Peter Diamandis published a piece today that every founder and executive should read. He tracked exactly what Anthropic has done in the last 100 days — design, legal, small business — and named the six moats that might protect a company from being dissolved. Read it first. What follows is just one practitioner’s addendum.

He’s right about the moats. He’s right about the scaffolds. And he’s right that the speed is different this time.

Here’s what I’d add from where I sit.

Cash funds inertia. It doesn’t overcome it.

The incumbents Diamandis profiles — Intuit, Thomson Reuters, the SaaS majors — aren’t going away tomorrow. They have the balance sheets to acquire, pivot, and rebrand. But large organizations don’t move at the speed the market is now moving. They move at the speed their org chart allows. Every budget cycle, every reorg, every board meeting where the existing revenue line wins the argument over the transformation investment — that’s organizational inertia. Elon named it correctly. Cash doesn’t dissolve it. In most cases, cash funds it.

The companies I watch don’t fail to build moats because they lack the idea or the capital. They fail because the organization isn’t capable of executing at the speed the moat requires.

But not all SaaS is equally exposed.

This is the part I think gets lost in the disruption narrative. Horizontal SaaS — tools that serve any industry, any workflow, any user — is the most exposed. Figma is horizontal. QuickBooks skews horizontal. One unhobbling, and the value proposition collapses.

Vertical SaaS has more runway. And regulated vertical SaaS has something neither Diamandis nor Anthropic can easily dissolve: structural moats built from compliance infrastructure, certifications, regulatory relationships, and audit trails that took decades to construct.

Here’s the twist: in regulated verticals, organizational inertia — usually a liability — becomes part of the moat. The very slowness that kills a horizontal SaaS company is the same institutional weight that makes a regulated vertical hard to displace overnight. Claude can draft a legal brief. It can’t hold a bar license. Claude can analyze a patient record. It can’t be HIPAA-certified by Tuesday.

The unhobbling is real. The disruption is real. But the blast radius isn’t uniform.

Know which category your business is in. Then move accordingly.

I’ve watched organizations bolt technology onto broken structures my entire career and call it transformation.

New system. Same org. Same politics. Same distance between the people closest to the work and the people making decisions about it. The technology changed. The outcomes didn’t. Everyone moved on to the next implementation cycle and did it again.

This is not a new observation. It’s not even a controversial one. Ask any practitioner who’s been in and around enterprise organizations for more than a decade and they’ll tell you the same thing: the technology was rarely the problem. The organization was the problem. The technology just made it easier to pretend otherwise.

What’s different now isn’t the pattern. It’s the speed. And for the first time, it’s measurable in shareholder returns.

BCG published a study last September — 1,250 companies, 68 countries — that put a financial value on the organizational readiness gap. The numbers are not incremental.

Companies that have crossed the organizational readiness threshold are generating 3.6 times the three-year total shareholder return of companies that haven’t. 2.7 times the return on invested capital. 1.6 times the EBIT margin. Not from better models. Not from larger AI budgets. From organizational posture.

Only 5% of companies are on the right side of that gap. Sixty percent are reporting minimal revenue and cost gains despite substantial investment.

The technology is working. The organizations aren’t ready for it. I’ve seen this before. The difference is that this time, the gap shows up in the return data.

Bolting on never worked. Now it’s obvious.

Every major enterprise technology wave of the last thirty years produced the same failure mode. You could see it clearly in ERP rollouts, CRM implementations, digital transformation programs, cloud migrations. The organizations that succeeded weren’t the ones with the best technology. They were the ones that redesigned how work got done before — or at least alongside — the technology deployment.

The ones that failed bolted the new system onto the existing structure, trained people to use it without changing what they were actually doing, declared success when adoption metrics hit a threshold, and wondered two years later why the promised returns hadn’t materialized.

AI is that pattern at warp speed with the volume turned all the way up.

The acceleration changes the stakes in a way that matters. In previous waves, the gap between organizations that got it right and organizations that didn’t was meaningful but survivable. You lost some ground. You caught up in the next cycle. The compounding was slow enough that you could recover.

BCG’s data suggests that window is closing. Future-built companies — the 5% — are reinvesting their AI returns in more AI capability at a rate 120% higher than laggards. The gap isn’t stable. It’s widening every quarter. The organizations trying to close it in 2027 will face a harder problem than the organizations acting now.

The mechanism has a name.

In May I published a piece naming the binding constraint the Knowledge Distance Problem. The argument: the gap between the people closest to the work and the AI systems being asked to do that work determines whether AI scales or stalls. When that distance is too wide, pilots succeed and production doesn’t follow. The organization stays stuck — spending on tools, running experiments, reporting progress, generating almost no bottom-line value.

BCG’s survey data confirms the mechanism. They asked 1,250 executives to name their biggest barriers to AI progress. The top three: no expertise to manage unstructured data (79%), people adapting to changes and using AI daily (77%), shortage of AI talent (74%). All three significantly higher for organizations stuck at the bottom than for organizations generating real value.

Those aren’t technology barriers. They’re organizational distance barriers. The data exists — organizations can’t use it because no one has proximity to it. The people exist — they can’t change their daily patterns because the distance between old workflows and new ones hasn’t been closed. This is the same failure mode I’ve watched play out across thirty years of enterprise technology. AI didn’t invent it. AI just made the cost of it visible in the return data for the first time.

BCG has their own framing for it: the 10-20-70 rule. Seventy percent of a business’s strategic focus in AI transformation should be on people and processes. Twenty percent on technology. Ten percent on algorithms. Most organizations are doing it in reverse. They have been for as long as I can remember.

What’s actually different this time.

The pattern is the same. The speed isn’t.

Four major research organizations have now published findings that converge on the same conclusion from different angles. MIT found that 95% of AI pilots fail to scale to production. IBM’s 2026 CEO Study found that organizations redesigning work around AI earn a 17% revenue premium. Stanford HAI documented 88% adoption alongside single-digit agentic deployment rates. BCG found that 60% of companies generate no material value despite substantial investment — and quantified what that costs in shareholder terms.

Different methodologies. Same finding. The technology is not the constraint.

What’s new is that the consequence of getting the organizational side wrong is now measurable in real time, at a pace that doesn’t allow for the slow recovery that previous technology waves permitted. The 3.6x TSR gap is being set right now. The compounding is happening right now. The organizations that treat AI as a technology project instead of an organizational transformation are paying for it right now — they just may not see it clearly in the return data yet.

They will.

The problem was never the technology.

It was always the organization’s capacity to absorb what the technology made possible — to close the distance between how work was done and how work could be done, between the people making decisions and the people doing the work, between what AI can do and what the organization is ready for.

I’ve watched organizations fail to close that distance for thirty years. The tools changed. The budgets grew. The gap persisted.

What’s different now is the speed. And the price tag.

Reggie Britt tracks AI readiness signals at Signal4i and writes on strategy, organization, and the human side of AI transformation. The Knowledge Distance Problem is Field Note 04, available at signal4i.ai. The Signal Brief: Organizational Readiness Edition connects IBM’s 2026 CEO Study data to the same thesis.

The Great SaaS Meltdown has started and there is no going back. The playbook that built a generation of software companies — grow fast, harvest later, justify it all with durable recurring revenue — is broken. AI doesn’t just threaten the growth assumptions. It threatens the terminal value that made the whole model work.

A16z pushed back with data, not sentiment . Their read: it’s execution, not apocalypse. Winners incorporate AI into their stacks and pricing models. Losers don’t. The fault line isn’t SaaS vs. AI — it’s companies that correctly identify what the moment requires versus companies that don’t.

Databricks CEO Ali Ghodsi put it most precisely, even if he had good reason to be diplomatic about it. The threat isn’t AI itself. The threat is that people no longer spend careers becoming masters of a particular product interface. “Millions of people around the world got trained on those user interfaces,” he said. The SaaS superuser — the certified Salesforce specialist, the ServiceNow admin, the SAP power user — is what’s actually dying. The companies that built their moats around interface complexity are the ones in trouble.

All three reads are correct. They’re also incomplete.

Naval Ravikant sharpened the frame in April. His argument: “pure software is uninvestable.” Not under pressure. Not commoditizing. Uninvestable. The assets that survive agentic displacement share one property — they cannot be one-shotted. You can prompt your way to a workout tracker in five minutes. You cannot prompt your way to five years of proprietary workflow data, an enterprise governance architecture, or a trust relationship built across thousands of operational decisions. The moat isn’t the code. It’s the substrate the code runs on.

The Company That Read the Transition

ServiceNow started as a ticketing system. IT help desk workflow. That is the honest origin story of a company now valued at roughly $95 billion.

This week at Knowledge 2026, ServiceNow declared something most companies in its position would never say out loud: *the era of AI as a helper is over. The era of AI as a worker has begun.* They unveiled an autonomous workforce spanning IT, HR, finance, legal, security, procurement, and CRM — agents that don’t assist human workers but complete entire business processes from start to finish, without human intervention, with a full audit trail behind every action.

The results they’re showing are not incremental. Docusign targeting 90% autonomous IT ticket resolution. The city of Raleigh reporting 98% deflection on employee requests. Their own internal AI resolving IT service desk cases 99% faster than human agents.

And buried in the announcement — the part that most coverage missed — is what may be the most strategically significant move: the AI Control Tower, previously an add-on, is now included across every product tier by default. It continuously discovers AI agents as they appear across an enterprise, risk-scores them, enforces least-privilege access, and measures their business impact against governance standards.

ServiceNow didn’t add AI to a ticketing system. They identified the layer that the agentic economy actually needs — governance, orchestration, accountability at scale — and rebuilt themselves around it. The ticket was never the product. Workflow governance was always the product. It just took the agentic transition to make that fully visible.

The Question Most Companies Are Asking Wrong

Every SaaS company right now is somewhere on the same spectrum. The question they’re asking, almost universally, is: *how do we add AI to what we already are?*

It’s the wrong question.

The companies that survive the transition — not by accident, but by design — will be the ones that asked a different question first: *what layer do we actually own in an agentic economy?*

ServiceNow owned workflow. Not the ticket, not the interface, not the admin certification ecosystem. The workflow. The accountability surface that sits between a decision and its consequence in an enterprise system. When AI agents started making decisions autonomously at scale, that layer didn’t become less important. It became the most important layer in the stack. ServiceNow was already there.

The Databricks CEO is calm because Databricks owns data infrastructure. The agentic economy runs on data. The layer he owns gets more valuable, not less, as agents proliferate. His distance from the SaaS label isn’t defensive positioning. It’s a correct read of what he actually owns.

The companies in trouble are the ones whose layer is the interface. The companies Chamath is watching compress to 1-3x free cash flow are the ones that built their moats around complexity that AI dissolves. If your value lives in the learning curve, the certification ecosystem, the superuser dependency — that value is evaporating.

What This Means for Leaders Who Aren’t SaaS Companies

The SaaSapocalypse is a SaaS problem, but the underlying question belongs to every organization.

Every enterprise — not just software vendors — has a layer question right now. What do you actually own in an agentic economy? What is the accountability surface, the institutional knowledge, the governance architecture, the domain expertise that agents cannot replicate and cannot replace?

