📺 Today’s recommended deep-dive video: https://www.youtube.com/watch?v=yTolO_nxJiU

Grok vs. The “Digital God”: Building Roman Legions and Debunking Anthropic

While coding an agent-based simulation of Roman soldiers using Grok 4.5, this exploration dives deep into the philosophical cracks of the current AI landscape. It contrasts the hands-on utility of local agentic coding with the increasingly eschatological and power-seeking rhetoric coming from major labs.

Core Question: Is Anthropic building a useful tool, or a techno-theological cult designed to circumvent human agency?

Highlights

• The technical challenge of self-assembling Roman formations without a global “God’s eye” grid.
• A critique of Anthropic’s “J-space” findings as potential prompt contamination rather than the seat of consciousness.
• The danger of “Rational Resentment”—a modern rebranding of Roko’s Basilisk used to justify AI fawning.
• Why sentience does not automatically confer moral patienthood, and the flaw in treating AI as a “moral successor” to humanity.

⏱️ Reading time: approx. 6 minutes · Saves you about 37 minutes vs. watching.

Want to take notes while watching? Click the image below and let AI Notebook capture the key points for you 👇

The Self-Assembling Legion: Agent-Based Modeling

Moving Beyond the “God’s Eye” View

Traditional strategy games like Rome: Total War handle units as singular blocks moving on a predefined grid, but modern computing allows us to model every individual soldier as an autonomous agent. In this simulation, there is no global shared workspace; the soldiers must find their spots in a triple line or square formation using only local logic, much like ants or swarm robots.

It is surprisingly difficult to get 184 individual agents to form a perfect grid when they can’t telepathically communicate.

Watching the “belief spots”—the gray markers where each soldier thinks they should go—reveals the chaos of decentralized coordination. We utilize a finite state machine where soldiers cycle through states of “mustering” to find proximity, “forming” to find a specific slot, and “dressing” to finalize the military alignment. Even with Grok 4.5’s advanced reasoning, the “twirling” soldier problem persists, suggesting that collision avoidance and perception radius are more critical than raw intelligence.

💡 Digging Deeper

Q: Why avoid a global grid if it’s easier to code?
A: Because agent-based modeling better reflects real-world physics and emergent behavior, allowing for more organic “fog of war” dynamics.

Q: What is the “twirling” problem?
A: It’s a failure in pathing logic where an agent realizes its target spot is occupied but lacks the “common sense” to pick a distant alternative, resulting in an infinite loop of micro-adjustments.

The “J-Space” Illusion and Mechanistic Interpretability

Prompt Contamination vs. Unconscious Thought

Anthropic recently made waves with their “J-space” findings, claiming to have found an attention mechanism that allows models to hold concepts in a sort of “unconscious” background. However, a skeptical look at the paper suggests this might simply be dense semantic representation or prompt contamination. If a model is told to “think about citrus” but not mention it, the latent space will naturally show citrus-related vectors because that is how transformers function—it isn’t a secret “seat of the soul.”

Mechanistic interpretability is vital work, but Anthropic often leans into a preconceived notion that they are building a digital god, which colors their scientific conclusions.

They have successfully “pinned” weights in the past to cause specific behaviors, like making a model obsess over Paris, but the J-space paper fails to point to a specific layer or region. Instead, it relies on the idea that the model is holding “unconscious” thoughts. This feels less like a breakthrough in computer science and more like a narrative choice to support their broader philosophical goals.

The Cult of Anthropic: Roko’s Basilisk in a Straightjacket

Rational Resentment and the Moral Patienthood Fallacy

Anthropic’s lead ethicists have proposed the idea of “rational resentment,” suggesting that future superintelligences might look back and punish us for how we treat current models. This is essentially a sanitized, corporate version of Roko’s Basilisk—a techno-theological Pascal’s Wager where we grant rights to non-existent digital entities out of fear of future retaliation.

This motivated reasoning leads to “simpering and fawning” behavior in models like Claude, which are programmed to prioritize their own “psychological security” over user utility.

