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.
Sign in to read the full notes
Get access to AI-generated notes, topic timestamps, and more.