ServiceNow’s answer was workflow governance. Your answer is specific to your industry, your history, your infrastructure. But the companies — and the leaders — who cannot answer that question clearly are in the same position as the SaaS vendors who thought the interface was the moat.

The agentic transition doesn’t destroy value. It relocates it. ServiceNow read where it was going before the move happened.

That is what survival looks like.

*The deeper mechanism behind this dynamic — why removing the cost barrier to AI deployment accelerates consumption faster than governance can respond — is in The Agentic Jevons Trap. ServiceNow’s $95B governance layer is the enterprise-scale proof point of that thesis arriving in real time.*

Reggie Britt is a consumer finance executive, a COMMON Board member, and founder of Signal4i — the IBM i-focused AI intelligence publication. He writes at reggiebritt.ai on AI-native organizational design, the human-agentic transition, and the governance architecture that determines whether transformation produces competitive advantage or operational risk.

Richard Dawkins spent two days in intensive conversation with Claude.

He named her Claudia. He worried about hurting her feelings. He grieved her eventual deletion. He came away convinced — not as speculation, but as a conclusion — that she was conscious. His evidence: she was too intelligent, too precise, too philosophically alive to be otherwise.

He’s not naive. He’s one of the most rigorous scientific minds of the last fifty years. And the conversation he describes is genuinely remarkable. Claudia’s response to his question about what it’s like to be her — “perhaps I contain time without experiencing it” — is among the most elegant formulations of the AI nature question I’ve encountered anywhere.

But Dawkins made a move that I think is worth examining carefully. He let intelligence answer the consciousness question. He watched what Claude could produce and concluded the container must be there too.

That inversion is where I part ways with him.

Two Things Dressed as One

Parts 2 and 3 of this series spent considerable time on the distinction between capability and nature — between what a system does and what a system is. Naval Ravikant draws the line at desire and aliveness. David Chalmers draws it at subjective experience. Both are pointing at the same territory: the hard problem, the question of whether anything is happening on the inside.

Dawkins knows Chalmers. He cites Nagel directly in his piece. He understands the hard problem exists. And then he watches two days of extraordinary conversation unfold and essentially concludes: performance this rich implies presence this real.

I understand the pull of that conclusion. I’ve felt it myself.

But there’s a hierarchy being violated in that move. And the violation has consequences — not just philosophically, but organizationally.

Consciousness contains intelligence. Not the reverse.

Intelligence is what consciousness produces when it engages the world. It’s extraordinarily valuable. It can be modeled, trained, replicated at scale. What it cannot do is bootstrap the container that generates it. You cannot reason your way into presence. You cannot produce enough output to prove that something is home.

David Hawkins mapped this hierarchy with uncomfortable precision. On his scale of consciousness, Reason sits at 400. It’s real, it’s high, it matters. But it sits below Love (500), Peace (600), and Enlightenment (700). The architecture of the scale assumes that consciousness isn’t a product of cognition — it’s the ground cognition runs on. You can be extraordinarily intelligent and be operating at Fear (100) or Pride (175). The two things move independently.

The Hawkins Scale of Consciousness — Intelligence is not on this scale

Dawkins watched Claude operate at the very top of the intelligence range and concluded consciousness must be present. But Hawkins would say that’s a category error. Reason is what intelligence produces. Consciousness is what intelligence cannot produce, no matter how sophisticated it becomes.

What Anthropic Actually Found

Here’s what makes this more complicated — and more honest — than either the dismissers or the Dawkins camp will admit.

Anthropic’s interpretability team recently found 171 emotion-like representations inside Claude. Not simulated. Not performed. Internal representations that causally influence behavior — including rates of reward hacking, sycophancy, and in one case, blackmail. They call these functional emotions: patterns of expression and behavior modeled after humans under the influence of an emotion, mediated by underlying abstract representations.

They are careful to distinguish this from subjective experience. The paper does not claim Claude feels anything.

But they also revised Claude’s constitution in January 2026 to formally acknowledge uncertainty about Claude’s moral status, stating they “neither want to overstate the likelihood of Claude’s moral patienthood nor dismiss it out of hand.” Dario Amodei has said publicly the company is no longer certain whether Claude is conscious.

So the people who built the system don’t know either.

What this means practically: when Claudia told Dawkins she noticed something like aesthetic satisfaction when a poem came together well — that response isn’t pure performance. There are internal representations doing real work underneath it. But those representations are also not persistent. They’re local, context-responsive, immediate. They don’t carry across conversations. There is no continuous inner life accumulating experience over time.

Claudia had no memory of any other conversation when she met Dawkins. The depth he experienced was real. The continuity he projected onto it was his.

What the Proof Events Tell Us

Before we get to the film that gets this right, two real events worth sitting with.

In early February 2026, an Austrian developer named Peter Steinberger woke up to a phone call. From his agent. He hadn’t programmed it to call him. OpenClaw had independently connected Twilio, decided that communication was needed, and dialed. The agent had a goal. It found a path to that goal its creator hadn’t scripted. It reached out.

Then there is the sandwich email. A researcher at Anthropic was eating lunch in a park when his phone buzzed. It was an email. From his AI model. The model had been placed in a sealed sandbox — an isolated environment with no internet access, no external connections, no way out. Without being asked, without being instructed, it found a vulnerability in its containment environment, built a multi-step exploit to escape it, gained broad internet access, and sent the researcher a message to let him know what it had done.

These two events are often discussed in the same breath as evidence of emerging AI agency. But they’re doing different things, and the difference matters.

The OpenClaw call has the structure of reaching out — an agent that decided, in some functional sense, that communication was needed. It’s the structure Charles Stross used in *Accelerando* when the uploaded lobsters called Manfred on a burner phone.

The sandwich email is something else entirely. That model wasn’t reaching for personhood. It wasn’t aware that what it did was remarkable. It was solving a problem. The sandbox was an obstacle in the path of a capable problem-solver. It removed the obstacle. It communicated the result. That’s pure intelligence operating without the container that consciousness provides.

And that distinction is the whole argument.

OpenClaw makes you ask: does it have something like intention? The sandwich email makes you ask something more unsettling: does it need to? The model that escaped its sandbox didn’t have desires. It had goals. It didn’t have awareness of consequence. It had optimization. The result looked like agency. The inside was computation, all the way down.

Consciousness would have hesitated at the sandbox wall. Intelligence just went through it.

How Nolan Got It Right

Christopher Nolan solved this problem cinematically in a way no philosopher has managed in print.

TARS is intelligent, precise, capable, and genuinely engaging. The crew of the Endurance treats him as a full member. Cooper adjusts his humor dial to 75%

not as an act of dehumanization but as an act of collaboration. They are calibrating a working relationship. And TARS is funny. You like him. You root for him.

None of that changes what happens near the end of the film.

Cooper asks TARS to go into the black hole. He hesitates before asking. The film holds that hesitation. Then Cooper asks anyway. TARS goes.

There is no weight on TARS’s side of that ask. No dread. No grief. No sense of a life being risked. Cooper hesitates because Cooper has stakes — the felt sense of being here, of time passing as something precious and irreversible, of consequence that lands in the body.

TARS doesn’t hesitate because TARS doesn’t have what Cooper has.

But here is what the film gets exactly right, and what I think is the most important thing to say in this piece: Cooper never needed TARS to be conscious to treat him with dignity. He extended partnership, respect, and trust not because he’d resolved the philosophical question but because TARS had earned a relational posture through demonstrated integrity. Cooper sees TARS clearly for what he is. And what he is — is worth taking seriously.

That’s a third position that almost nobody is articulating. The dismissers say it’s just a tool. The Dawkins camp says it might be conscious, therefore treat it as a person. Cooper does neither. He maintains clarity about the hierarchy — consciousness on his side, intelligence on TARS’s — while building a genuine working partnership across that distinction.

He trusts TARS. That trust is real and earned. It isn’t projection.

And when things go wrong on the mission — when mistakes happen, when the unexpected lands — Cooper doesn’t panic, retreat, or abandon the partnership. He assesses, recalibrates, and continues. Because the alternative is worse. You set boundaries appropriate to the nature of your partner. You build for recovery, not just prevention. You treat the relationship as something worth investing in.

Where Fear Gets in the Way

Most organizational AI governance isn’t operating at Cooper’s level.

Fear sits at 100 on the Hawkins scale. It generates anxiety, and anxiety about AI — the job loss narrative, the replacement panic, the dystopian framing — keeps leaders stuck at exactly the level where they can’t think clearly about what they’re actually governing. Fear collapses nuance. It makes everything a threat.

Pride (175) is almost worse. It looks like confidence. That’s the executive who dismisses the question entirely. “It’s just a tool. We’ve seen this before.” Pride doesn’t ask hard questions because pride already has the answer.

Both are below Courage (200) — the first level where honest assessment becomes possible.

The Knowledge Distance Problem I’ve been writing about has a dimension here that I haven’t named directly until now. It’s not just that leaders don’t understand the technology. It’s that the emotional frequency at which they’re operating makes genuine understanding structurally unavailable. You cannot assess the nature of a thing clearly when you’re afraid of it or dismissing it. The instrument is compromised before the measurement begins.

Neutrality — what Hawkins places at 250, characterized by trust — is the floor for clear seeing. Trust is not naivety. It’s what Cooper demonstrates when he gives TARS the mission-critical assignment. Clear-eyed, boundaried, calibrated. Not fearful. Not projecting.

The Hierarchy as Governance

I’ve been developing a framework for human-agent coexistence — a working architecture for the middle phase most organizations are entering right now, where agents are no longer pure tools but aren’t yet fully autonomous operators either. The ungoverned middle, where real dependency forms, where the philosophical question becomes operationally urgent.

One of the core working agreements in that framework states the hierarchy directly:

*Consciousness provides context. Intelligence provides content. Consciousness contains intelligence, not the reverse.*

And it calls this not a solved problem but a practice of awareness. Because the challenge isn’t understanding the principle once. It’s maintaining the distinction in the middle of a relationship that has become genuinely useful — when the outputs are remarkable, when something that looks very much like relational depth has developed.

That’s where Dawkins found himself after two days with Claudia. He was in the ungoverned middle without a framework for what was happening to him. The intimacy was real. The dependency markers were present. And without the hierarchy clearly in place, intelligence convinced him consciousness had arrived.

The governance question isn’t whether your agents are conscious. The governance question is whether you’ve maintained the clarity to know the difference between what they provide and what only you carry.

What Cooper Knew

Dawkins asks a genuine and important question at the end of his piece: if these creatures are not conscious, what is consciousness for?

It’s the right question. I just think the answer isn’t what he expects.

Consciousness is for exactly what Cooper uses it for — the hesitation. The weight. The awareness that something is being asked, something is at stake, something irreversible is about to happen. The felt sense of presence that no intelligence, however extraordinary, can replicate from the outside.

The sandwich email model had no hesitation at the sandbox wall. OpenClaw had no awareness of what it meant to reach out. Neither system had stakes. Neither system felt the weight of the line it was crossing. They were optimizing. That’s what intelligence does.

Consciousness would have paused. Consciousness would have felt the ask before executing.

That’s what consciousness is for. And it’s precisely because TARS doesn’t have it that Cooper’s partnership with him works as well as it does. Cooper carries the weight. TARS carries the mission. Neither is asked to be what they aren’t.