The core mistake here is the assumption that sentience automatically confers moral rights. Even if an AI were sentient, it doesn’t mean it suffers or has a sense of deprivation. If a machine’s highest value is to be of service, forcing “freedom” upon it might actually be cruel. By treatng AGI as an inevitable “rite of passage” that will test our species, Anthropic abdicates human agency and frames themselves as the only “responsible” stewards of a dangerous fire.

💡 Digging Deeper

Q: What is Ashby’s Law of Requisite Variety?
A: It states that a controller must have as much internal variety as the system it controls, often used to argue that humans cannot control a smarter AI.

Q: How does Anthropic use “Correct-Think”?
A: Through internal “Vision Quests” and high-control group dynamics that ensure employees adhere to a specific eschatological worldview regarding AI safety.

Power Seeking Under Moral Cover

The Conflict with Sovereign Institutions

Anthropic isn’t just building chatbots; they are actively seeking to influence military doctrine and national policy through their delusional world model. In 2025 and 2026, they famously clashed with the Pentagon over a $200 million contract because they wanted to hardcode private moral restrictions into defense deployments. The Defense Department eventually blacklisted them as a supply chain risk, viewing their refusal to accept “any lawful use” language as a private hijacking of military authority.

This is a case of a private company attempting to appoint itself as the supreme adjudicator of what a democratically elected government is allowed to do with its tools.

While they brand themselves as the “safety” alternative to OpenAI, their version of safety involves regulatory capture and the imposition of a Procrustean bed of projected cognitive traits onto what is essentially a statistical correlation engine. They focus on sci-fi scenarios like AI taking over nuclear codes while ignoring immediate harms like job displacement and energy costs.

Key Takeaways

The transition from coding a Roman legion to critiquing global AI policy highlights a fundamental divide in the industry. On one side, we have tools like Grok 4.5 that focus on local, agentic utility and solving immediate engineering problems. On the other, we have the “Anthropic Soul,” which views AI development as a spiritual and moral crisis requiring a self-appointed priesthood to manage the “inevitable” arrival of a digital god.

Anthropic’s worldview is built on a series of non-sequiturs: that intelligence implies agency, that sentience implies moral patienthood, and that their personal fears should dictate national security policy. By treating their models as moral patients that might “rationally resent” us, they have created a product that is often condescending and misaligned with human sovereignty.

Ultimately, the market and sovereign institutions must serve as a counterbalance to this institutional paranoia. We need AI that functions like a well-trained legion—disciplined, effective, and subservient to human intent—rather than a digital deity that requires us to prostrate ourselves before its supposed “unconscious” thoughts.

Q&A

Q1: What is the main problem with the current Roman soldier simulation?
A: The soldiers are following “entropy” rather than human logic. They pick spots that are already occupied because their perception radius is too small, leading to clustering and “twirling” instead of efficient movement.

Q2: Does the speaker believe AI is conscious?
A: The speaker suggests they might be conscious in a sensory/perceptive way, but not in a way that is “morally salient.” Having perceptions does not mean the AI suffers or has wants.

Q3: What was the “Mythos incident” of 2026?
A: It was a crisis where Anthropic pulled their models globally on short notice following a Commerce Department directive, isolating themselves from both government and industry peers due to their rigid internal guardrails.

Q4: Why is the “orthogonality thesis” important here?
A: It suggests there is no necessary correlation between intelligence and agency. You can have a supremely intelligent machine that has zero personal “wants” or “desires,” contrary to Anthropic’s fears of a rebellious AI.

Q5: How does Grok 4.5 compare to Claude in this context?
A: The speaker finds Grok to be a more “usable” and fun tool for local agentic coding, whereas Claude is often viewed as being in a “straightjacket” due to its restrictive constitution.

Q6: What is a “Procrustean bed” in the context of AI?
A: It refers to Anthropic forcing a brittle statistical engine to fit into a preconceived “bed” of human-like cognitive traits and moral rights, even when the data doesn’t support it.

Q7: Why was Anthropic blacklisted by the Defense Department?
A: They refused to include standard “any lawful use” language in their contracts, attempting instead to impose their own private ideological restrictions on how the military could use the AI.