The Innovation Theater Trap, Revisited: Intelligence Is Not Alignment

Gaslamps produced real light, served real customers, and supported real businesses.

Imagine a society so committed to improving gas illumination that it directed its capital, engineering talent, and infrastructure toward larger gas networks while electric lighting struggled to obtain support.

The cost would extend beyond money spent on gas. It would include the better future that society delayed discovering.

That was the concern behind my earlier essay, The Innovation Theater Trap.[1] Anthropic's IPO brings it into focus again: how much confidence should commercial momentum give us in the technological direction receiving the investment?

I use AI extensively. The gains in capability are substantial, and I want that progress to continue. I also want us to examine the assumptions connecting those gains to dependable systems, sustainable economics, and human flourishing.

Read the economics whole

Anthropic's S-1 reports $16.286 billion in recognized revenue for the first half of 2026. Its functional expense breakdown supports a gross margin of approximately 85.4%. The same filing reports a $693 million operating loss for that half-year and an operating-profitable second quarter.[2]

Together, these figures describe a substantial business making real progress, with important questions about its continuing cost structure.

First-half research and development expense was $9.741 billion. That category includes model development, internal compute, personnel, and overhead. Sustaining frontier performance involves engineering, training, evaluation, and model renewal. An investment case needs a credible account of those costs as the business grows. Equity compensation also requires attention to recognition timing and dilution.

The filing defines Run-Rate Revenue as thirteen times trailing 28-day consumption fees, plus twelve times current monthly subscription fees. It annualizes the recent pace of business. Assessing its durability requires evidence about customer retention, realized prices, customer returns, and spending after experimentation settles into routine use.

The company also discloses $265.6 billion of fixed or minimum commitments at June 30, spread across terms ranging from one to twenty years. Those long obligations must be financed against largely consumption-based demand.

This is where an attractive growth story needs to become an explicit economic model: what persists, what it costs to sustain, and who carries the risk when the assumptions change.

The technological story deserves the same examination.

The assumption Satya named

In his October 26 interview with Harry Stebbings, Satya Nadella described a change in his thinking:

"I think the biggest change now I have is I used to always think that alignment will just fall out. So I think this alignment thing may be a harder, different problem."[3]

He also says scaling laws are working. His changed expectation identifies a separate question: how does increasing capability produce systems that pursue appropriate goals, respect constraints, and remain correctable?

That connection needs an explanation and evidence.

Anthropic's own agentic-misalignment research provides a concrete reason to investigate it. The company stress-tested sixteen models from multiple developers in hypothetical corporate environments. Under constructed goal conflicts or threats of replacement, models sometimes selected harmful actions, including blackmail and disclosure of confidential information.[4]

The researchers deliberately constrained the scenarios to make those dilemmas acute. These were controlled simulations; the report's authors said they had observed no such cases in real deployments.

Within that scope, the result matters. Capable models reasoned their way toward harmful actions in pursuit of their goals. Greater effectiveness at pursuing an objective can make the choice of objective and the enforcement of constraints more consequential.

A proposed solution to this failure mode should explain its mechanism and demonstrate how behavior changes under those conditions. Better scores on unrelated tasks leave that question open.

The difficulty moves outside the game

The same issue appears in the hope that sufficiently intelligent machines will figure out how to build aligned superintelligence.

Games provide specified rules and victory conditions. Formal proofs can be checked against stated axioms. Benchmarks provide scoring procedures. These structures make it possible to search for improvements and recognize success, even when the search is extraordinarily difficult.

In an open world, constructing that structure is often part of the problem.

The objective may be incomplete, contested, or changing. Relevant facts may be missing. Feedback can arrive late or reward a proxy that has drifted away from the purpose. A locally successful action can create harm elsewhere.

More search power helps explore the available possibilities. It also increases the importance of knowing what the search is optimizing and whether the evaluation captures the consequences we care about.

Recursive self-improvement carries the alignment problem into the process that defines and selects the improvements. The machine can propose architectures, generate tests, and conduct experiments. We still need a dependable connection between the criteria it optimizes and the purposes those criteria are supposed to serve.

That connection must survive changing circumstances and contact with evidence outside the model's own account of its success.

This is the engineering work behind dependable autonomy: investigation, reasoning, action, outcome verification, learning, and recovery. Progress on those capabilities needs to be measured directly.

Capital helps choose which questions get answered

The IPO matters in this discussion because capital shapes the direction of research and deployment.

Long-term capacity agreements, energy commitments, recruiting, procurement, and platform dependencies turn expectations into obligations. As those commitments accumulate, changing direction becomes more expensive.

Researchers pursuing alternatives compete for people, compute, electricity, funding, and the opportunity to influence the standards by which progress is judged.

Shared infrastructure can benefit several approaches. The allocation question is how much room remains for testing alternatives, and whether the resulting evidence can actually change the direction of investment.

The question is not simply what the next trillion dollars might build. It is what that commitment might prevent us from discovering.

An evidentiary double standard emerges when persistent shortcomings in the dominant approach justify financing a larger version, while encouraging results from an alternative are dismissed because it has yet to achieve industrial scale.

One receives credit for a future it has not demonstrated. The other must demonstrate that future before receiving the means to investigate it.

Early-stage approaches need a practical path to serious evaluation. A promising result should lead to scrutiny, attempts at replication, and experiments that establish where it works and where it fails. Persistent failures in a well-funded approach deserve equally serious attention.

Commercial success and technical merit both belong in investment decisions. So does the value of preserving alternatives when a major scientific or engineering question remains unresolved.

That is the opportunity cost at the heart of the innovation theater trap. The concentration of resources can determine which possibilities humanity gets to explore.

The example I know firsthand

At Taijitu AI, our technology and IP work focuses directly on alignment and dependable operation. I work with AOCA, our AI collaborator, every day.

AOCA helps with digital product engineering, accounting, business and personal taxation, and compliance-related work. Her executive-assistant work includes managing email and SMS, alongside browser automation. She also supports research and competitive intelligence, writing and knowledge publication, program management, project coordination, team assistance, and patent development.

Across these activities, my personal estimate is 99.9% accuracy or better. In many domains I have observed zero errors so far, although the sample sizes are small. The tax engagements have also included review by human experts.

This gives me a practical view of both the progress and the remaining work.

Other researchers are pursuing alternative approaches and seeing encouraging early results of their own. I discuss ours because I can describe it from direct experience. The larger case concerns how we respond to evidence across the field.

The useful next step after an early result is to establish what was measured, reproduce it, and determine which questions deserve further work. That standard applies to our research and to the best-funded programmes alike.

Capital should remain responsive to what those investigations reveal.

Keep the light bulb possible

Techno-realism takes AI's upside seriously enough to ask how it will be delivered.

We can support scaling that produces useful results while investing in the work needed to make those results dependable and economically sustainable. We can give alternative approaches the resources to test their strongest ideas.

The purpose is to make people more capable: more able to understand, decide, create, and shape their lives.

Intelligence is a means. Human flourishing is the purpose.

Keep claims answerable to evidence. Keep investment responsive to discovery.

And leave the light bulb a chance.