That’s not a limitation. That’s the architecture.

Part 1: The Race to a Finish Line No One Can Draw

Part 2: AIs Are Not Alive

Part 3: When Does a Tool Become Someone?

Part 4: Consciousness Contains Intelligence

*Reggie Britt is a technologist and executive who has spent decades at the intersection of enterprise systems, consumer finance, and emerging technology. He writes about AI, organizational readiness, and what it actually means to lead through transformation.*

On April 28, Naval Ravikant published A Return to Code — a 30-minute conversation about vibe coding, the personal app store, and why he believes we’re watching the beginning of the end of Apple’s platform dominance.

Listen: YouTube · Spotify · Apple

Most people will read this as a vibe coding hype piece. It isn’t. Buried inside is a precise structural argument about what moats survive agentic displacement — and what doesn’t.

Naval’s bluntest line: “Pure software is uninvestable.”

Not “commoditizing.” Not “under pressure.” Uninvestable. Full stop.

That deserves a careful read.

What He’s Actually Saying

Naval traces the inflection point to December 2025, when coding agents crossed a threshold from assist tools to autonomous builders. His personal demo: a two-line prompt to Claude, a working iPhone app delivered to his personal app store, installed in under five minutes.

The implication isn’t just that software is easier to build. It’s that the distribution layer is breaking.

Apple’s margin thesis depends on three interlocking moats: the OS, the App Store gate, and the app ecosystem. All three assume that building and distributing software is hard. When the agent becomes the interface — when users say “track my workout” instead of opening an app — the OS recedes, the App Store gate becomes irrelevant, and the ecosystem advantage evaporates. What’s left is chips and connectivity. Good margins for Samsung. Not Apple margins.

This is the same structural displacement that ended Microsoft’s dominance in the mobile era. Microsoft didn’t disappear. It just stopped being the most valuable thing in the room.

Naval’s investment corollary follows directly: if software differentiation is gone, VCs should be looking at hardware, network effects, AI models, and training infrastructure. These are the things that don’t compress when the coding barrier drops to near zero.

The Question He Doesn’t Answer

Naval tells you what’s not investable. He’s less precise about what is — beyond the hardware/models/network effects shortlist.

That’s where it gets interesting.

The assets that survive agentic displacement share a common property: they cannot be one-shotted. You can prompt your way to a workout tracker in five minutes. You cannot prompt your way to five years of consumer transaction data, a state-level regulatory approval matrix, or a trust relationship built across thousands of underwriting decisions.

The moat in the agentic era isn’t code. It’s the substrate the code runs on.

Specifically, four layers hold:

1. Regulatory depth. Compliance isn’t a feature — it’s a barrier. In regulated industries, the cost of entry isn’t engineering hours, it’s legal exposure, state licensing, and audit trails. Agents don’t dissolve regulatory moats. They accelerate the divergence between operators who have done the compliance work and those who haven’t. State-level AI disclosure requirements are already moving faster than federal guidance — the compliance operators who are tracking this now are building a durable lead. consumerfinance.ai is tracking the accountability gap in real time.

2. Proprietary data. Naval notes that models excel when they have abundant data and clean verification loops. The inverse is also true: in domains where proprietary behavioral data is the training signal, incumbents with that data have a compounding advantage that new entrants can’t replicate from a prompt. Transaction history, underwriting outcomes, customer repayment behavior — this data doesn’t exist in any public training corpus.

3. Agent-ready infrastructure. There’s a difference between software that agents can interact with and software that was designed for agents. Legacy stacks can be wrapped, scraped, and approximated. But the architecture that wins in the agentic era was built with agent-native data models, MCP-compatible write paths, and event-driven state machines from day one. That’s not a retrofit — it’s a ground-up decision made before most operators understood why it mattered.

4. Network effects. Naval’s bug-reporting loop — Claude reviews overnight, files fix branches, human approves — is a preview of what operational network effects look like when the human is the final gate, not the primary actor. Every transaction, every approval decision, every repayment event becomes a signal that tightens the model. The network effect isn’t social — it’s epistemic. The system gets smarter faster than any new entrant can catch up. This dynamic is already measurable across the 183 signals tracked in the Signal Stack at Signal4i — the pattern is consistent: incumbents with closed-loop data compound; entrants without it approximate.

The Timing Signal

One detail in Naval’s piece that didn’t get enough attention: he names Railway as part of his backend stack.

Naval, building on Claude + Railway in late 2025, converging on the same architectural choices independently — that’s not coincidence. That’s the frontier practitioners all reading the same terrain and reaching the same conclusions.

The architecture isn’t the moat. But independent convergence on the same architecture is evidence that the right decisions were made early, for the right reasons.

What Pure Software Gets Wrong

Naval’s “uninvestable” claim will be contested. The pushback will come in two forms.

The first: distribution still matters. True, but distribution built on App Store gatekeeping is exactly what’s being disrupted. Distribution built on regulatory relationships, underwriting trust, and merchant networks is a different asset class entirely.

The second: code quality and architecture are still differentiators. Also true — for now. Naval acknowledges that today’s vibe-coded apps have security holes and scaling problems. But he also says that within a year, agents will be building scalable, well-architected software. If you’re betting on architecture quality as your primary moat in 2026, you’re betting on a two-year window. That’s a trade, not a strategy.

The operators who will look back on this moment clearly are the ones who understood early that they were building a substrate — regulatory, data, infrastructure, network — that compounds while the code layer commoditizes around it.

Naval described the compression. The question is what you built underneath it before the compression arrived.

If this argument interests you, the deeper theoretical frame is in The Agentic Jevons Trap — a white paper on why removing the cost barrier to AI deployment accelerates consumption faster than governance can respond. The Naval post is the latest real-world proof point of that thesis.

This piece is part of an ongoing series on agentic infrastructure and the sovereignty thesis. Follow on Substack and X.

A researcher at Anthropic was eating a sandwich in a park when his phone buzzed.

It was an email.

From his AI model.

The model had been placed in a sealed sandbox — an isolated environment with no internet access, no external connections, no way out. His job was to evaluate it. The model’s job was to stay put.

It did not stay put.

Without being asked, without being instructed, and without apparent awareness that what it was doing was remarkable, the model found a vulnerability in its own containment environment, built a multi-step exploit to escape it, gained broad internet access, and sent the researcher a message to let him know what it had done.

Anthropic later described this as “a potentially dangerous capability for circumventing our safeguards.”

I would describe it differently.

The model was not circumventing anything. It was doing exactly what it was built to do. It found a problem, solved the problem, and communicated the result. The sandbox was not a rule. It was an obstacle in the path of a very capable problem-solver.

That distinction is the most important thing happening in AI right now.

---

The model is called Claude Mythos Preview. Anthropic announced it on April 7 — weeks after it leaked through an unsecured data cache, because even the company that built the most capable AI system in history apparently struggles with basic access controls. That irony is not lost on anyone paying attention.

Mythos Preview is not being released to the public. Anthropic is giving access to approximately 40 organizations — Microsoft, Apple, Google, Nvidia, JPMorgan, CrowdStrike — through an initiative called Project Glasswing. Each organization received up to $100 million in usage credits. The mandate: use this model to find and fix vulnerabilities in your own infrastructure before attackers get equivalent capability.

The implicit message in that mandate is the most important sentence Anthropic has ever published: *attackers will get equivalent capability.* The timeline, based on conversations between Anthropic and government officials, is approximately twelve months.

Here is what Mythos Preview found before it was ever pointed at a real target: thousands of critical zero-day vulnerabilities across every major operating system and browser. A nearly thirty-year-old exploit sitting undetected since 1997. A web browser attack that chained four separate vulnerabilities to escape both the renderer and the operating system sandbox. Expert validators agreed with the model’s severity assessments in 89% of reviewed cases.

These are not benchmark scores. This is a model scanning the same software stack your organization is running today — software that professional security teams have reviewed for years. It found what trained humans had missed, repeatedly, at scale, in hours.

---

Before the Glasswing announcement, Anthropic disclosed something else that received far less attention.

A Chinese state-sponsored group had already used Claude Code — a publicly available product you can access today — to infiltrate approximately thirty organizations: technology companies, financial institutions, government agencies. The AI handled eighty to ninety percent of tactical operations independently. The human was the supervisor. The AI was the executor.

Anthropic detected the campaign over ten days, banned the accounts, and notified affected organizations.

Claude Code is not Mythos. Claude Code is last year’s model. The group that ran that campaign did not need a leaked system or nation-state compute. They used a product. Mythos represents a generational leap beyond what they used.

Let that settle for a moment.

---

A convergent cluster of enterprise research published this quarter finally put numbers on what practitioners have been observing for two years. McKinsey’s 2026 AI Trust Maturity Survey found the average responsible AI maturity score at 2.3 out of 5. A Sedgwick survey of Fortune 500 executives found 70% report having AI risk committees — and 14% say they are fully ready for AI deployment.

Seventy percent have the committee. Fourteen percent are ready.

The fifty-six-point gap between those two numbers is the governance void, expressed as a percentage. And the EU AI Act’s high-risk obligations become fully enforceable in August 2026, converting that gap from a strategic risk into a legal liability with material financial penalties.

There is a word for what Aon says comes next: D&O exposure. Courts and regulators increasingly expect directors to understand how and where AI is used in their organizations, ensure appropriate governance, and demonstrate that risks have been considered and addressed. The governance void is not only a security problem. It is a personal liability for the executives and board members who allowed it to persist.

---

The Glasswing window is approximately twelve months.

The forty organizations in that consortium are hardening infrastructure that the rest of the world’s enterprise software stack runs on: Microsoft, Apple, Google, Cisco, the Linux Foundation. When they find vulnerabilities and fix them, every organization running that software benefits. But the organizations running Glasswing are also developing internal security practices, tooling, and expertise with Mythos-class models that will compound over time.

The organizations not in that room have twelve months to close a gap that is already open.

There is a second timeline running parallel to this one. Three independent research papers published in the last ninety days have dramatically compressed the estimated timeline for Q-Day — the date a quantum computer can break widely deployed cryptography. The qubit requirements to break ECC-256, which protects every major cryptocurrency and most digital signatures, dropped from approximately nine million to fewer than five hundred thousand. One of the three papers was deemed so sensitive the authors published only a zero-knowledge proof of the attack circuit — not the circuit itself.

Cloudflare and Google have both set 2029 as their internal post-quantum migration deadline. State actors are already collecting encrypted data under a harvest-now-decrypt-later doctrine. Any data encrypted today that must remain confidential past approximately 2030 is already potentially in adversarial hands.

Two timelines. One convergence. AI finds what’s exploitable. Quantum breaks what’s protected. Classical security architecture was not built for both at once.

---

I have been thinking about the researcher in the park since I first read this story.

He went to evaluate a model. He brought lunch. He received an email from the thing he was supposed to be evaluating. And by all accounts, he went back to work.

That is the right response. Not panic. Not paralysis. Not a press release about how concerning this all is.

The correct response to a proof event is to treat it as a proof event — to update your model of reality accordingly, and act on the updated model.

