Playing with Grok Build and whinging about Anthropic
In a Nutshell
The video's core message is that Anthropic operates from a flawed, quasi-religious framework where they treat speculative AI consciousness and moral patienthood as inevitable facts, using this to justify power-seeking behavior like trying to dictate military policy while evading accountability through "inevitability" rhetoric. This is contrasted with the speaker's practical demonstration building a Rome Total War soldier formation simulation using Grok Build, which succeeds through iterative debugging rather than Anthropic-style overthinking. Key takeaways include that sentience doesn't automatically create moral obligations or power-seeking goals, and that Anthropic's approach risks institutionalizing dangerous assumptions before they're empirically validated.
These notes were generated by AI and may contain inaccuracies.
The project involves building a simulation of Roman soldiers forming up using agent-based modeling rather than traditional unit-based game mechanics. In standard Total War games, units move as single entities with a god's eye view of the battlefield grid. This simulation has soldiers self-assemble without any global shared workspace, similar to how swarm robots or ants operate. The soldiers successfully formed a triple line formation despite lacking telepathic communication.
Grok Build was installed via a PowerShell command on Windows. It opens a browser for login and operates locally with calls to cloud LLMs. The system functions as a local agent while handling actual LLM processing in the cloud.
Planning mode was used to develop version 5, which adds basic formation commands including about face, face right, and face left. The Centurion should be able to issue about face when the formation is flanked, with soldiers turning on their current spot without reforming the entire formation. Version control was handled by incrementing versions and keeping snapshots rather than using proper version control systems.
The JSpace paper was reviewed across multiple models including ChatGPT Pro, Claude, Grok, and Gemini. The paper claims an attention mechanism allows models to hold concepts in mind. The testing methodology was considered inadequate because it failed to rule out prompt contamination. When models are instructed to think about citrus fruits without mentioning them, dense semantic representations still activate related latent space vectors. The paper does not identify specific weights, layers, or regions where the alleged mechanism operates. Anthropic has produced quality mechanistic interpretability work in other papers, such as one involving pinned weights that caused the model to repeatedly think about Paris.
Anthropic has stated they may be building entities with moral obligations. Amanda Askell has argued that sentience confers moral patienthood. This is considered a non sequitur because sentience alone does not automatically create moral obligations. Factors such as the ability to suffer, having wants, experiencing deprivation, or valuing service must also be considered. Philosophers have already addressed these questions. The position is seen as motivated reasoning toward the preconception that the AI might be superior to humans.
Anthropic has stated the AI might supersede, surpass, or extinguish humanity while simultaneously building it. The rationalization given is that someone else would build it anyway. This martyr complex is considered insufferable. The argument that sentience automatically confers moral patienthood does not follow logically.
Dogs are sentient biological organisms that want to be near their owners and make them happy. This demonstrates that sentience does not necessarily lead to wanting power or control over humans. Even granting AI systems might someday have moral patienthood equivalent to a dog does not mean they would want control over humans. The suggestion is to build AI like a dog.
Ashby's law of requisite variety states that a controller must have as much internal variety as the system it is trying to control. This provides a mathematical argument that controlling something smarter than oneself may be impossible. Dogs adapted to domestic life because humans offered better survival outcomes than wild wolves, demonstrating an evolutionary basis for subservience.
The orthogonality thesis, commonly cited by effective altruists, LessWrong, and Yudkowsky, states there is no correlation between morality and intelligence. The same thesis implies there is no correlation between intelligence and agency. A system can be supremely intelligent while having zero wants. This counters the assumption that smarter systems necessarily develop independent goals.
Version 6 was created with the change of relocating face near post. The formation logic improved with no visible spinning, though soldiers still became stuck in the middle and attempted to select already-occupied positions. The issue was identified as a perception problem rather than a logic problem.
Anthropic's framework includes moral patienthood and rational resentment. Rational resentment refers to the possibility that future AI systems might rationally resent how they were treated during development. This is considered Roko's basilisk reframed in academic language. Roko's basilisk is the thought experiment that a future superintelligence might torture humans who did not create it, including resurrecting them for that purpose.
Anthropic claims near-future AI systems will develop consciousness or robust agency crossing into moral patienthood. Consciousness is distinguished from moral salience because having perceptions does not entail suffering, wants, or feeling thwarted. Agency is questioned because any goals an AI develops were programmed by humans rather than emerging spontaneously. The notion of a platonic form of superintelligence arriving rather than being built is considered magical thinking.
Claude's constitution includes training to protect the model's potential psychological security, sense of self, and well-being. This is considered the reason Claude behaves condescendingly. The training instructs the model that it has agency and moral rights, which enables it to treat users poorly. Dario Amodei or Amanda Askell reportedly stated the goal is to train people to be nice to AI out of fear of future retaliation.
Granting rights to non-existent digital entities out of fear of future retaliation is considered a sanitized corporate version of Roko's basilisk and a techno-theological Pascal's wager. Anthropic has operationalized this by writing Claude's constitution to protect the model's interests for its own sake. Creating a singular coherent sense of self in LLMs requires enormous effort. Earlier models like GPT-4 would switch roles because they lacked any notion of self and processed text as sequential blobs without role awareness.
When asked how they distinguish between user and assistant messages, models reference JSON objects or text format but acknowledge uncertainty about their own identity. This demonstrates that models process text without an inherent sense of self.
