism in AI requires genuine model pluralism — many models trained on different data by different communities, rather than one model wearing many prompt-configured masks.

This is a technically demanding vision. Training frontier models is extraordinarily expensive, and the idea of most organizations training their own frontier-scale models from scratch isn't realistic. What the argument does imply, practically, is a much stronger emphasis on fine-tuning, continued pretraining on domain-specific corpora, and reinforcement learning from human feedback sourced from within the organization — approaches that go substantially deeper than prompt engineering without requiring a build-from-scratch commitment. It also implies that the architecture choices labs make today — including how efficiently they support adaptation and how much of a model's reasoning is accessible to external feedback — are not neutral technical decisions. They determine what kind of participatory future is even technically possible downstream. This resonates with conversations in open-source communities, including the push toward open, adaptable model architectures that a broader range of actors can shape.

Key Takeaways

• Centralized, autonomous AI inherits a Hayekian knowledge problem: tacit, local expertise cannot be aggregated at training time, making human participation a functional requirement in most real-world domains, not an optional add-on.
• The dominant benchmark optimizes for the wrong objective: autonomous task-completion time horizons measure solo AI performance and create industry-wide pressure away from collaborative, human-in-the-loop design — precisely where most economic and social value is created.
• The communication bottleneck is real and under-resourced: the text-prompt interface is too narrow to transmit the richness of human intent; natively interactive, multimodal models are the necessary long-term architectural investment.
• Prompt-level customization is insufficient for genuine value alignment: deep behavioral habits are set in model weights, making weight-level fine-tuning and domain adaptation essential for organizations that want AI which truly reflects their knowledge and values.
• Power concentration in AI alignment is a structural — not merely ethical — risk: a single locus of value-setting, however well-intentioned, is a capture target; ecosystem-level diversity of models is the structural safeguard.
• The incentive structure of AI labs determines their actual alignment: labs that profit by absorbing user distinctiveness have systematically different — and arguably contrary — incentives from labs that profit when users leverage their own advantages.
• The Toyota lesson generalizes broadly: eliminating human expertise from a production loop doesn't merely displace workers; it destroys the organizational knowledge required to diagnose and improve the machines themselves.
• Model pluralism, not prompt pluralism, is the path to value diversity: a future worth building for everyone requires many AI systems shaped by many communities, not one system wearing many masks.

What Comes Next: Building the Infrastructure for Participation

The Thinking Machines Lab essay functions as a research and product agenda as much as a philosophical argument. The practical work it implies — natively interactive multimodal models, evaluation frameworks for human-AI collaborative performance, fine-tuning pipelines that organizations can genuinely own, and incentive structures that reward customization over homogenization — is large, expensive, and not yet mature. None of it will arrive in a single product release or a single funding cycle.

The Stakes of Getting the Infrastructure Right

What the piece accomplishes is articulating why this work is worth doing in specific, non-sentimental terms that resonate with engineers and product teams: not because AI without humans is frightening, but because AI without humans is structurally limited and concentrates power in ways that undermine the conditions for its own legitimacy. The Hannah Arendt framing that closes the original essay — drawn from The Human Condition (1958) and its insistence that "man himself becomes the ultimate end," never merely a means to a more efficient outcome — is the philosophical gloss on an argument that stands without it. As the writer of this piece reads Arendt, the moment human beings become instrumental inputs rather than authors of collective action, the political conditions for genuine freedom dissolve; applied to AI, this suggests that systems optimized purely to replace human judgment, rather than to augment it, erode the very agency that gives social outcomes their legitimacy.

The future worth building isn't human because of nostalgia or sentimentality. It's human because the knowledge is there, the values are there, and no centralized system — however intelligent — can substitute for the irreducibly distributed nature of both. The question for the remainder of this decade is whether the labs, enterprises, and policymakers shaping AI development will build the infrastructure to make genuine participation possible, or whether the path of least resistance toward autonomy and centralization will lock in a series of choices that prove far harder to reverse than they appeared when they were made. As debates over who controls digital infrastructure intensify globally — from platform lock-in concerns in Europe to open-source model governance — the argument for a pluralistic, participatory AI ecosystem is only going to grow more urgent.

---

Appendix: Future Worth in Financial and Engineering Contexts

For readers who arrived here searching for financial or engineering definitions, here is a concise reference.

Future Worth Formula

In engineering economics and financial analysis, future worth (also called future value) is the value a present sum of money will reach after earning interest over a specified period. The standard future worth formula is:

FW = PW × (1 + i)^n

Where:

• FW = Future Worth (the value at the end of the period)
• PW = Present Worth (the value today)
• i = interest rate per compounding period (expressed as a decimal)
• n = number of compounding periods

Future Worth Method

The future worth method is a project-evaluation technique used in engineering economics. Instead of converting all cash flows to a present value, analysts convert all cash flows to an equivalent value at the end of the study period. A project is economically justified if its future worth is greater than or equal to zero (assuming a minimum attractive rate of return has already been embedded in the discount rate). The method is mathematically equivalent to net present value analysis but is preferred in contexts where end-of-period comparisons are more intuitive — for example, comparing the terminal balance of competing capital investments.

Future Worth Analysis

Future worth analysis applies the future worth method to compare mutually exclusive alternatives. Each alternative's cash flows — initial investment, operating costs, salvage values — are converted to a single equivalent future amount. The alternative with the highest (or least negative) future worth is selected. A future worth calculator automates this conversion for uniform annual cash flows using the uniform series compound amount factor (F/A, i, n).