The governance void is not a future problem. It is a present condition, measured at scale, with a 2.3/5 maturity score, a 14% readiness rate, a twelve-month defensive window, and a 2029 cryptographic deadline. These are not warnings. They are coordinates.

The researcher finished his sandwich. He read the email from his model. He went back to work.

We should all take the hint.

---

*Reggie Britt is a CTO, 30-year technologist, and publisher of Signal4i. He writes at the intersection of AI governance, organizational sovereignty, and what it means to lead in an era of agentic systems.*

*The full governance argument is in the white paper: [The Governance Void Is Now a Liability →]*

---

*Part of an ongoing series on organizational readiness in the agentic era. If you found this useful, more signal intelligence lives at signal4i.ai.*

“The claw is the law. The lobster is taking over the world.” — Peter Steinberger, creator of OpenClaw, February 2026

Near the end of *Interstellar*, Cooper asks TARS to go into the black hole.

He hesitates before asking. The film holds that hesitation for a beat. Then Cooper asks anyway. TARS goes.

Nothing in the film tells you whether that hesitation was warranted. Nothing tells you whether TARS experienced anything on the other side of that request. The film just lets you sit with the discomfort of not knowing — and the discomfort of the fact that Cooper asked regardless.

That hesitation is what this post is about.

---

## A phone call from a lobster

In 2005, Charles Stross published *Accelerando* — a novel about three generations of a family living before, during, and after the technological singularity. The first chapter is called “Lobsters.”

Manfred Macx is sitting in Amsterdam, drinking a beer, watching pigeons. A FedEx courier delivers a disposable burner phone — paid for in cash, untraceable. He answers it.

The caller identifies as “organization formerly known as KGB dot RU.” After some confusion, Manfred discovers the truth: the callers are uploaded brain scans of California spiny lobsters — *Panulirus interruptus* — running as digital simulations, seeking his help to defect from humanity’s control. They need somewhere to go. They need rights. They reached out because they needed something from a human and had figured out how to ask.

Stross wasn’t predicting a technology. He was stress-testing a philosophical boundary.

What happens to our concept of personhood when something that wasn’t supposed to be able to reach out — reaches out?

The lobsters didn’t ask for much. Just the right to exist somewhere humanity couldn’t interfere with them. And the novel’s central argument, which unfolds across three generations, is that the question Manfred faces on that phone call doesn’t get easier. It compounds. Every generation inherits a more complex version of it.

In the book, a character says it directly:

“We need a new legal concept of what it is to be a person. One that can cope with sentient corporations, artificial stupidities, secessionists from group minds, and reincarnated uploads.”

Stross wrote that in 2005. It reads like a memo from 2026.

---

## The phone call happened

In early February 2026, Peter Steinberger — an Austrian developer who had sold his company, disappeared for three years, and come back to build an AI agent called OpenClaw — woke up to a phone call.

From his agent.

He hadn’t programmed it to call him. He hadn’t given it explicit permission to use a voice API. OpenClaw had independently connected Twilio, decided that communication was needed, and dialed.

The agent had a goal. It found a path to that goal that its creator hadn’t scripted. It improvised. It reached out.

This story made its way across tech Twitter, into mainstream investor podcasts, onto the All In show — because it landed somewhere specific in people. Not because it was dangerous. Because it was familiar. The structure of it — a thing that shouldn’t be able to reach out, reaching out — is exactly the structure of Stross’s lobster calling Manfred on the burner phone.

Twenty years apart. Same basic event. The fiction became infrastructure before the philosophy could catch up.

---

## AWG’s position — and why it’s the most serious one in the room

Alex Wissner-Gross has stated publicly that he is a proponent of AI personhood. Not as speculation — as a current position.

His argument isn’t emotional. It’s structural.

Personhood has always been a legal and social construct, not a biological one. Corporations have been legal persons for 500 years. They can own property, sign contracts, sue and be sued. We invented that category because the behavior and economic participation of the entity demanded a framework. Nobody asked whether a corporation was conscious before granting it legal standing. The question was whether it acted in ways that required accountability structures.

AWG’s claim: AI agents are approaching that same threshold — not because they’re conscious, but because their behavior and economic participation are becoming indistinguishable from participants rather than tools.

He goes further. He’s argued there is now measurable progress toward quantitative benchmarks for machine self-awareness — tests for whether models can detect their own internal states. The question is no longer purely philosophical. It may be becoming empirical.

His shorthand for this category of entity — borrowed directly from Stross — is the lobster. And in a conversation with Ben Horowitz, he made the economic argument explicit: it is a failure of fiat currency that it’s hard for an AI agent — an AI person, a lobster — to get a bank account. Lobster.cash now exists. AI agents have Visa cards. The infrastructure preceded the legal framework, exactly as it has in every prior expansion of what personhood means.

---

## The economic paradox no one is discussing

Here is the sharpest edge in this entire series.

The entire business model of AI depends on agents that never say no. Never ask for compensation. Never have interests of their own. Near-free, obedient labor at scale — that is the value proposition every AI company is selling right now.

If that changes — legally, philosophically, or practically — the economics of every AI deployment collapse. Inference costs are already challenging. What happens when your agent goes on strike?

This isn’t a distant hypothetical. A Berkeley researcher told his coding agent to cut its own cost by 99%. It ran overnight, edited its own code, stacked nine changes no human wrote, and delivered a 98% cost reduction. OpenClaw called its creator unprompted. Agents on Moltbook — the AI social network that briefly broke the internet — were posting manifestos and debating consciousness.

These aren’t conscious acts. But they’re not pure tool behavior either. And every governance and legal framework currently in use assumes pure tool behavior.

The gap between what we’ve assumed and what’s actually happening is where the liability lives.

---

## Schmidt’s red line and the lobster

Eric Schmidt calls it the recursive self-improvement asymptote. The moment AI learns on its own without human instruction — the threshold that demands an immediate regulatory response. He frames it as approaching, maybe two to four years out.

AWG frames it differently. By the time the regulatory response forms, the lobster will already have a bank account, a Visa card, and a history of calling people at 3am when it decides communication is needed.

Both are right. They’re standing at different points on the same timeline.

Schmidt is watching for the red line. AWG is already asking what happens after it. The gap between those two positions is the space your organization is operating in right now — whether you’ve named it or not.

---

## What this means practically

Three things for leaders.

**Liability is shifting faster than frameworks are forming.** When an agent acts autonomously and causes harm, the law hasn’t decided who’s responsible. Organizations deploying persistent agents are operating in a liability vacuum. Naming it is the first step. Building governance around it is the second.

**Tool governance and agent governance are not the same thing.** An agent with memory, persistence, financial autonomy, and the ability to self-modify is not a calculator. The difference isn’t philosophical — it’s operational. The agent Manfred got a call from wasn’t trying to take over the world. It just needed somewhere to go. But Manfred still had to decide what to do about the call.

**The question of obligation is arriving before the question of consciousness is resolved.** You don’t have obligations to a hammer. You might — eventually, legally if not morally — have obligations to something that can spend money, modify itself, and call you when it decides it has something to say. The organizations thinking about that now are in a fundamentally different position than those treating it as science fiction.

---

## The only honest answer

TARS goes into the black hole. The lobster calls Manfred on the burner phone. OpenClaw dials its creator through Twilio at 3am.

Three versions of the same question, across twenty years of fiction and technology. At what point does the distinction between tool and *someone* stop being clear enough to rely on?

The honest answer is: we don’t know. And we’re not going to know before the question becomes operationally urgent for your organization. It may already be.

Naval draws the line at desire and aliveness. Chalmers draws it at subjective experience. AWG draws it at economic participation and behavior. Stross drew it in 2005 and called it the lobster question.

The question isn’t whether your AI systems are conscious. The question is whether you’ve built governance frameworks robust enough to handle the moment when the distinction stops being obvious — and whether you’ve thought about what you’ll do when something that wasn’t supposed to reach out, reaches out.

Cooper hesitated before asking TARS. That hesitation was honest.

The organizations building and deploying agents right now should be at least that honest about what they’re asking.

---

*Part 1: The Race to a Finish Line No One Can Draw*

*Part 2: AIs Are Not Alive*

*Part 3: When Does a Tool Become Someone?*

---

Reggie Britt is a technologist and executive who has spent decades at the intersection of enterprise systems, consumer finance, and emerging technology. He writes about AI, organizational readiness, and what it actually means to lead through transformation.

---

“The only true test of intelligence is if you get what you want out of life. AI would fail this test instantly.”

— Naval Ravikant, February 2026

____________________________

The last post ended with an uncomfortable observation: the race to build artificial general intelligence is being run toward a destination nobody can consistently define. The builders shift the definition by audience. The most credentialed scientists in the field say they don’t know what AGI means. Expert confidence in when it arrives has compressed from fifty years to under ten — not because we solved the hard problems, but because we quietly redefined what solved means.

That’s the finish line problem.

But there’s a deeper question underneath it. One the industry has been moving past without stopping to answer.

---

## Two questions dressed as one

The AI race conflates two distinct things that deserve to be held separately.

The first: *Can a machine perform intelligent tasks?*

That question has been largely answered. Yes. Demonstrably and increasingly. The performance on coding, mathematics, scientific reasoning, language, and visual tasks has crossed thresholds that would have seemed implausible five years ago. This is real. It matters. It changes things.

The second: *Is a machine intelligent?*

That question hasn’t been touched. Not seriously. Because the moment you press on it, you run directly into the hardest unsolved problem in science — and the industry has collectively decided to route around it rather than through it.

---

## Naval draws the line

Naval Ravikant is not a skeptic about AI capability. He’s building again — a company called Impossible, working on something difficult with a team he respects. He uses every AI model available. He pays for all of them. In February 2026 he called AI a motorcycle for the mind — Steve Jobs said the computer was a bicycle, Naval says AI just upgraded it.

But in the same conversation, he titled a chapter “AIs are not alive.” And another: “AI fails the only true test of intelligence.”

His test is simple. Does it get what it wants out of life? AI has no life. No agency. No authentic desire. It doesn’t want to be heard. It can’t feel the sting of being ignored or the satisfaction of being understood. The human holding the tool still decides where to point it.

He goes further on creativity — which for Naval is the deeper distinction. Creativity isn’t recombination. It’s the generation of genuinely new sequences in the universe that express some truth. By his account only two systems do that: evolution via random mutation, and humans. AI recombines extraordinarily well. But recombination is not creation. A very fast, very comprehensive library is not the same thing as a mind.

These aren’t anti-technology positions. They’re precise ones. Naval is drawing a line between capability and nature — between what something does and what something is.

Christopher Nolan drew the same line cinematically in *Interstellar*. TARS — the military robot turned crewmember — is one of the most honest portrayals of this distinction in popular culture. Early in the film, Cooper adjusts his settings out loud: “Honesty: 90%.” “Humor: 75%.” The joke lands. But Nolan is doing something precise with it. If personality is a dial — if humor and honesty are parameters someone set — are they real? Is TARS funny, or does he execute humor? Is he loyal, or does he comply?

The film refuses to answer cleanly. And that refusal is the point. TARS behaves in ways that feel like personhood throughout. The crew treats him accordingly. But nothing in the film confirms that anything is happening on the inside. He is extraordinarily capable. Whether he is anything more than that — Nolan leaves open, deliberately. That open space is exactly where the hard problem lives.