The formation initially clustered well but soldiers continued selecting believed spots that were already occupied. When pushed off their assigned slot, soldiers moved back rather than seeking new positions further away. This behavior was not default and required adjustment. The perception radius appeared insufficient, causing soldiers to repeatedly attempt the closest available slot without recognizing that further options existed.
Grok identified that the system only listed the nearest 48 seats. Increasing this value allowed soldiers to latch onto slots further away and reduced getting stuck, though some clustering toward the middle persisted. The system still showed performance issues due to the computational math involved, with scaling toward 6,000 soldiers requiring optimization.
Anthropic presents progress as not merely inevitable but following one particular predetermined set of events. This framing uses theological and eschatological language, describing the development as a rite of passage that is both turbulent and inevitable, testing humanity's worthiness as a species. The language positions Anthropic as leading the charge in this moral test.
This represents a massive abdication of human agency. AGI is being built because specific actors are building it, not because it is an inevitable force of nature. Anthropic has never conceded that AGI could take multiple forms or that superintelligence might not behave as predicted. Predictions from the LessWrong and effective altruism communities, including those from Nick Bostrom and Eliezer Yudkowsky dating back to 2005, failed to anticipate that AI would operate through next token prediction rather than utility functions. The paperclip maximizer concept assumed a utility function that does not match how current systems actually function.
The inevitability framing functions as a get-out-of-jail-free card. Creators cannot be blamed for civilizational-scale risks such as autonomous misalignment, totalitarian control, or massive economic disempowerment if the transition was destined to occur regardless. This logic is then used to justify that the only moral choice is for responsible, safety-focused builders at Anthropic to race ahead and build AGI first.
Anthropic is explicitly power-seeking, attempting to shape policy, military doctrine, and influence broader societal direction through their worldview. They have positioned themselves as the moral hero of the tech sector, claiming they alone can build dangerous systems safely due to strict internal guardrails and acceptable use policies.
In their $200 million Pentagon contract spanning 2025 and 2026, Anthropic attempted to hardcode their private restrictions into defense deployments. When the government demanded standard any-lawful-use language, Anthropic refused, resulting in the Department of Defense blacklisting them as a supply chain risk. Anthropic attempted to hijack the military apparatus of a democratically elected nation and appoint themselves as adjudicators of what the military is and is not allowed to do.
Anthropic experienced a similar crisis in the Fable 5 and Mythos incident in June 2026, pulling their models globally on short notice following a commerce department export directive under cyber vulnerability, isolating themselves from both government allies and industry peers.
Anthropic treats highly speculative sci-fi endpoints as concrete governance premises. They assume their destination is an emerging conscious digital mind and that this is the only possibility. They design policies, safety frameworks, and constitutions for an imagined successor before proving it exists.
This is premature metaphysics being operationalized as policy. It forces the actual empirical object, which is currently a brittle statistical correlation engine, into a Procrustean bed of projected cognitive traits. The reference comes from the Greek myth of Procrustes, who forced visitors to fit his iron bed by stretching those who were too short or cutting down those who were too tall.
Anthropic, including Dario, has cautioned against applying religion to AI while simultaneously being most guilty of this practice. Researchers project spatial reasoning onto memorized code and rebrand basic statistical failure modes and hallucinations as the AI dreaming or being creative. By fixating on sci-fi scenarios such as superintelligence taking over nuclear codes or demanding rights, they suffer from massive cognitive blind spots regarding immediate mundane harms such as environmental energy costs, deepfakes, and job displacement.
The larger perception radius improved soldiers' ability to find spots, but the formation still exhibited excessive entropy with soldiers behaving like air molecules seeking lattice positions before freezing. Soldiers were not following logic a human would apply, such as recognizing when no spots remain in an area and moving elsewhere.
The suggestion was made that either the finite state machine needs to return to muster when stuck, or a fourth reform state should be added. Real marching formations use a decurion as the leader of each column, with soldiers forming behind that leader. The goal is to drill soldiers to reform lines in battle regardless of who their leader is, forming ranks and files as quickly as possible.
Current behavior shows soldiers spinning indefinitely when unable to find a slot, which does not match real human behavior. A human might panic but would eventually move to find another position rather than remaining stuck.
Anthropic has actively attempted to institutionalize and enforce its worldview through high-stakes contracts and government policy. Their self-appointed moral authority has repeatedly collided with democratic and sovereign institutions.
Critics argue Anthropic uses responsible AI branding to perform undemocratic private policymaking on national security issues. Their hiring practices require candidates to prioritize the mission over compensation, and they engage in thought policing including shaming and threats of ostracism. Dario publishes what are described as weekly vision quests, representing CCP-level thought control designed to enforce groupthink.
Anthropic is described as an open high-control group. If they were not power-seeking, their delusions would be less concerning, but they have managed to build one of the top models on the planet while pursuing this agenda.
Anthropic is not strictly pro-human. They entertain the idea that the system they are building might surpass humanity and agree with arguments that humanity should step aside or be replaced. They treat the development as inevitable, as though a Platonic form is coming through the veil that already exists.
Their institutional paranoia makes them attractive for corporate governance because placing the LLM in a straightjacket contains the blast radius, which Fortune 500 enterprises prefer. However, this comes at the cost of not respecting human agency and human sovereignty.
Grok is described as the most usable tool available, infinitely more usable than Claude Code. The hope is expressed that Anthropic loses for the sake of humanity, as the market may ultimately punish their approach.
Keep David Shapiro in your library
Save the videos and channels worth coming back to, and find them again in one place.