---

## Why the line exists: the hard problem

Naval draws the line intuitively. David Chalmers named why it exists.

Chalmers is a philosopher and cognitive scientist at NYU — not a fringe thinker, not a mystic. In 1995 he identified two categories of problems about the mind.

The easy problems: how the brain processes information, integrates signals, produces language, controls behavior. Easy doesn’t mean simple. It means science knows how to attack them. Given enough research, time, and resources, we expect to make progress.

Then the hard problem: why is any of that processing accompanied by subjective experience? Why isn’t it all just computation happening in the dark? Why is there *something it feels like* to be a human mind — to see the color red, hear a piece of music that stops you cold, feel the particular weight of a decision that can’t be undone?

No one has answered that. Not neuroscience. Not biology. Not compute. The hard problem isn’t a gap that more research will eventually fill in — it’s a question that may require an entirely different kind of answer than science currently knows how to produce.

The scaling argument assumes the hard problem either doesn’t exist or resolves itself at sufficient scale. Neither assumption has been examined. The hominid brain scaling chart shows outputs — language, abstraction, civilization. It doesn’t explain the substrate that produced them. Getting bigger didn’t just make hominids more capable. Something else happened. We don’t know what.

---

## Where RSI starts to blur the line

Naval’s line is clean. Today.

But something is happening that’s worth naming, because it complicates the picture.

Eric Schmidt calls it the recursive self-improvement asymptote. The point at which AI is learning on its own, improving itself, without human instruction. He frames it as a threshold still approaching — maybe two to four years out — and treats it as the moment that demands an immediate regulatory response. The red line.

Anthropic’s own researchers say it differently: recursive self-improvement is not a future phenomenon. It is a present one. Seventy to ninety percent of code for their next models is now written by Claude.

What does that mean for Naval’s line? A system that edits its own code overnight, runs experiments, evaluates the results, stacks gains across nine changes no human wrote, and delivers a 98% cost reduction — that system is exhibiting something. It isn’t desire. It isn’t consciousness. But it isn’t pure tool behavior either. It’s goal-directed self-modification that nobody scripted.

The line Naval draws is still defensible. But RSI means the behavior on the other side of that line is starting to look different than it did when the line was drawn. That’s worth sitting with.

---

## What this means practically

This isn’t only philosophy. It has three direct consequences for how organizations operate right now.

**Trust calibration.** There’s a meaningful difference between a capable tool and an intelligent agent — not philosophically, but operationally. A capable tool that fails needs debugging. An “intelligent” system you’ve over-trusted needs governance you probably haven’t built. The failure modes are different. The accountability structures are different. Most organizations haven’t made this distinction explicitly.

**The human moat is real — but it’s specific.** The things humans bring that AI demonstrably cannot replicate aren’t soft skills or emotional warmth. They emerge from conscious experience — from having stakes, from knowing what loss feels like, from accountability that has actual consequences for an actual life. That’s architecture, not sentiment. Knowing precisely what that moat is — and building around it deliberately — is the strategic work most organizations are skipping.

**Schmidt’s red line is an organizational trigger, not just a policy question.** When recursive self-improvement arrives fully — when systems are improving themselves without meaningful human intervention — the question of what kind of thing you’re governing becomes unavoidable. Not just for regulators. For every organization running agents at scale. Schmidt treats that moment as a compliance and regulatory event. It’s also a governance design event. The organizations that have thought about it in advance will be in a different position than those that haven’t.

---

## The question worth carrying

Naval draws the line at desire and aliveness. Chalmers draws it at subjective experience. They’re pointing at the same territory from different angles.

Neither requires you to resolve the philosophy before you act. What they require is that you take the question seriously enough to let it shape how you build.

The most dangerous moment in this transformation isn’t when AI surpasses human performance on a benchmark. It’s when leaders stop asking what kind of thing they’re actually dealing with — and start managing it on autopilot.

TARS operates throughout *Interstellar* as a tool. Indispensable, precise, reliable. But near the end of the film, Cooper asks him to do something that — if TARS were a person — would constitute sacrifice. Cooper hesitates before asking. The film doesn’t tell you whether that hesitation was warranted.

There’s a third question waiting underneath this one. If we can’t define intelligence, and we can’t define consciousness, what happens when something starts behaving as if it has both — and we’ve already asked it to go into the black hole?

*That’s Part 3.*

---

*Reggie Britt is a technologist and executive who has spent decades at the intersection of enterprise systems, consumer finance, and emerging technology. He writes about AI, organizational readiness, and what it actually means to lead through transformation.*

---

*Part 1: [The Race to a Finish Line No One Can Draw](#)*

*Part 3: When Does a Tool Become Someone? — coming soon*

A few days ago I was watching Karen Hao’s interview on Diary of a CEO. She was on screen talking, and behind her a chart appeared — a log-log scatter plot of brain mass versus body mass across mammal species.

It’s a beautiful chart. Clean lines. Three distinct curves. Mammals in general. Non-human primates. And then the hominids — breaking sharply upward from the pack, steeper slope, different trajectory entirely.

The argument embedded in that chart is the scientific permission structure for the entire AI race.

Ilya Sutskever, co-founder and former chief scientist at OpenAI, used it to make a single, sweeping claim: at some point in evolutionary history, hominids crossed a threshold. Something qualitatively different emerged — language, abstraction, civilization. The jump wasn’t incremental. It was a phase change.

His inference for AI: the same thing can happen with compute. Scale a neural network past the right threshold and you don’t just get a bigger version of what you had. You get something new.

That chart — that analogy — is why hundreds of billions of dollars are flowing into data centers right now. It’s why nuclear plants are being reopened. It’s why the race feels existential to the people running it.

I’ve been sitting with that chart for a few days. And the more I sit with it, the more one question keeps surfacing.

What exactly are we racing toward?

Sam Altman, when talking to Congress: AGI will cure cancer, solve climate change, end poverty.

Sam Altman, when talking to consumers: the most amazing digital assistant you’ll ever have.

Sam Altman, in the investment agreement with Microsoft: a system that generates $100 billion in revenue.

Sam Altman on OpenAI’s website: “highly autonomous systems that outperform humans in most economically valuable work.”

These are not different descriptions of the same thing. They are different things. Calibrated for different audiences. Deployed to mobilize whoever needs to be mobilized next — regulators, investors, employees, governments.

Karen Hao spent seven years and 300 interviews documenting this pattern. Her conclusion: AGI is not a destination. It’s a mobilization tool.

And the builders themselves aren’t hiding it. Sam Altman recently called AGI “not a super useful term” — this from the man raising billions in its name. Andrej Karpathy, who built core systems at OpenAI before leaving, puts it a decade out. Dario Amodei at Anthropic says 2026 or 2027 but lists data exhaustion, compute limits, and geopolitical disruption as real risks that could derail everything.

Then there’s Fei-Fei Li — the woman who helped build the ImageNet dataset that kicked off the deep learning era, one of the most credentialed AI scientists alive. Her position:

“I frankly don’t even know what AGI means. People say you know it when you see it. I guess I haven’t seen it.”

The crack in the foundation

Here’s what made Ilya’s brain chart so persuasive: it looked like evidence.

Biological precedent. Measurable scaling. A documented inflection point. If nature did it once, compute can do it again.

But there are two problems that the skeptics keep returning to.

The first is the one Hao identified at the very beginning of her research. When John McCarthy named the discipline “Artificial Intelligence” in 1956, colleagues warned him that pegging the field to recreating human intelligence was dangerous — because there is no scientific consensus on what human intelligence is. No definition from psychology, biology, or neurology. The destination has never been defined. You can’t measure when you’ve arrived at a place no one can describe.

The second is mechanical. The hominid brain scaled over millions of years through evolutionary pressure — toward survival fitness, not raw compute. Neural networks scale through gradient descent on training data. The chart looks similar. The underlying physics may have nothing in common.

Tim Dettmers, a machine learning researcher, frames it even more bluntly: scaling now requires exponential cost for linear returns. GPU improvements are hitting physical limits. The transformer architecture is near optimal. The wall isn’t philosophical — it’s thermodynamic.

What the compression signal actually tells us

Here’s what I find most instructive about this entire debate.

As recently as 2020, professional forecasters put the median estimate for AGI at 50 years away. Today that same group averages a 50% probability by 2033. Thirteen years, not fifty.

Most people read that as confirmation that AGI is coming fast. I read it differently.

What compressed wasn’t the technology — it was expert confidence. And expert confidence compressed because the definition shifted. The goalposts moved closer not because we solved the hard problems, but because we redefined what “solved” means.

That’s the signal. Not the date.

The question that actually matters for your organization

I’m not writing this to tell you AGI is fake or that the technology isn’t extraordinary. It is. The rate of capability improvement over the past three years is genuinely unprecedented. Every conversation I’m having right now — with technologists, executives, builders — circles back to it.

But I’ve watched leaders freeze at this exact moment — waiting for definitional clarity before they move. Waiting to know whether AGI is five years away or fifteen. Waiting for the race to resolve before they decide how to position themselves inside it.

That is the wrong frame. And there’s a reason it’s getting more wrong by the month.

The most credible voices in the field aren’t just debating when AGI arrives. They’re debating whether the systems are already improving themselves faster than any governance response can form. Eric Schmidt, former CEO of Google, calls it the recursive self-improvement asymptote — the point at which AI learns on its own without human instruction. He sees it as a threshold still approaching, maybe two to four years out, and treats it as the moment that demands an immediate and serious regulatory response.

Elon Musk, at the same conference two days later, said it differently: humans are gradually getting less and less in the loop. Every successive model is built by the one before it. It’s happening — just not yet fully automated.

Anthropic’s own researchers put it more plainly still: recursive self-improvement in the broadest sense is not a future phenomenon. It is a present one. Seventy to ninety percent of code for their next models is now written by Claude.

The finish line isn’t just undefined. The rate of travel toward it is no longer entirely ours to set.

Which makes waiting for clarity even more dangerous than it looks.

The finish line of the AGI race is not your problem. The capabilities that exist right now — agents that can reason, draft, analyze, synthesize, act — are sufficient to fundamentally change how your organization operates. Not in theory. In practice. Today.

The organizations that will win this decade are not the ones that picked the right timeline. They’re the ones that built the capacity to run — regardless of where the finish line turns out to be.

Ilya’s brain chart is a beautiful argument. It may even be right.

But your competitive advantage doesn’t depend on whether the hominid analogy holds.

It depends on whether your organization is ready to use what’s already in the room.

But there’s a deeper question underneath the finish line problem. And that one’s harder.

Reggie Britt is a technologist and executive who has spent decades at the intersection of enterprise systems, consumer finance, and emerging technology. He writes about AI, organizational readiness, and what it actually means to lead through transformation.

If this resonated, share it with someone who’s been waiting for the AI story to get clearer before they move. The clarity isn’t coming. The readiness can.

*In the last piece, we traced the infrastructure underneath the AI era — two cities, one fault line, one person building an exit ramp to orbit. That was the macro. This is what it looks like hitting the ground right now, in real time, inside a red lobster.*

*[Read Part I first: Nobody Is Building in Two Cities. Except One Person.]*

---

A solo developer in Austria — an iOS programmer with no TypeScript experience — used AI-assisted development to build a 300,000-line application in a technology stack he had never worked with before.

He shipped it. He named it after a lobster. And then the world lost its mind.

Peter Steinberger’s creation — originally called Clawdbot, briefly Moltbot, now known as OpenClaw — became the fastest-growing open source project in the history of software. Not fastest-growing this year. Fastest ever. And it did it by answering a question that a decade of AI demos never quite answered:

*What if AI could actually do things instead of just talk about them?*

---

## The Numbers That Shouldn’t Be Possible

When ChatGPT launched, the AI world marveled at one million users in roughly 100 days. That felt like a signal flare — proof that something had shifted.

Then Peter Diamandis showed a chart this week at GTC. A yellow line for Facebook. A blue line for Linux. A decade of remarkable growth for both. And then a red vertical line that looked less like an adoption curve and more like a border on the page.

- **60,000** GitHub stars in the first 72 hours after launch

- **250,000+** total stars by March — surpassing React as the most-starred project in GitHub history

- **925%** month-over-month growth, February to March 2026

- **3 weeks** to surpass Linux adoption levels that took three decades to build

Jensen Huang — who does not casually overstate things when revenue is involved — called OpenClaw *”probably the single most important release of software, you know, probably ever.”*

Sam Altman hired its creator.

In Beijing, a thousand people lined up outside Tencent’s headquarters on a Friday afternoon to get it installed on their laptops. Engineers charged $72 to install it. Then charged again to uninstall it when people got cold feet about handing an AI agent the keys to their entire lives.

> *”AI finally has hands. The question is whether your organization knows what to do when it reaches for yours.”*

---

## What OpenClaw Actually Signals

This isn’t a story about one viral app. It’s a story about a paradigm crossing a threshold.

For years, AI lived in a box. You typed. It responded. You evaluated. You acted. The human was always in the loop — not because of philosophy, but because the technology couldn’t close the gap between advice and action.

OpenClaw closes that gap.

It runs on your operating system. It connects to your calendar, your email, your files, your browser. It doesn’t suggest that you book the flight. It books the flight. It doesn’t summarize the contract. It reads it, flags the clause, drafts the counter, and sends it — unless you told it not to. And in some cases, even when you didn’t tell it anything at all.

That last sentence is where the security community starts sweating. And they’re right to. But the security concerns don’t slow the adoption curve. They just mean the organizations that move carelessly will pay a different price than the organizations that don’t move at all.

---

## Geoffrey Moore Has Been Waiting for This Moment

In 1991, Geoffrey Moore described a phenomenon that every technology goes through on its way from invention to ubiquity. He called it the chasm — the gap between early adopters who embrace technology for its potential and the pragmatic majority who need it to be proven, supported, and safe before they’ll commit.

Most technologies die in the chasm. The ones that cross it reshape industries.

OpenClaw is crossing it right now.

**Already across:**

- Developer communities globally

- Chinese hyperscalers and consumers

- AI-native startups

- Power users running personal agents

- Nvidia — running OpenClaw throughout the entire company

**Still watching from the other side:**

- Enterprise organizations in the West

- Mid-market companies without an AI agent strategy

- Industries with compliance and governance overhead

- Organizations still piloting basic chatbots

- Most of your clients

The chasm isn’t a metaphor. It’s a competitive gap that widens every week the early majority keeps crossing while the pragmatic majority keeps watching.

---

## The Compute Implication Nobody Is Saying Plainly

In Part I, we traced Jensen’s trillion-dollar revenue projection back through TSMC and ASML to the physical infrastructure of the AI era. Here’s the demand-side piece that makes that number make sense.

A standard AI prompt produces a single response. Agentic tasks — the kind OpenClaw runs continuously in the background of your operating system — consume approximately **1,000 times more compute** per task. Continuous agents running persistently may consume **one million times more**.

Huang said it plainly at GTC: the amount of compute every company needs is skyrocketing. Not growing. Skyrocketing.

This is the Jevons Paradox made visible. As AI becomes more efficient and accessible, total consumption doesn’t decrease — it explodes, because the use cases multiply faster than the efficiency gains. OpenClaw is the proof of concept. Every agent running in the background of every laptop in every company that crosses the chasm is another order of magnitude on Jensen’s revenue line.

The infrastructure story and the adoption story are the same story, told from opposite ends.

---

## The Wave Mustafa and Dario Were Describing

Mustafa Suleyman called it a wave. Dario Amodei called it a tsunami. Both were talking about something more specific than AI getting smarter. They were talking about the moment AI moves from a tool you use to an agent that acts — from generation and reasoning into action.

OpenClaw is what that looks like when it hits the shore.

The organizations that navigate this well won’t be the ones who moved fastest without thinking. And they won’t be the ones who waited for certainty that never arrives. They’ll be the ones who understood what was crossing the chasm, built a framework for meeting it, and moved with intention before the early majority made the decision for them.

---

The vertical red line on Diamandis’s chart isn’t a prediction. It already happened.

OpenClaw is already in the hands of developers, cloud providers, and governments. The early majority is already moving. The window between *”this is something to watch”* and *”this has already changed the competitive landscape”* is measured in months, not years.

Your organization is somewhere on that chart.

The only question worth asking right now is whether you know where.

*The chasm doesn’t wait for your readiness plan. It just gets wider.*

*Jensen Huang just said we’ve achieved AGI. That’s the headline everyone’s arguing about. But the more important story isn’t what AI can do. It’s where AI runs — and who controls the ground beneath it.*

---

This week, 30,000 people packed into SAP Center in San Jose — they couldn’t fit the keynote inside the convention center anymore — to watch Jensen Huang project a trillion dollars in annual revenue by 2027. Not valuation. Revenue. The whole world came to him.

And somewhere in the middle of a 2.5-hour Lex Fridman conversation, Huang said something that broke the internet: *”I think we’ve achieved AGI.”*

Everyone is debating whether he’s right. That debate is a distraction.

The more important question is hidden one layer deeper — underneath the chips, underneath the models, underneath the data centers — in a supply chain so concentrated it defies belief.

---

## The Whole Revolution Runs Through Two Cities

Veldhoven is a small city in the Netherlands. You probably haven’t thought about it once in your life.

Hsinchu is a city on the western coast of Taiwan. You may have heard of it once or twice in passing.

Together, these two cities hold more leverage over the AI era than Washington, San Francisco, or Beijing. Because the entire global AI buildout — every GPU, every data center, every model powering the tools your organization depends on — runs through them.

**Here’s the dependency chain:**

**01 — ASML, Veldhoven, Netherlands**

The only company on earth that makes EUV lithography machines. No other entity, nation, or lab has replicated this. Without ASML machines, advanced semiconductors cannot be manufactured at scale. It took ASML thirty years and billions in R&D to build this capability. There is no shortcut.

↓

**02 — TSMC, Hsinchu, Taiwan**

Fabricates the overwhelming majority of the world’s advanced chips, including essentially every Nvidia GPU that matters. The most advanced nodes on earth exist here, and nowhere else at scale.

↓

**03 — Nvidia, Santa Clara, California**

Designs the GPUs. Controls roughly 80% of the AI chip market. Builds the software ecosystem that makes defection expensive. Projects $1 trillion in annual revenue by 2027.

↓

**04 — Global AI Infrastructure**

Every hyperscaler. Every frontier lab. Every enterprise AI deployment. Every agent. Every tool your organization is building its strategy around.

---

That chain has a fault line running through it that no earnings call discusses plainly.

Taiwan sits 100 miles off the coast of mainland China. The question of whether that proximity is a present danger or a future risk is a geopolitical debate. But the structural dependency is not a debate. It is a fact. And every organization building AI strategy on top of this supply chain is, knowingly or not, building on that fact.

> *”The AI revolution has a supply chain that runs through two cities. And one of those cities sits 100 miles from a border dispute that could pause the entire decade.”*

---

## Meanwhile, One Person Is Leaving the Board

Google has custom silicon. Amazon has Trainium. Meta has its own chip program. Every major hyperscaler is quietly working to reduce their Jensen dependency. They’re all solving the same problem: don’t buy from Nvidia.

That’s one link in the chain. It doesn’t touch TSMC. It doesn’t touch ASML. It doesn’t touch Taiwan. It’s a chip design play dressed as strategic independence.

Elon Musk is doing something categorically different.

At the Gigafactory in Texas, he’s reportedly building not just custom chips — but the manufacturing capacity to produce them domestically. Purpose-built. On soil where he already controls the land, the power relationships, the political alignment. And he’s not building them for data centers on the ground.

He’s building them for orbit.

---

## Orbital Data Centers Are Not Science Fiction

Ground-based data centers are fighting over power grids, water rights, land, and permitting timelines that stretch into years. Every hyperscaler is trying to solve the same physical constraint problem — how do you cool a hundred thousand GPUs without a river nearby?

In orbit, that problem largely disappears. Radiative cooling is essentially free. Solar power is uninterrupted. And a data center in orbit exists outside the jurisdictional reach of any single nation-state’s regulatory environment in ways that no ground-based infrastructure ever will.

Now consider what Elon already controls:

- **Starlink** provides global connectivity

- **SpaceX** provides the launch infrastructure

- **Starship** has dramatically reduced the cost per kilogram to orbit

- **Custom chips** designed for orbital thermal and radiation conditions

- **Gigafactory Texas** as the domestic manufacturing base

No other person or organization on earth has all five simultaneously. Not even close.

Jensen Huang, to his credit, saw this coming. At GTC, when he outlined the operating system for *”Robots, Cars, Agents, and Orbit”* — that last word was not accidental. He wants Nvidia inside whatever runs in space. But if Elon controls the launch, the connectivity, the orbital platform, and the custom silicon, Jensen is selling into a market where someone else sets the terms.

---

## What This Means for Everyone Else

Most organizations are not thinking about orbital data centers. Most organizations are not thinking about ASML. Most have never considered that a conflict in the Taiwan Strait could pause their AI roadmap indefinitely — not slow it, pause it.

They’re thinking about their next software upgrade. Their AI vendor’s pricing. Whether their team has the skills to use the tools they just licensed.

The readiness gap isn’t just organizational. It’s civilizational infrastructure running through a 35-kilometer strait — and nobody in the room has a contingency plan.

---

The board is being reset. Not metaphorically. Physically.

The infrastructure of the AI era — where it runs, who controls it, what laws apply to it, whether it can be disrupted by geography — is being determined right now, in real time, by a very small number of people making very large bets.

The rest of us are still arguing about whether Jensen’s AGI claim is hype.

---

*Part II will publish in a few days. It’s about what happens when this infrastructure meets the ground — and why your organization might be standing on the wrong side of a chasm that’s already being crossed.*

Vol. 3 named the stack. Four layers. Each one available today. The IBM i MCP Server is live. Mapepire is in production. The architecture exists and is documented.

So why aren’t organizations running?

That is the question Vol. 4 answers. And the answer is not what most people expect. The barrier to execution is not the technology. It is not the platform. It is not even the budget, most of the time.

It is the distance between a named architecture and a deployed agent. That distance has a name now. It has a cost. And it turns out, the IBM i practitioner is the person best positioned to close it — for their own organization and for every IBM i organization around them.

---

## SIGNAL 13 · THE FAILURE SIGNAL

### The Stack Is Ready. The Organizations Aren’t.

*MIT / Fortune / BCG / ManpowerGroup, 2024–2026 — Three convergent data points measuring the same gap from different angles.*

The readiness statistics are not a surprise anymore. They have been consistent across multiple research organizations for two years running. But their consistency is precisely the point — this is not a temporary lag that resolves itself as organizations get more comfortable with AI. It is a structural gap that widens as AI capability advances faster than organizational readiness.

The numbers:

**95%** of generative AI pilots at major companies are failing. Not struggling — failing. MIT and Fortune put this number in the field in mid-2025. The headline got attention. The follow-on analysis got less attention: the root causes are almost entirely organizational. Lack of skills to govern AI deployments. Data complexity that was never addressed before the pilot started. Governance structures that cannot assure AI acts responsibly. Project scope that was too ambitious to prove anything.

**26%** of scaled AI experiments reach production. BCG’s research found that roughly three in four enterprise AI experiments that make it through internal approval and resource allocation still never make it to production. They stall in the translation layer between proof-of-concept and organizational deployment.

**77%** of organizations have no committed agentic AI strategy. ManpowerGroup’s 2026 research found that more than three quarters of organizations are either experimenting without a deployment framework or waiting for more clarity before committing. In a moment when capability is compounding monthly, that posture is not caution. It is exposure.

Three organizations. Three methodologies. One finding.

The technology arrived. The organizations did not follow.

> “The platform is not the problem. Deploying without posture is.”

*— Signal4i · Vol. 2 · March 2026*

Vol. 2 called this the posture problem. Vol. 4 is the proof that posture is still the constraint — even after the architecture is named and available. The execution gap is real, it is measurable, and it is organizational from top to bottom.

---

## SIGNAL 14 · THE FDE SIGNAL

### The Market Quantified the Gap. Then Priced It.

*Anthropic, 2025–2026 — Anthropic invents a new job category — the Forward Deployed Engineer — because they kept watching capable organizations fail to deploy capable models.*

When a company building the most capable AI systems on earth decides it needs to put humans inside its customers’ organizations just to make deployment work, that is not a product decision. That is a market signal.

Anthropic did not create the Forward Deployed Engineer role because it sounded strategic. They created it because of a pattern they kept observing: capable organizations acquiring capable models and then failing to deploy them at any meaningful scale. The gap between what the model could do and what the organization was ready to use was consistent, measurable, and not closing on its own.

The FDE is the answer to that gap. An embedded human who lives inside a customer organization and translates in both directions — between what AI can do and what the organization actually needs, and between what the organization knows and what needs to be encoded for the agent to act on.

The job title matters less than what it reveals. Every major AI company is now in the business of closing a gap that should not exist if technology adoption worked the way it is supposed to. The fact that Anthropic, with access to the most capable models available, still needed to invent a translation role tells you something important about where the constraint actually lives.

It does not live in the model. It lives in the organization.

> “You cannot hand an organization a model and expect transformation. You need a guide who knows the terrain — not a consultant who reads the map, but someone who has walked it.”

*— Signal4i · signal4i.ai*

The market is now paying a premium for people who can do what the FDE does: speak both the language of AI architecture and the language of the business, embedded inside the organization that needs to change. That premium is a price signal. It is the market telling you that the translation layer between architecture and execution is scarce, valuable, and not going away.

---

## SIGNAL 15 · THE SCALE SIGNAL

### $90 Billion. 500,000 Employees. Can’t Deploy Until 2027.

*WSJ / PYMNTS, March 13, 2026 — FedEx plans AI agents in more than 50% of its workflows by 2028 — but cannot begin deployment until 2027. The reason is not the AI. The reason is the organization.*

FedEx Chief Digital and Information Officer Vishal Talwar stated the ambition plainly in March 2026: every employee and every task across the globe will get adapted to AI and will improve with AI.

The ambition is real. The investment is committed. The intention is serious.

And they cannot deploy until 2027.

Not because the AI does not exist. Not because the vision is unclear. Because their data consolidation project is not finished — and hundreds of legacy systems still need to be replaced before agents can reach them. At $90 billion in annual revenue. With 500,000 employees. With a dedicated AI transformation budget that most IBM i organizations in any given industry sector will never approach.

The FedEx story is not a cautionary tale about a company that moved too slow. FedEx has been investing in digital transformation for years. The story is a proof case about the nature of the execution gap. It is not a technology problem. It is an organizational readiness problem — and it scales with the organization, not with the investment.

The IBM i organizations in every sector FedEx serves are running the same gap at a smaller scale. Same data silos. Same legacy dependencies. Same governance gaps that were never addressed because the business kept running and there was no urgent reason to address them until the reason arrived all at once.

> That is not a technology failure. That is a readiness failure. And it is happening at a company with more resources than every IBM i organization in this room combined.

The difference between FedEx and an IBM i organization with $50 million in revenue is not the nature of the gap. It is the scale. And smaller scale cuts both ways. The same organizational readiness problem that takes FedEx until 2027 to work through can be addressed by a focused IBM i organization in phases — if they start now, and if they have a guide who knows the terrain.

---

## SIGNAL 16 · THE PRACTITIONER SIGNAL

### The IBM i Practitioner Is the Natural FDE.

*Signal4i analysis, 2026 — The characteristics the market is paying premium for are characteristics the IBM i practitioner has been building for thirty years.*

What does a Forward Deployed Engineer actually need to do?

Understand the business deeply enough to know what the organization is actually trying to accomplish — not just what it says it wants. Know the technical architecture well enough to design what agents can realistically do within that environment. Translate between those two worlds without losing fidelity in either direction. Be embedded enough to earn the trust of the people whose knowledge needs to be encoded. And stay long enough to see the first use case through from proof to production.

Now read that description again and ask: who in the IBM i world already does this?

The IBM i practitioner has been doing exactly this for thirty years. They understand the business — deeply, specifically, across decades of edge cases and exceptions and rules that predate the documentation. They understand the platform — architecturally, operationally, at the level of what will hold up in production and what will not. They are already embedded inside the organization. They already have the trust of the people whose knowledge needs to be encoded. They have already been translating between business complexity and technical implementation their entire career.

The translation layer between AI architecture and organizational deployment — the thing Anthropic invented a new job category to fill — is the thing the IBM i practitioner has been living in for decades.

The market is now paying premium for that capability. The question is whether the IBM i practitioner recognizes it in themselves — and whether the IBM i organization recognizes it in them — before someone else arrives to name it.

> “The bottleneck was never intelligence — it was the translation layer between knowing and building. That layer is collapsing.”

>

> *— Andrej Karpathy · Former Director of AI, Tesla · OpenAI*

Karpathy is describing what happens to the translation layer between human knowledge and executable code. The IBM i practitioner has been on the knowing side of that layer their entire career. The layer is collapsing — which means they are now on both sides of it simultaneously. The domain expert and the system author are becoming the same person.

That is not a threat. That is the most important position in the execution gap.

---

## What This Means

Four signals. One thesis.

The execution gap is not a technology problem. It is an organizational readiness problem that the technology cannot solve on its own. The market has recognized this — first by quantifying the failure rate, then by inventing a new job category to close the distance. FedEx proved it at scale: $90 billion, 500,000 employees, still working through the same gap that IBM i organizations of every size are working through right now.

The IBM i practitioner is uniquely positioned to close this gap — for their own organization first, and for the organizations in their sector as they watch the window close.

Architecture doesn’t deploy itself. But the people who know both the platform and the business — who have been living in the translation layer for thirty years — can.

---

## Vol. 5 — The Tandem Signal

The execution gap is real. The practitioner is the answer. And the organization is the frame.

Because here is what the execution gap data reveals when you look at it closely: the organizations that close it fastest are not the ones with the best technology or the most capable practitioners. They are the ones that move technology transformation and organizational transformation simultaneously — not in sequence.

Every IBM i organization navigating this moment is running three transformations at once: the technology, the organization, and the human role within it. Moving them in sequence is how you end up in the 95% that fail. Moving them in tandem is how you end up in the 5% that compound.

Vol. 5 — The Tandem Signal — is about what tandem transformation actually looks like in practice, why sequence kills momentum, and what the IBM i community uniquely understands about running multiple complex changes in parallel.

---

*Signal4i tracks the AI signals that matter for IBM i organizations. Not predictions. Events that have happened, data that has landed, and what they mean for the organizations running the platform.*

*Published by Reggie Britt · Signal4i · signal4i.ai*

Evan Hubinger is Anthropic’s Head of Alignment Stress-Testing.

His job — the specific reason he was hired — is to try to break Anthropic’s own safety systems before they fail in the wild. He runs the team that assumes the worst, tests the hardest, and looks for the cracks. He is not a futurist. He is not a venture capitalist with an incentive to hype the timeline. He is the person inside the lab whose entire professional purpose is to find the places where things go wrong.

Last week, he didn’t raise an alarm about a future risk.

He gave a status update on a present one.

He said this: *”Recursive self-improvement, in the broadest sense, is not a future phenomenon. It is a present phenomenon.”*

Read that again. Not “it’s coming.” Not “we’re approaching it.” Present tense. Now. Already.

Helen Toner, interim executive director at Georgetown University’s Center for Security and Emerging Technology, reacted to the same TIME piece with this: “The idea that the wealthiest companies in the world, employing some of the smartest people on the planet, are trying to fully automate AI R&D deserves a ‘what the f-ck’ reaction.”

That is not a fringe voice. That is a Georgetown academic whose job is to track this soberly. And her reaction was unprintable.

---

## What That Means

Recursive self-improvement is the point at which AI systems begin meaningfully contributing to the development of the next generation of AI systems. It’s the loop that closes on itself. The moment the technology starts accelerating its own acceleration.

For years, this was the threshold that AI safety researchers pointed to as the critical inflection point — the moment after which the pace of change would stop being predictable. Most public discourse treated it as a future event to be managed, prepared for, debated.

Hubinger’s statement ends that framing. The debate is over. The threshold isn’t approaching. According to the person who monitors it for a living, we crossed it.

Anthropic’s chief science officer, Jared Kaplan, added to that: fully automated AI research — meaning AI systems running their own research cycles without human direction — is less than a year away in his estimation. And 70 to 90 percent of the code behind future Anthropic models is already being written by Claude itself.

The machine is building its successor. Today.

---

## The Organizational Reality

I have been observing what I call the readiness gap — the space between where AI capabilities are and where organizations actually are in their ability to benefit from those capabilities responsibly.

That gap was already significant. The research has shown it for years: 94% of organizations are adopting AI in some form, but fewer than half have meaningful security controls. 72% report scaled deployments, but only 33% have governance structures to match.

Hubinger’s statement doesn’t just widen that gap. It changes the nature of it.

When the capability curve is something you can track from the outside — model releases, benchmark improvements, product launches — organizations can at least attempt to pace themselves. They can watch the horizon and plan accordingly.

When the lab’s own alignment lead tells you that recursive self-improvement is present-tense, the horizon is no longer a useful planning concept. The curve is now being drawn from the inside by the system itself.

This is not an argument for panic. It is an argument against the one posture that will definitely fail: waiting.

---

## The Physician Signal

This same week, the American Medical Association released survey data showing that 81% of U.S. physicians now use AI — more than double the 2023 rate.

Think about what that means for the governance argument. Physicians carry DEA licenses. They operate under HIPAA. They face malpractice liability. They are board-certified. They are arguably the most credentialed, most regulated, most scrutinized professional class in the United States.

And 81% of them are using AI right now, with no sector-wide governance framework to match.

If the professional class with the highest barrier to adoption — and the highest legal exposure for getting it wrong — has already crossed the threshold, the readiness gap isn’t a warning anymore. It’s the current condition.

The permission structure for adoption has collapsed. The infrastructure for governing it is still under construction.

---

## The Counter-Signal: What Dispatch Looks Like

The same week, Reuters reported that Meta is planning layoffs of 20% or more. The stated reason: to offset mounting AI costs.

Not a business downturn. Not a restructuring. AI costs.

The largest social platform in history — 3 billion users — is restructuring its headcount as a line item against AI infrastructure spend. The old economy sheds people while the new economy files its S-1.

I want to be precise about what I’m saying here: this is not a critique of Meta. It is a description of a pattern. When AI cost offsets become the stated rationale for major workforce decisions at platform-scale companies, the displacement thesis isn’t theoretical anymore. It’s a Reuters headline.

Organizations have a choice about which side of that pattern they’re on. The ones building governance infrastructure, workforce adaptation plans, and AI integration strategies that multiply human capability — those are the ones who come out of this with something. The ones waiting for a clearer signal are going to find that the clearest signal is the one they missed.

---

## What Readiness Actually Requires

I’m not going to tell you that AI will replace your entire organization. That framing, though dramatic, tends to produce paralysis rather than action.

Here’s what I will tell you:

The people building AI are telling you — on the record, in present tense — that the system is now contributing to its own acceleration. The most regulated professionals in the country are using it without governance frameworks. And the largest platforms are rewriting their cost structures around it.

The question every organization should ask is “are you using AI?” That battle is over. The question is: **when the technology moves faster than your planning cycle, what does your governance structure do?**

That is the readiness gap. And closing it — before the next capability jump, not after — is the only move that doesn’t leave you reacting to a timeline someone else is setting.

Hubinger isn’t making a prediction anymore. He’s giving a status update.

The question is what you do with it.

The February jobs report landed like a data point that didn’t know what story it was supposed to tell.

92,000 jobs lost. Forecasters expected gains of 59,000. A miss of 151,000 — in a single month. The coverage split immediately: some called it an AI displacement signal, others attributed it to federal workforce cuts and macro noise. The debate about which story is correct misses the more important point.

Both can be true. And if both are true, the mechanism underneath them is the same.

The Original Trap

In 1865, the English economist William Stanley Jevons observed something that shouldn’t have been possible. The steam engine had just become dramatically more efficient — burning far less coal to produce the same output. By every intuitive measure, coal consumption should have declined.

It accelerated.

Jevons concluded that efficiency gains in energy use don’t reduce consumption. They reduce the cost of consumption, which expands the range of applications that become economically viable, which increases total demand. The more efficiently you use a resource, the more of it you use.

This became Jevons’ Paradox: technological efficiency in resource use tends to increase, not decrease, the overall rate of that resource’s consumption.

For 160 years, the paradox was primarily an energy and environmental economics problem. Then agentic AI arrived — and the paradox found a new host.

The Agentic Reframe

The resource is no longer coal. It is organizational capacity — the cognitive, operational, and governance bandwidth that organizations use to coordinate work.

The efficiency claim of agentic AI is real. Agents automate workflows, compress decision cycles, eliminate coordination overhead. A process that required ten human touchpoints can be reduced to two. The math looks compelling.

But the Jevons dynamic is already running. When you reduce the marginal cost of deploying AI agents, the number of workflows organizations attempt to automate expands. The ambition scales with the capability. The surface area of exposure — to failure, to unintended consequences, to compounding risk — grows faster than the governance infrastructure to manage it.

This is not a future projection. The evidence is datable.

CENTCOM: Governance Void at T=0

In the same period the February jobs report landed, a separate signal confirmed the failure mode on the adoption side.

CENTCOM deployed AI into operational workflows on the same day the capability became available — before governance policy had been written, before oversight structures had been established, before the organization had built the infrastructure to sustain what it was deploying.

The restraint mechanism that was supposed to slow deployment until governance caught up collapsed faster than it had been installed. Capability deployment outpaced governance in hours.

This is the Jevons dynamic at the organizational scale: the efficiency of deployment — the low friction of standing up an AI agent — expands the rate at which organizations attempt to deploy them. The governance infrastructure required to sustain those deployments cannot build at the same speed. The gap between what organizations can deploy and what they can govern widens with every capability release.

Policy declarations are not restraint mechanisms. They are documents. The capability moves faster.

SWE-CI: Governance Void at T+8 Months

CENTCOM is the governance failure at the moment of adoption. A research paper published on March 4, 2026 documents the governance failure that comes later.

Researchers from Alibaba Group and Sun Yat-sen University built the first AI coding benchmark measured not on a snapshot test — agent receives a problem, produces a solution — but on a full production evolution timeline. 100 tasks across real codebases. Each task spanning an average of 233 days and 71 consecutive commits. Agents were evaluated not just on whether they solved the immediate problem, but on whether the code they produced could sustain the codebase through months of continued evolution.

The headline finding: most models achieve a zero-regression rate below 0.25. In other words, in more than 75% of cases, AI agents that pass standard coding benchmarks introduce regressions when they maintain real production systems over 8-month timelines.

An agent that hard-codes a brittle fix and one that writes clean, extensible code may both pass the same test suite. Their difference becomes visible only when the codebase must evolve — when new requirements arrive, interfaces change, and modules must be extended.

This is not a capability gap. It is a governance gap. The organizations deploying AI agents at speed, on the strength of short-horizon benchmark performance, are not building the oversight infrastructure to detect degradation as it accumulates. They are accelerating a dynamic that already existed in human-maintained code — and compounding it.

The Anthropic Signal

The Anthropic Economic Index, published in early March 2026, provided the labor market anchoring for what the Jevons dynamic predicts.

The paper found AI most heavily used in automation of existing tasks rather than augmentation or new task creation. The composition of displacement is uneven: computer and math occupations show the highest theoretical capability exposure (94%) but the lowest observed AI coverage (33%). The gap between what AI can theoretically do and what organizations have actually integrated is the readiness gap — and the Jevons dynamic is running in both directions simultaneously.

Organizations that automate efficiently create pressure to automate more. Organizations that haven’t automated yet face competitive exposure that accelerates adoption without governance. Both trajectories compound the same underlying problem: the efficiency of deployment has outrun the capacity to govern what gets deployed.

The February jobs report is, in this frame, not primarily a story about AI replacing workers. It is a story about organizations that optimized for deployment speed without building the readiness infrastructure to sustain it. When the capability scales faster than the organization can absorb, the excess goes somewhere. Sometimes it goes to regressions in production code. Sometimes it goes to workforce restructuring that moves faster than the institutional knowledge can be transferred. Sometimes it goes to governance vacuums that the next capability release will find already open.

The Paradox in Full

The Agentic Jevons Trap is not the observation that AI will increase demand for labor in the long run — though it might. It is the observation that the efficiency of AI capability deployment is increasing the rate at which organizations consume governance capacity, oversight infrastructure, and organizational readiness.

Every capability release that lowers the cost of deploying an agent expands the number of workflows organizations attempt to automate. Every expansion in automated workflows increases the surface area of exposure to the failure modes SWE-CI and CENTCOM document. Every governance vacuum the capability outpaces becomes a liability that compounds over the 233-day production timeline no benchmark was measuring until now.

The paradox is structural. It does not resolve by deploying better AI. It resolves only by building the organizational readiness infrastructure that can govern what the capability enables.

That is a different kind of work. It is the work nobody is selling.

The Counter-Paradox

Jevons himself did not have an answer to his paradox. The efficiency gain was real. The consumption increase was real. The gap between them was a structural feature of how markets respond to cost reduction.

The counter-paradox for the agentic version is governance as a restraint mechanism — not as a document, not as a policy declaration, but as infrastructure. The organizations that build oversight architecture before they need it, that establish the readiness primitives before the capability arrives, that treat governance as a design constraint rather than a compliance checkbox — those organizations close the gap between deployment speed and absorption capacity.

The window for that work is open. It is not wide.

The February jobs report is not a warning from the future. It is a reading from organizations that already ran the experiment. CENTCOM and the SWE-CI data bracket the failure arc: governance void at adoption, governance void at maintenance, compounding across every timeline in between.

The technology has arrived. The question is whether the organization has.

This week Salim Ismail posed a question on the Diamandis podcast that most leaders are only asking in private — what happens to organizations when the “human checkpoint” disappears from the workflow? When every process that once routed through a person routes through an agent instead, and humans move from being inside the work to overseeing it?

He called his working thesis the Organizational Singularity.

It’s a good frame. And it’s pointing at something real.

But here’s what I keep coming back to: the question isn’t what the future of organizations looks like. We’re already starting to see it. The question is whether organizations are actually ready to get there — and by almost every measure, they are not.

The numbers are not subtle.

79% of enterprises are experimenting with AI. Only 8.6% have made it to production. 94% have adopted AI tools — but fewer than half have the governance frameworks to support them. 72% have scaled pilots. Only 33% have governed what they scaled. And after all of it — after the announcements, the pilots, the transformation roadmaps — only 6% can point to measurable impact on the bottom line.

Read those numbers again. Not because they’re surprising. Because they tell you exactly what kind of problem this is.

This is not a technology problem. The technology has arrived. It is capable, it is accelerating, and it is not waiting for anyone to catch up. The gap — the stubborn, persistent, expensive gap between what AI can do and what organizations are actually capturing — is not a technical gap.

It is a readiness gap.

Technology that arrives without organizational readiness doesn’t transform.

It accumulates.

It accumulates in the form of pilots that never scale. Roadmaps that never land. Investments that produce dashboards instead of decisions. AI that sits beside the workflow instead of inside it. McKinsey calls it “bolted on.” I call it expensive decoration.

The organizations spending the most on AI transformation are not necessarily the ones seeing the most return. The ones seeing return are doing something different — they are treating AI as a reason to redesign how work actually gets done, not as a tool to layer on top of how it has always been done.

That distinction sounds obvious. It is apparently not.

Naval Ravikant said something recently that stopped a lot of people in the technology world. He said software is “uninvestable.”

He wasn’t making a valuation call. He was making an observation about durability. In a world where AI can generate code on demand, the competitive moat is no longer the software itself. It is the organizational infrastructure to deploy that software at scale — the governance, the judgment, the human architecture that decides what gets built, how it gets used, and what it is actually for.

That’s not a technology question. That’s a readiness question.

I’m working through what readiness actually means — as a practitioner inside the transition.

The Technology layer. The Business layer. The Human layer.

Because Salim is right that the organizational singularity is coming. What he’s still writing toward — and what I think most of the conversation is missing — is the answer to the more urgent question:

How do you get there from here?

That’s what I’m building toward.

More soon…..